Changelog — Loop Engineering¶
All notable changes to LOOP_ENGINEERING.md and the docs/ knowledge base are recorded here.
Format follows Keep a Changelog.
Versioning follows Semantic Versioning:
- MAJOR — existing doc removed/renamed, or LOOP_ENGINEERING.md index restructured
- MINOR — ≥1 new docs/*.md file created (new concept documented)
- PATCH — existing docs updated or new findings added to digest with no new doc files
- None — zero findings and zero doc changes (no version cut; the run still commits its digest section)
[Unreleased]¶
Added¶
Changed¶
[3.7.9] — 20261019 18:21¶
Backlog A19 — PATCH: no new docs/*.md. The 2026-10-19 run swept every X source with Chrome logged
out of X, published v3.7.7 and exited 0; the only notice was a line inside the digest. Stage A now reads
X's login state before dispatching anything, and the wrapper stops the run, with its own exit code, before
anything is published. Proven by running the real wrapper against stubbed sessions; not yet exercised by
a real tracker run (the logged-out coverage records and the skill's preflight are prose a model follows —
the wrapper's defence is that a missing or malformed x_login fails the gate).
Added¶
scripts/assert-x-login.sh— the wrapper's reading of the artifact's newx_loginfield. Only a positivelogged-in(stamped today, with evidence, uncontradicted by the coverage records) passes;logged-out,unknown, a missing or malformed field and an unreadable file all stop the run.scripts/verify-x-login-guard.sh, wired intoguards.yml— 21 decision cases on fixture artifacts, 15 cases that run the realrun-loop-news.shend to end against stubbedclaude/script/osascript, 6 static placement/contract checks, 6 cases that run the skill's own preflight snippet under node, and 7 mutants (of the wrapper and of the script) that must each be caught.- Wrapper exit 8: Chrome is not logged in to X, or that could not be shown. It fires right after
Stage A, before the retry logic and before Stage B, keeps the Stage-A artifact so a re-run resumes, and
notifies.
--allow-x-logged-out(forwarded byrun-loop-news-now.sh) publishes without X for that run only: X sources are recordedblocked: x-login, and the digest's Coverage section opens by saying X was not swept by the operator's choice.
Changed¶
fetch-loop-newsPhase 1 step 4: one read-only browser call onx.com/homebefore any X source is dispatched; the result is recorded asx_login. Logged out or unknown: no X source is swept, each gets ablocked: x-logincoverage record, and Stage A stops unless the operator's override file is present.integrate-loop-news: the Coverage subsection states an overridden X gap on its first line.run-loop-news-now.shforwards its remaining arguments to the wrapper (before, they were ignored) and explains exit 8.docs/09-headless-mode.md(the preflight pattern, its limits),scripts/SCHEDULING.md(exit 8, the login step, the override),README.md,plans/20260904_2053-open-work-backlog.md(A19struck,A21opened).
[3.7.8] — 20261019 17:29¶
Retry pass on the 2026-10-19 run — PATCH: no new docs/*.md, 18 docs extended, 13 docs corrected.
The run swept X with a logged-out browser (0 of 11 x sources fully covered); after logging in, the 13 X
sources were re-swept (now 8 of 11 complete, the rest partial with their gaps stated) and 193 new
findings were added to the digest. 59 Tier 1-2 findings were integrated into docs by four writers working
from primaries, each independently verified claim by claim (219 claims: 208 verified, 11 corrected, none
fabricated). A fact-check of the original run's doc changes (312 of about 631 claims) found 21 wording or
attribution errors, now fixed — one contradicted claim, one invented quotation, one reverted timeout.
Evidence: plans/20261019_1657-retry-of-20261019-run-evidence.md.
Changed¶
docs/21-context-vs-loop-engineering.md,docs/01-paradigm-shift.md,docs/20-loop-maturity-model.md,docs/32-reading-list.md— a dated October 2026 discourse snapshot (definitions, critiques, the Laforge/Venema deck, the June wave dated from The Register); promotional posts labelled as suchdocs/16-memory-patterns.md— Pattern L, a lessons file with a count-based promotion rule (magpie)docs/25-long-running-agents.md,docs/26-factory-model.md,docs/28-routines.md,docs/36-development-workflow.md— OpenClaw's test-pruning loop, the dark-factory label against its primaries, a practitioner triage routine, the lix loop with its PR figures checkeddocs/24-harness-patterns.md,docs/11-cost-control.md,docs/12-hooks.md,docs/13-context-management.md,docs/19-mcp-security.md— SoL-Pi (from the paper and repo), the Apodex ranking reversal, the multi-harness RL guide, Claude Code mods, blade-code, Strands Boxdocs/23-fleet-engineering.md,docs/29-background-agents.md,docs/37-session-architecture.md,docs/38-agent-teams.md— OpenClaw's shared gateway and inter-session receipt collapseLOOP_ENGINEERING_NEWS.md— a hand-authored "Retry pass" entry (coverage after the retry, 193 rows, held-out claims, corrections); eleven rows of the 15:02 run carry bracketed correctionsKB_GAPS.md— a retry-pass gap block; a note on the 15:02 block that its X gap is now partly closed
Fixed¶
docs/23: hearting v3.35.0 was said to replace blanket timeouts; its PR adds an earlier stop and keeps the hard limitdocs/24: HarnessRouter v0.31.0's 3-minute silence timeout was taken back in v0.31.1; the doc now says sodocs/07,docs/17: a quotation not on its cited page, and two quotes that differed from their sourcedocs/21: two unconfirmable LinkedIn claims relabelled;docs/36: two of the lix PR figures are narrower than the article reads- sixteen smaller wording, date, attribution and link fixes across
docs/01 04 13 19 23 24 26 27 28 36
Not changed¶
- The
v3.7.7entry above stands as written; its X coverage statement ("0 of 11") is superseded by this entry, not rewritten.
[3.7.7] — 20261019 15:30¶
Tracker run 2026-10-19 15:02 UTC — PATCH: no new docs/*.md, 29 existing docs updated, 339 digest
rows (366 in the artifact; 8 already published, 19 folded). First run after a 21-day gap
(last_run_date 2026-09-28). Every x sweep was partial (logged-out browser; 0 of 11 fully
covered) — see the digest's Coverage section. Doc agents worked from the search stage's captured
summaries and fetched a primary only for verbatim quotes and figures; second-hand and vendor claims are
labelled as such in the docs.
Changed¶
docs/04-verification.md— review loops that converge (Anthropic's own code-review process and five open-source ledgers/gates) and measuring whether a guard, gate or judge works (GateGuard's null pilot, caliper, Wilson-interval gate evals, harness test-adequacy papers)docs/24-harness-patterns.md,docs/22-learned-orchestration.md— one literature map of the late-2026 harness-evolution papers inside Self-Improving Harnesses; Harness², RunTune, Agent Lightning; UHP/HarnessRouter at the harness boundarydocs/33-agent-security-hardening.md,docs/19-mcp-security.md,docs/08-permissions.md— out-of-process runtime enforcement, containment failures that went around tool-level controls, monitors and logs as attack surface, state-file injection laundering, approval-to-execution bindingdocs/11-cost-control.md— the 5.5-generation model/price table, turns-not-observations cost, routing and decision models, hard budget caps, measured team-cost multiplesdocs/26-factory-model.md,docs/23-fleet-engineering.md— Scion and the merge-volume bottleneck; what open-source factory floors converge on; Kubernetes-for-agents; eight unattended-fleet recovery patternsdocs/38-agent-teams.md,docs/39-dynamic-workflows.md,docs/07-subagents.md— measured evidence on whether teams pay off; flows in Managed Agents; per-subagenteffortdocs/28-routines.md,docs/12-hooks.md,docs/13-context-management.md,docs/16-memory-patterns.md,docs/29-background-agents.md,docs/37-session-architecture.md,docs/03-building-blocks.md— Anthropic's six rules for a reference automation, mods, idle compaction and Fork-and-Flush, git-backed memory with a Dreaming pass, other vendors' always-on agentsdocs/27-loop-contract.md,docs/30-goal-engineering.md,docs/25-long-running-agents.md,docs/20-loop-maturity-model.md— Guards That Failed Open (eight repos), writing a condition/goalcan judge, control-plane continuity, score what is provendocs/21-context-vs-loop-engineering.md,docs/17-failure-patterns.md,docs/36-development-workflow.md,docs/01-paradigm-shift.md— the vocabulary formalised and contested; review-loop drift; back pressuredocs/32-reading-list.md— added two Anthropic first-party entries, removed one (group stays at 5); LoopX link follows the repo renameLOOP_ENGINEERING.md— summary clauses on 23 index rows; cross-links docs/04 ↔ docs/17, docs/17 → docs/27, docs/24 UHP bullet → its new subsectionSOURCES.md,README.md— 89 rows (was 87): three added (claude.dev/blog, anthropic.com/engineering, chaseai-yt/claudex-loop), the two LoopX rows merged, five rows switched torss, renamed/moved repos and theclaude.com/blogredirect corrected, ~50 dated cadence/dormancy notes; the tracker's expected-keys set changes with these rows (13 keys removed, 15 added — listed in the run's digest section). README counts corrected to matchKB_GAPS.md— 2026-10-19 gap block (15 open gaps + structural candidates deliberately left for a decision)
[3.7.6] — 20260929 19:00¶
Backlog A17: closes the root cause behind the 2026-09-28 run's 7 partial X sweeps —
fetch-loop-news ran every browser-bound sweep at once against one shared Chrome, and its
Pass-2 stop rule could not tell a stalled tab from the end of a timeline. Shipped after three
rounds of adversarial review by fresh subagents (35 findings total, all applied or fixed); not
yet exercised by a real tracker run.
Changed¶
fetch-loop-news: browser-bound sources (x,x-search,linkedin) run at most 3 at a time; everything else at most 15 — unbounded, the non-browser lane alone would saturate Claude Code's 20-concurrent-subagent cap, the same failure one lane over.- A page counts as stalled, not ended, unless it stops growing with no loading spinner and (for a zero-post page) shows X's own empty-search message; day-range searches cover the rest of the window after a stall.
- Pinned posts are detected by their own DOM label plus a second confirming post, not by date order alone, so one out-of-order pin can no longer end a scroll early.
- Every source (
x,x-search,linkedin,phase-3:x-general-search) gets a keyed coverage record (complete/partial/sampled, keyed on the unique<type>:<Handle/URL>— Actor names repeat), re-checked when banked and again byintegrate-loop-news;x-search,linkedinand the Phase 3 search must read a minimum post count to count assampled. Partial browser sources are re-run once, resumably; every source of every type that never returned gets a "not swept" record. integrate-loop-newspublishes a### Coveragedigest section built from the expected set of sources, not from whichever records happen to exist, so a partial, missing or unrecorded sweep stays visible instead of reading as a complete run.plans/20260904_2053-open-work-backlog.md,RESUME.md:A17marked shipped;CLAUDE.md's "Open work" pointer back to two open items (approved by the maintainer).
[3.7.5] — 20260929 16:05¶
A hand-run retry of everything the 3.7.4 tracker run (2026-09-28 09:52 UTC) left
incomplete — 7 X sweeps partial on a contended shared browser, @kachar136 thread
expansion, 7 findings cited by profile URL instead of post URL, and three open verifications —
in three verified rounds, then integrated, reviewed and critic-checked. Digest:
## Retry pass — 2026-09-28 13:23 UTC (hand-authored, not a tracker run). Evidence pack:
plans/20260928_1401-retry-of-20260928-run-evidence.md.
Added¶
docs/34-loop-patterns.md— Thin Loop, an eighth named pattern (deep-read from cobusgreyling/loop-engineering); the pattern count moves seven → eight indocs/10,docs/11,docs/20andLOOP_ENGINEERING.mddocs/04-verification.md— "Formal Models as a Bug-Finding Loop" (bcherny's Lean/TLA+ verification of the Claude Agent SDK, self-reported); OpenClaw's test-audit authoring gate; the libuv fix series; kachar136's convergence signaldocs/31-claude-tag.md— first-party case study: claude.ai's performance sprint run in Tagdocs/25-long-running-agents.md— Anthropic's ART genome-mining campaign from its primary sources (post, preprint), with Nature News's outside caveat- New material in
docs/05(the AGENTS.md-only-with-telemetry gate),docs/11,docs/17,docs/19,docs/23,docs/24,docs/27anddocs/29 SOURCES.md—www.anthropic.com/newstracked (a first-party post reached the digest only secondhand);README.mdsource counts follow- Backlog
A17—fetch-loop-newsruns every browser sweep at once, and its X Pass-2 stop rule cannot tell a stalled tab from the end of the timeline (the root cause of the partial sweeps)
Changed¶
docs/07,docs/11— Opus 5.5 blog quotes made verbatim; "3.3x / 2.6x" is aggregate Claude Code usage March–September 2026, not a comparison with Opus 4.6;docs/11cites the prompt-caching page that confirms effort switching keeps the cache (with its exclusions)docs/17— the Delete gate "lowercased after matching" (was "before"); the 20% quote moved to its real source postdocs/29— Autopilot-on-OpenClaw cited first-party (Microsoft's Autopilot CVP)docs/04,docs/17,docs/19,docs/27,docs/31— 7 X citations point at the exact postLOOP_ENGINEERING_NEWS.md— the 3.7.4 run's rows corrected in place by appended notes (kirillk primary source, cobus "conflict", Opus 5.5 figures, Autopilot, SQLite/goal, 400k/20%, Tag expansion); seven link cells now point at the resolved post (disclosed there)KB_GAPS.md— cobus "conflicting metadata" filled (there was no conflict); four gaps loggedRESUME.md— handover header for 20260928CLAUDE.md— "Open work" now names three §5 items (A17added); approved by the maintainer
[3.7.4] — 20260928 10:27¶
Added¶
Changed¶
docs/24-harness-patterns.md— three independent meta-harness/harness-router products (openrig, orchestra) converged alongside omnigent/HarnessRouter in one week; Archon v0.11.0 (workflow packs, event-triggered workflows, forge plugins); four new arXiv self-improving-harness entries (Harness-Zero, RRSI, Agensh, Growing Harness); a multi-generation-improvement verifier distinct from an RSI claim (ruvnet/metaharness); a model-score-vs-agent-score benchmark methodology (HDS6/Apodex, @akshay_pachaar); a context-isolated adversarial reviewer (milosCvetkovicDev); a completion-recovery fix and a harness-thinness design rule (dmlguq456/hearting; ikangai.com)docs/27-loop-contract.md— "write the done condition, not the prompt" (kachar136); huvii174's coded per-item close (loop-close.mjs, v0.18.2/0.18.3); a three-way closure-eligibility gate (eugenelim); a prefilter self-match fix (InjayTseng); Kanevry v5.3.0's PSA-006 gate-verdict-without-receipts claim class and process-group managementdocs/17-failure-patterns.md— a real capitalized-Delete-bypasses-lowercased-gate instance (kachar136); State rot's multi-loop shared-STATE.md anti-pattern (cobusgreyling); a quantified stopping-target lesson (steipete/OpenClaw); METR's live per-action monitor defeated by self-approvaldocs/04-verification.md— mechanically-enforced per-provider read-only reviewers (navels/neal); plan-stage echo-chamber gating (claudex-loop); a multi-model-refute practitioner quote (kachar136); a continue-on-unverifiable-check design tension flagged (eugenelim); a repair-round dispatch-assertion fix (eugenelim)docs/19-mcp-security.md— an eval harness's own-connector leakage case study (edonadei/caliper); MCP tool-selection-at-scale reliability gap (kachar136/Orama)docs/23-fleet-engineering.md— a synchronous-SQLite-at-scale lesson (steipete); a named 20+-agent coordination threshold (explainx.ai/Paperclip)docs/28-routines.md,docs/29-background-agents.md— NanoClaw and durable scheduled automations (Untrivial-ai) as Routines comparators; Microsoft Autopilot-on-OpenClaw confirmed (steipete); Copilot Agent 365 identity (The New Stack); checkpoint resumability with an explicit retention rule (the-open-engine/zeroshot)docs/31-claude-tag.md— @bcherny's own Tag usage (>50% of PRs) with four concrete loop-spec promptsdocs/21-context-vs-loop-engineering.md— a late-Sep LinkedIn/X convergence wave incl. a real production loop (Halldor Fannar, NVIDIA) and an independently-arrived-at harness/loop/graph diagnostic split (Steve Nouri)docs/33-agent-security-hardening.md— process-vs-path isolation-testing caveat (loopx-project); a Docker-socket bind-mount self-disclosure (daonhan/ralph); a named network-perimeter-bypass incident (The New Stack)docs/16-memory-patterns.md— self-reported vs. host-observed subagent activity (loopx-project/loopx)docs/11-cost-control.md,docs/07-subagents.md— Opus 5.5's forked-subagent cache inheritance and mid-session effort switching (Anthropic, unconfirmed against the cache-invalidation reference page)docs/25-long-running-agents.md— an 8-day self-improving research loop (AIDE²)docs/26-factory-model.md— Anthropic's own official Code Modernization plugin (PROVEN/PARTLY PROVEN/NOT PROVEN certificates, blast-radius-tiered stopping)docs/32-reading-list.md— added "Loop Engineering" (Halldor Fannar) to Loops in ProductionSOURCES.md— 7 new tracked sources (HarnessRouter, mvschwarz/openrig, DrSeedon/orchestra, cocodedk/loop-engineering, ujconsulting/claudex-loop, @its_meseba, Halldor Fannar)KB_GAPS.md— two new "tracked but never deep-read" gaps logged (holny's §7/§8 content; milosCvetkovicDev's full production book), plus a same-URL metadata discrepancy flagged forcobusgreyling/loop-engineering
[3.7.3] — 20260921 19:59¶
Added¶
Changed¶
docs/24-harness-patterns.md— Brownfield Agentic Engineering pattern (Osmani); AWS Strands Harness, Writer Agent Harness, True Forge, and a second LangChain data point added to the quantified harness>model corpus; Teleport CTF-prompt worked example under the removal test; Meta-Harness (Stanford IRIS Lab), RobustSGPO, PRISM, and a second AHE-style ablation added to the self-improving-harnesses cluster; claude-uds-bridge added to cross-model division of labor; claude-deep-loop v1.24.1, hearting interim manifests, and a new mechanical drift-detection section (loop-sync) addeddocs/26-factory-model.md— Warp's "Wilson" factory (2,000 PRs/month) added to Named Factory Deployments; Anthropic's own CI-scaling and "When AI Builds Itself" figures added as cross-provider corroborationdocs/04-verification.md— a second high-volume verification-is-the-constraint case study (pstack); a verification-chain demotion-pass fix (affaan-m/ecc)docs/16-memory-patterns.md— loopx recall-filter receipts; eugenelim's versioned work-item-capture contract added to Pattern Idocs/25-long-running-agents.md— Stellar Colosseum (math/TCS research harness) and a cross-vendor ChatGPT Astra computer-use example added to Real-World Scaledocs/33-agent-security-hardening.md— Perplexity CobbleDB build-vs-operate blast_radius example; new Maker/Verifier Process Isolation section (zeroshot)docs/23-fleet-engineering.md— ruflo ADR-385 steerable-swarm update; new GitHub Copilot Rust rewrite case study alongside Bundocs/37-session-architecture.md— Zed Delta added as an industry data point on handoffs carrying conversation, not just diffsdocs/13-context-management.md— Simon Willison's compaction-unrecoverability findingdocs/27-loop-contract.md— huvii174 v0.16.0's deterministic verdict-enum writerdocs/29-background-agents.md— @bcherny's "Projects" description added, flagged unconfirmed against Coworkdocs/05-claude-md.md— new AGENTS.md fallback section (v2.1.277)docs/17-failure-patterns.md— new Industry Infrastructure section (SAFE / Nvidia)SOURCES.md— 11 new tracked sources; ruvnet, Hanako, Akshay Pachaar, and the GitHub harness-search rows updated with 2026-09-21 findingsKB_GAPS.md— two new gaps logged (MCP-as-skill-surface coverage; an unsourced autonomy-limit stat pair)
[3.7.2] — 20260921 19:06¶
The on-demand launcher could not authenticate, because it reproduced the plist rather than the
launchd environment. bash scripts/run-loop-news-now.sh died on every attempt with
Not logged in · Please run /login while the same CLI worked interactively. The login was never the
problem. PATCH — a bug fix in a script plus the doc sync it forces.
Changed¶
scripts/run-loop-news-now.shnow suppliesUSERandLOGNAME, derived fromid -un— the same source launchd derives them from — in addition to the plist'sPATH/HOME. The plist is not the whole environment: launchd synthesizes variables a job never declares. Measured with a throwaway GUI agent whoseEnvironmentVariablesset onlyHOMEandPATH; it receivedLOGNAME SHELL TMPDIR USER SSH_AUTH_SOCK XPC_FLAGS XPC_SERVICE_NAME OSLogRateLimit PWD SHLVLon top. Replaying the plist alone built an environment stricter than launchd's, so the launcher failed where a scheduled run succeeded — the exact inverse of the interactive-PATH-superset problem the launcher exists to prevent.- Root cause, bisected with a negative control:
PATH/HOME/TERMalone →Not logged in;+USER→ works;+LOGNAMEalone → stillNot logged in. WithoutUSERthe CLI cannot reach its Keychain credential. The symptom misdirects: it reads as an expired session, so the natural fix is to re-authenticate, which changes nothing. Re-login was tried first here and did not help. - Three "identical by construction" claims corrected —
CLAUDE.md,scripts/SCHEDULING.mdandRESUME.mdall asserted the two modes ran in the same environment. They now say closely reproduced, not identical, and name what is deliberately not reproduced (SSH_AUTH_SOCK,XPC_*,OSLogRateLimit,PWD,SHLVL,SHELL) and why. An overclaim of fidelity is what makes "it worked when I ran it by hand" read as evidence about the scheduled path when it is not. --checknow printsUSER/LOGNAMEand says they come fromid -unrather than the plist, so the command that exists to show the run environment actually shows it.
Added¶
docs/09gains Reproducing the scheduled environment: the plist is not the whole environment — the measured variable list, the bisect table with its negative control, and why a launcher should state its fidelity gap instead of claiming there is none.docs/17gains a Reproduction stricter than the original row: a reproduction built from a job's own declaration can be narrower than production, and the resulting failure misdirects.
Notes¶
- Not fixed, and not a regression: the launchd-internal variables remain unreproducible from outside launchd. The launcher now lists them rather than pretending otherwise.
scripts/com.luca.loop-news.plistis unchanged. It correctly does not declareUSER— launchd provides it. The bug was in the launcher's assumption, not the record.
[3.7.1] — 20260921 16:16¶
D4 resolved: the freshness watchdog becomes a 14-day staleness nudge, single-homed. Once runs
went on demand in [3.7.0], tracker-watchdog stopped measuring health and started measuring how
recently anyone chose to sweep — it failed ten consecutive scheduled runs, 20260912–20260921.
That is the Notification fatigue pattern this KB documents in docs/17, running in our own CI.
PATCH on both tables: no doc file created, removed or renamed, and no new capability — a CI
threshold plus a documentation sync.
Changed¶
MAX_AGE_HOURS48 → 336 (14 days), and the number now has one home:scripts/check-digest-freshness.sh. The workflow'senv:block is deleted, not set to 336. The backlog's "one-line change" framing of option (b) was wrong and is corrected in place:scripts/run-loop-news-now.sh --statuscalls the script with no env, so editing only the workflow would have left the local status command judging against 48 while CI judged against 336 — the same two-homes drift that got option (d) rejected, merely relocated. The env var remains an honoured override at every call site.- The STALE message no longer misdirects. It told the operator to run
launchctl print gui/$(id -u)/com.luca.loop-news— a LaunchAgent deliberately disabled since[3.7.0]. At 48h the alarm fired constantly and wrong advice was cheap; at 336h it fires fortnightly, so every firing must be actionable on its own. It now namesrun-loop-news-now.shfirst and reaches forlaunchctlonly if--statusreports SCHEDULED. Six survey agents read those lines and cleared them; only the adjudicating pass caught it. scripts/SCHEDULING.mdcarries the revert step in both directions — switching back to scheduled must set 336 → 48 in the same commit, or (b) quietly becomes (a)'s blind spot, exactly asD4warned. It also records that a red run under ON DEMAND is a true positive whose fix is a sweep, not a wider threshold; widening until a check stops firing is this repo's "a check that cannot tell must fail" rule broken in the other direction.- Docs synced to on-demand runs —
docs/34's tracker card still readTrigger: Daily cron (launchd), 05:00 localandCadence: One run per day, andREADME.mdhad the same shape in three places. This drift dated from[3.7.0]and was not caused byD4.docs/17's Notification fatigue row gains this repo as its own worked example.
Added¶
A15(backlog §5, open) — teach--statusto warn whenMAX_AGE_HOURSdisagrees with the live mode. Deferred deliberately: new code on a rarely-exercised path, and single-homing the number removed the larger half of the risk.A16(backlog §5, open) — nothing off-machine distinguishes "no sweep ran" from "a sweep ran and published nothing". Verified while reviewingD4:run-loop-news.shpushes only on success andTEMP_BRANCHcheckpoints stay local, so a failed run leaves zero off-machine trace. Under the schedule the digest-age check covered this because a run was due; on demand it cannot.D4fixed the threshold, not this.
Notes¶
- Verified in real CI, not only locally: the branch push fired
tracker-watchdog, which reportedFRESH: newest digest entry is 300h old (limit 336h)with noenv:override in the log (run35624147361). - A count in this release's own first commit message is wrong and stands uncorrected as history: it says "11 consecutive red scheduled runs since 20260912". There are 11 consecutive scheduled failures, but they span 20260911–20260921; the 20260911 red was a true positive under the schedule and is how the pause was noticed at all. Only ten belong to the on-demand era.
[3.7.0] — 20260912 08:49¶
The tracker is now run on demand rather than on a schedule — decided by the user 20260912. The
schedule is not removed: both modes are supported and switching is two commands either way.
MINOR on the [3.5.0] precedent — a new runnable capability, where the project's own tier
table is doc-scoped so the global new-capability rule governs.
Added¶
scripts/run-loop-news-now.sh— runs a sweep in the exact environment a scheduled run would have had. It readsPATH,HOME,WorkingDirectory, the log paths and the invocation out ofcom.luca.loop-news.plistat run time and starts the wrapper underenv -iwith exactly those, so the two modes are identical by construction, not by intent. Output isteed: the bytes reachinglogs/launchd.logare what launchd would have captured, and you also watch it live. Exit codes pass through unchanged.--status— which mode is live, whether the agent is loaded or disabled, the newest publish and day log, and digest freshness. Its absence is why a deliberate pause read as a three-day outage on 20260912; it asks launchd and git directly instead of inferring from a stale digest.--check— prints the resolved environment and runs nothing, naming what it does not prove.
Why a launcher rather than "just run the wrapper". An interactive shell has sourced your
profile, so its PATH is a superset of the recorded one; a run that succeeds in a terminal
can fail on the scheduled path, and you would not find out until you switched. This repo already
lost eight weeks to a binary-resolution bug of that shape.
Verified, not asserted. A probe driven through the launcher reproduced the recorded PATH,
HOME and working directory exactly; an exported interactive variable did not survive
env -i; the probe's output landed in the log file as well as on screen; exit codes 0, 3, 6 and
7 propagated unchanged through the tee; a missing plist failed closed at 2. That probe caught
a real flaw — the launcher sanity-checked $HERE/run-loop-news.sh while exec'ing whatever
ProgramArguments names, a guard checking something other than the thing it runs. Fixed to check
the recorded target.
Changed¶
scripts/com.luca.loop-news.plistis now the single definition of the run environment, not merely a schedule. Both modes read it; deleting it breaks on-demand runs too. Stated in the file.scripts/SCHEDULING.mdrewritten around the two modes, with symmetric switch procedures in both directions — including theenable-before-bootstraptrap that cost a diagnosis on 20260912, where abootstrapalone inherits the disabled flag and the job silently stays dead.CLAUDE.mdandRESUME.mdso a cold read learns the active mode without digging.
Known limits, recorded rather than solved¶
tracker-watchdogno longer measures health under on-demand runs. It fails when the newest digest is older than 48h, which under a schedule meant a run had silently failed. On demand it measures when you last chose to run a sweep, so it goes red within two days of any pause and stays red — the Notification Fatigue pattern this KB documents indocs/17. Left unchanged deliberately: it is CI config on a public repo, and restoring the schedule makes it correct again with no edit. Options are inRESUME.md§1.
[3.6.3] — 20260909 05:02¶
Added¶
- Loop news run 2026-09-09: OpenAI's own primary-source "harness engineering" case study and
Codex-agent-loop internals (docs/24, docs/02); an eleven-harness academic survey (arXiv
2609.00006); a distinct, disambiguated Harness-Bench (arXiv 2605.27922); Anthropic's
commerce-agents skills-vs-subagents heuristic (docs/03); Google Cloud's first official-vendor
"loop engineering" social post with a live corroborating codelab (docs/21); a three-era
harness-evolution periodization (docs/22); recency-weighted loop-maturity scoring (docs/20);
new failure-pattern rows — verifier theater's unreachable-flag variant, Guard that cannot
fail, Missing value gate (docs/17); OAuth2 machine-to-machine auth + a cache-path traversal
fix, symlink-vs-
@importconfinement, and a cryptographic witness-manifest check (docs/33); two review-discipline lessons on trusting a sustained finding or a reviewer's own fix (docs/04); a forced-interleaving test for file-lock races (docs/38); the Anthropic multi-agent report's conformity sub-finding (docs/37); Claude Tag's human-gated Friday instruction-improvement routine (docs/31); a Stop-hook self-cap rule (docs/12); Context Hub (chub) as a documentation-freshness tool distinct from state-persistence patterns (docs/16); stateless MCP's audit-advantage framing (docs/19); org-wide fleet-scale adoption evidence (docs/23).
Changed¶
- Loop news run 2026-09-09: corrected two claims a source summary overstated before they
reached a doc — Cursor's "$60 billion bet" is a reported acquisition price, not a valuation
(docs/24), and OpenAI's Sherwin Wu's "10-20 parallel agents" describes his engineering org,
not his personal practice (docs/23). Upgraded the QM citation in docs/24 from a secondhand,
caveated panel recap to QM's own primary source, correcting "built on OpenClaw" to a shared
Pi/OpenCode/Codex/Claude Code core. Swapped
docs/32's Martin Dilger "Never Argue" entry for OpenAI's harness-engineering post (Harness Design & Architecture group was at its 5-entry cap).
[3.6.2] — 20260908 07:30¶
Backlog A14, the last open item, opened by A13's own review: the pre-flight and
failure-path guards each re-implemented "has origin/main gained one of our commits since
BASE_SHA?" as a bare git log "$BASE..$NEW" | grep -qE. That omits the ancestry precondition, so
over a rewritten history the range becomes "the new history minus the old one" and a stray matching
subject reads as we published. PATCH — no new doc file and no new script; the tier table's
new-capability rule does not fire.
Fixed¶
- One implementation of the delta question.
published_state()wrapsscripts/assert-published.sh, and all four call sites go through it — pre-flight guard, failure-path guard, the failure path's re-check, and the post-run assertion. No baregit log … | grep -qEsurvives anywhere in the wrapper. - The two guards' conservative directions are opposite, which the backlog item did not anticipate, so "cannot tell" could not take the same branch in both. A guard deciding skip the expensive stage? is conservative when it declines to skip; a guard deciding retry the push? is conservative when it declines to retry. Pre-flight cannot-tell marks the attempt failed rather than starting Stage B; the failure path re-reads once after the loop's own backoff and then exits 7 rather than retrying. Retrying the check is not retrying the push, so the safety property is untouched while a four-second network blip no longer costs the day.
- A fabricated-success path this change itself introduced, found by adversarial review. Both
caseblocks enumerated only 0 and 2 and swept everything else into a*)arm whose comment said "checked cleanly". Reproduced: with the script absent,bashreturns 127 and took that arm — so on the exact case the failure-path guard exists for (Stage B pushed, then the attempt was marked failed) the wrapper would have retried and committed the digest twice.REPO_ROOTis the primary checkout, which this repo's own rules warn other agents mutate mid-run. Both arms now name1)explicitly and fail closed on anything else. - The exit-7 justification was false. It claimed "the 48h freshness watchdog pages for it".
Measured:
check-digest-freshness.shruns at 09:00 UTC againstMAX_AGE_HOURS48 while the tracker fires 04:00–05:00 UTC, so a single lost day reads 29h and prints FRESH. Two consecutive misses (53h) are needed. The comment now states the real backstop and says the branch takes a possibly silent lost day knowingly — the same defect class as the corrections in[3.6.1]'s follow-up PR, in the one comment a future agent reads before relaxing the trade-off. - All four deterministic-failure markers now reset per attempt, not two.
CREDIT_BALANCEandUNKNOWN_COMMANDwere initialised once, globally, whilerun_claude()sets them by grepping the whole transcript — so a marker string in a successful stage's transcript could exit the run telling the operator to wait for a quota reset that never happened. Diagnostic accuracy, not a safety fix, but A14's new pre-flight path is another way to reach the wrong message.
Added¶
scripts/verify-publish-guard.shgrows to 21 checks. Case 10 pins one home for the delta question and matches the pipeline shape rather than one variable name — a reimplementation that copied the regex into a local walked past the name-pinned version — and strips comments first, since the file's own design-rationale comment quotes the bad pipeline verbatim and made the naive regex report a hit on a clean tree. Case 11 pins both cannot-tell branches; case 11b pins that both catch-all arms fail closed. Every one of those three was added because a mutant survived: deletingexit 7, deletingok=0, and restoring the 127 hole each left the harness fully green.- An exit-code table in
scripts/SCHEDULING.md, where the operator actually reads them — none existed anywhere.6and7are called out as the two that need someone to look at the repository rather than just re-run, and as pinning nobody off-machine inside 48h.
Changed¶
docs/09-headless-mode.mdpoint 4 described the retry guard as a two-way decision. It now carries the generalisable part: the question has three answers, and wiring both guards to one "cannot tell" reflex necessarily makes one of them wrong.scripts/run-loop-news.sh's header carries the full exit-code list;CLAUDE.md's open-work note records an empty backlog;RESUME.mdreplaced;A14struck.
[3.6.1] — 20260908 05:01¶
Loop news run 2026-09-08 04:00 UTC — 92 new findings after dedup (24 already covered under
differently-versioned arXiv URLs, no doc changes needed for those), 0 new docs, 8 existing docs
updated. No tag/release — this is a pipeline-cut version; see CLAUDE.md § Releases for the
manual-backfill convention.
Changed¶
docs/26-factory-model.md— Stripe's "Kai" internal agent platform added as a Named Factory Deployment (10,000+ weekly users, 86% adoption, "Projects" governance layer, a near-incident naming identity management and load-shedding as the fix)docs/11-cost-control.md— extended the OpenAI research-acceleration telemetry with The New Stack's follow-on reporting on the same source (Epoch AI task taxonomy, a July 20 outage that took OpenAI's training container service offline, an Aug 7 Astra safety-threshold access cut, the March 2028 automated-researcher goal)docs/24-harness-patterns.md— new "Agent Tool-Choice Bias" section (5,292-session tool-selection study: brand mentions vs. actual selection, per-agent web-search behavior); Task-Shaped DAG Orchestration extended with gate-first fan-out + capture-once/inject-many (~7.8x redundant-read savings); Self-Improving Harnesses extended with the YC panel (Prime Agent /refine, OpenJarvis, QM's 50-agent OpenClaw fleet)docs/04-verification.md— Verifier Integrity extended with rb-lite's no-shell reviewer enforcement and advisory-only skeptic (cannot gate a review round, only inform one)docs/10-fan-out.md— added a git-ref-based distributed loop lock for cross-worktree/cross-clone coordination (deterministic ref-sort race resolution, local-only degrade mode)docs/33-agent-security-hardening.md— added a session-start memory-quarantine gate against poisoned lesson files (flags instructions, credentials/URLs, or missing provenance)docs/16-memory-patterns.md— cross-referenced the new memory-quarantine gate in docs/33docs/32-reading-list.md— added Claire Vo's "How I AI: GPT-6 Astra is a banger..." to Loops in ProductionKB_GAPS.md— logged the GitHub-search buzzword-genre finding for "loop engineering"/"harness engineering" (19 of 30 newly-surfaced repos this run were thin copycats); logged a source-quality flag on Hanako's X account (now primarily promoting a crypto token) and a LinkedIn search SSO-wall stop from this run's Stage ASOURCES.md— flagged the Hanako source-quality concern inline; no new standing sources added (all 28 new candidate GitHub repos scored below the deep-read bar)
[3.6.0] — 20260907 10:18¶
Backlog A13, the last open item: the wrapper judged a run by whether claude -p exited 0.
That is true of a genuine publish and equally true of a session whose skill-internal guard aborted
— Phase 5c's build gate and Phase 5d's digest guard exit 1 the agent's own bash block, not the
process — so a run that published nothing still logged success. This is the eight-week-outage
shape, and closing it empties §3, §4 and §5 of the backlog. MINOR on the precedent [3.5.0]
set: the project's own tier table is doc-scoped, so the global new-capability rule governs.
Added¶
scripts/assert-published.sh— re-fetchesorigin/mainand requires a commit matchingOUR_COMMIT_REGEXinBASE_SHA..origin/main. Exit 0 published / 1 checked, no match / 2 could not check; callers treat 1 and 2 alike, because a check that cannot tell has confirmed nothing. Every cannot-tell path exits 2: bad arguments (wrong count, or any empty one), a non-positive-integer retry count, a non-git target, fetch failure after a bounded retry, an unresolvableorigin/main, aBASE_SHAthat is not an ancestor of the neworigin/main, agit logfailure, and a grep that itself fails on a malformed regex. An empty regex is refused up front —grep -qE ""matches any input, so it would have turned the whole script intoexit 0on a run that published nothing. No|| true, no empty default, no stale-ref fallback anywhere in it.scripts/verify-publish-guard.sh— 18 checks proving that script, invoked as it ships rather than as a copy pasted into the harness. Two clones of one bare origin, so "committed but never pushed" is constructable instead of vacuous. A sanity case that must read no-publish before any passing case is believed; both healthy shapes; the main-moved-but-not-by-us twin (this wrapper's older bug class); eight cannot-tell cases, three of which assert which guard fired and not merely that one did; and two static cases pinning the harness's regex against the wrapper's and the call site's arguments, ordering and its terminatingexit 6. Mutation-tested across two rounds: 12 mutants, 11 killed, 1 equivalent (disabling thegit logfailure branch changes no observable behaviour, because reaching it now requires a repository wheremerge-basesucceeds andlogdoes not — kept as defence in depth and recorded as uncovered, not claimed as tested).
Fixed¶
scripts/run-loop-news.shno longer treats Stage B's exit status as proof of publication. The call sits between the retry loop's failure exit and the artifact-retirement block, and above the run-complete line so that line is never printed and then contradicted. The placement is load-bearing:ARTIFACT_CONSUMED=1also disarmscleanup()'s resume-preservation copy offindings.json, so asserting after retirement would destroy Stage A's ~23-minute search on the exact failure it detects. New exit code 6 (0/1/3/4/5 were taken), and the verdict line reaches the day log on both paths, so a healthy run leaves positive evidence the check ran.- The first harness passed over an untested fail-closed branch, and the mutation pass is what
found it. Reverting
assert-published.sh'srev-parseguard to the unchecked form left all 14 checks green: no case madefetchsucceed whileorigin/mainstayed unresolvable, so an emptyNEW_MAIN_SHA— which collapses theBASE_SHA..NEW_MAIN_SHArange intoBASE_SHA..HEAD, silently asking about the local branch — was never exercised. Case 5c reproduces it with a single-branch clone (fetch origin mainsucceeds and writes FETCH_HEAD; the configured refspec never createsrefs/remotes/origin/main) and asserts its own fetch succeeded, so a broken fixture fails loudly rather than degrading into a duplicate of case 5. -
docs/09-headless-mode.mdsaid "The wrapper still judges success by exit code." True when written, falsified by this commit. Replaced with a point 5 covering the whole pattern — why a session's exit code proves nothing, why the check must precede resume-state teardown, and why "fail loudly" is only as loud as its channel. -
The first version of this check reported PUBLISHED on a rewritten history — found by adversarial review, reproduced against the shipped file.
git log A..Brequires only that both objects exist, not that A is an ancestor of B; after a force-pushBASE_SHAis normally still a loose object locally (git keeps unreachable objects forgc.pruneExpire, two weeks by default), so the range stopped meaning "what this run added" and started meaning "the whole new history minus the old one". One strayfeat: loop news runsubject in it yielded exit 0 on a run that published nothing. The code comment, this changelog andRESUME.mdall claimed that case exited 2, which is worse than the bug: the docs steered the next reader away from the hole. Fixed with amerge-base --is-ancestorguard, which cannot false-alarm —BASE_SHAis only everrev-parse origin/mainor a forward-rebase to a laterorigin/main. - Adding that guard silently un-covered two checks the first mutation round had killed. With
several guards able to return 2, a mutant that removes one is masked by the next, exits 2 anyway
and survives; the unchecked-
rev-parsemutant went from KILLED to SURVIVED with no test touched. The three cannot-tell cases now assert the diagnostic as well as the code. Case 6b's label was wrong too — it said "git log fails" when the ancestry guard reaches an unresolvable base first; observing it, rather than trusting the label, is what showed that. - Deleting the wrapper's
exit 6left the harness fully green. The call-site check greped four anchors around the call but never that its else branch terminates, and the wrapper runsset -uo pipefailwith-edeliberately absent, so the fall-through to "Run complete" and artifact retirement was silent — the outage shape reintroduced under a passing test. A fifth anchor now pinsexit 6between the call and the run-complete line. scripts/run-loop-news.sh'scleanup()comment claimed no test covers the wrapper. Now partly false, so it states precisely what the harness does and does not reach: it never executescleanup(), so the exit-6 preservation path is still read from the code, not run.
Changed¶
.github/workflows/guards.yml— both guard harnesses now run in CI on any PR or push touchingscripts/**or.claude/skills/**. Neither was invoked by anything before this;verify-digest-guard.shhad that gap from the day it shipped. A proof on disk that nobody executes decays silently the moment someone edits what it proves — the same defect the guards themselves exist to catch, one level up. The skills path is in the filter becauseverify-digest-guard.sh's subject is Phase 5d ofintegrate-loop-news/SKILL.md, which can be falsified without touchingscripts/at all.CLAUDE.md's open-work note was rewritten, not deleted. The note met its own stated delete condition (every tier empty), but deleting it outright would have left the backlog andKB_GAPS.mdunreachable from the entry point — the defect[3.5.0]had just fixed forRESUME.md. Two map rows added for the new scripts. The note namesA14rather than claiming an empty backlog — see below — and its own delete condition now says to replace it with a pointer rather than remove it.
Corrected 20260908: this bullet originally read "is now a one-line pointer" and "162 → 159
lines". Both were true of round 1 and false by the time the PR merged — round 2 rewrote the note
again to name A14, making it ten lines, and the file was 164 lines at merge. The prose
describing an edit was not re-read after the edit changed. A line count in a changelog goes stale
the moment the file is touched again, so it is dropped here rather than restated.
- docs/17-failure-patterns.md's silent-scheduler-death row now names the implementation of
the fix it prescribes; LOOP_ENGINEERING.md row 9 gained the wrapper shape;
RESUME.md replaced; A13 struck in the backlog with a status paragraph recording that
no items remain.
A14opened in the backlog's §3.A13's review found the wrapper's two sibling guards — the pre-flight "already published, skipping" check and the failure-path "failed AFTER publishing" check — sharing the same non-ancestor hazard, since both grep the same unguarded delta. It is not a fabricated-success path: both take the conservative action on a match, andassert-published.shthen exits 2 on the same delta, so the run exits 6 rather than reporting success. The residual is a misdiagnosis in the log. Recorded rather than folded in, because it changes unattended retry-path behaviour — the same reasonA13was recorded rather than done insideC4.
Known limits, recorded rather than solved¶
notify()is still desktop-only — an osascript popup plus a gitignored day log, both on the machine that failed. A caught non-publish stays invisible off-machine for up to 48h untilcheck-digest-freshness.shpages STALE.- The
git logfailure branch inassert-published.shis unreachable in practice and is therefore uncovered: reaching it needs a repository wheremerge-basesucceeds andlogdoes not. Kept as defence in depth, recorded as an equivalent mutant, not claimed as tested. - A transient network failure immediately after a genuine push now raises a false alarm. The three-try fetch retry reduces it; nothing eliminates it, and nothing should — the alternative is a stale-ref fallback, which is the class this change closes.
[3.5.0] — 20260907 09:25¶
A handover audit before clearing the session: a fresh agent, started cold from CLAUDE.md with no
context, tried to pick up the one open item and reported what it could not reach. Eight gaps, all
verified before acting. MINOR rather than PATCH because it adds a runnable check to the repo —
the project's own tier table is doc-scoped, so the global "new capability" rule governs.
Added¶
scripts/verify-digest-guard.sh— proves Phase 5d's digest guard fires, and only when it should.v3.4.2claimed that guard was "proven, not asserted"; the proof lived in a throwaway scratch repo and reached the repo only as prose in a digest entry, whichRESUME.mdlesson 6 ("anything a repo file instructs must live in the repo") forbids. Six cases: a sanity case that must abort before any passing case is believed, both healthy shapes (with and without a Phase 4d checkpoint), both broken shapes, and the unscoped guard failing case 4 to show the-- LOOP_ENGINEERING_NEWS.mdscoping is not cosmetic. Mutation-tested: reverting the guard to its unscoped form fails case 4a, and moving it before the--softreset fails case 1b — each mutant reproducing one of the two real bugs found inv3.4.2.
Fixed¶
RESUME.mdwas unreachable from the project entry point. Nothing inCLAUDE.mdor the backlog named it, so a session following the documented chain would never open it — losing the concurrency note and the three lessons most relevant toA13(guard placement, harness proof, line-number drift). Now a row in the repository map.CLAUDE.mdclaimed "One item is open" whileRESUME.mdnamed three more, two of them live inKB_GAPS.md§ Active Gaps. Now "one backlog item", pointing at the rest — an empty §4 and §5 is not an empty repo.A13omitted the ordering constraint that makes it unsafe to get wrong. The assertion must precede the artifact-retirement block:ARTIFACT_CONSUMED=1disarmscleanup()'s resume-preservation branch, so asserting after it would destroy Stage A's resume state on the exact failure being detected — against the repo's own every expensive stage must be resumable rule. Added, with the caveat thatnotify()is still desktop-only so "fail loudly" remains invisible off-machine for up to 48h, and with four cases to cover plus the harness to copy.scripts/run-loop-news.shcredited a test that does not exist — "(Caught by a test asserting the files on disk…)".git ls-filesfinds no test for the wrapper. Corrected to say it was a hand check, and to point atA13and the new harness.- A stale "found but not fixed" note in the backlog's retained step-10 status block — the
README source-type drift it describes was closed by
85f69f0, andkb-structure-check.sh§5 now reports "67 rows, 7 types, README agrees". It survived because it sits outside §3/§4/§5, where no tier-emptiness check would ever have cleared it. C4's closure described a fixed defect in the present tense ("two consecutive quiet days page a healthy tracker"). Untrue sincev3.4.2.- The live LaunchAgent still carried the comment
H14deleted.~/Library/LaunchAgents/held05:00 local (IST/UTC+1 summer) = 04:00 UTC— the operator-facing copy of the sentence that goes stale on 2026-10-25. Synced perSCHEDULING.md; comment-only,Hourstill 5, job still loaded, no reload needed. - A leftover branch (
20260907_0703-readme-source-count-fix) existed local and remote. Not an ancestor ofmain— squashed into85f69f0— sogit branch --mergedwould never have listed it. Every added line confirmed present onmainbefore deleting both refs.
[3.4.2] — 20260907 08:03¶
Backlog step 13 — C4 and H14, the last two items in §4 and §5. Both are shipped; the
backlog's §4 and §5 are now empty, and one new item (A13) was opened rather than silently done.
Changed¶
- A zero-finding tracker run now commits its digest section instead of discarding it (
C4). Decision taken by the user 20260907: commit the section, cut no release.integrate-loop-newsmandated writing the empty section in Phase 4 and then threw it away in Phase 5a, so a quiet day left no trace. The consequence was sharper than the item was filed as:check-digest-freshness.shreads the newest committed dated header against a 48-hour limit, so two consecutive quiet days paged a healthy tracker as STALE — a false alarm, not just a gap in the record.fetch-loop-newsalso deriveslast_run_datefrom that header, so skipping froze it and widened every later run's search window. The None tier still cuts no version; only the commit/skip decision changed. Phase 5d additionally requires every run to assert its staged diff is non-empty after the--softreset, before committing — a run that finds nothing staged never wrote the section and must fail rather than commit nothing and report success. The placement is the point: an adversarial review of this change caught the assertion sitting in 5b, before any staging, where it would have read empty on every correct quiet-day run and halted it — this repo's own defect class pointed the other way, a check that cannot tell and so fails always. Proven, not asserted (four cases in a scratch repo, with a sanity case confirming the harness could produce both outcomes): a healthy run passes whether or not Phase 4d checkpointed the digest; a run that never wrote the section fails; and the original 5b placement does halt a healthy checkpointed run. The check is also scoped toLOOP_ENGINEERING_NEWS.mdrather than the whole index — unscoped, it passed on a run that staged something else and never wrote the digest, which is the exact failure it exists to catch. The commit keeps thefeat: loop news runsubject prefix thatrun-loop-news.sh'sOUR_COMMIT_REGEXmatches, so the wrapper's double-publish guard still recognises it. - The DST-dependent scheduling comments now state the rule, not one regime's stamp (
H14).StartCalendarIntervalis local time, so the launchd job'sHour: 5is 04:00 UTC under IST and 05:00 UTC under GMT; two comments asserted the summer stamp as fixed and would have gone stale on 2026-10-25.scripts/com.luca.loop-news.plist:15and.github/workflows/tracker-watchdog.yml:15now name both regimes, so they do not expire again at the next changeover in either direction. The schedule itself is unchanged — this was doc accuracy, and the watchdog keeps 4–5h of margin either way. Historical04:00 UTCstamps (this file, the digest headers,run-loop-news.sh:59) record real runs and were deliberately left as written.
Fixed¶
- Three downstream claims that asserted the discarded-section behaviour — this file's release
tier legend (
:10),docs/34's KB-tracker Stop condition row, anddocs/09's retry-guard rationale.plans/split-fetch-loop-news.mdrepeats the superseded claim at:303,:514and:568, all deliberately left: that plan is retired and its own header declares the checklist historical, and rewriting a retired design document to match current behaviour would be worse than leaving it. (The first pass of this change said "one occurrence"; it is three.) - Two backlog rows that this work proved wrong.
C4'sD1half was already shipped — theSKILL.md:300line it cites no longer exists; Phase 5b has mandateddate -u '+%Y%m%d %H:%M'since theCLAUDE.mdUTC rule landed.H14namedCHANGELOG.md:867, which had drifted to the05:00 local (= 04:00 UTC)LaunchAgent bullet in the v2.x notes — a historical entry that stays as written. (Do not re-cite it by line number: every release prepends a section and moves it. It was:867when the row was filed and:1609before this release; grep the text.) Its live surface was two comments, not the two files named —docs/09,docs/34andscripts/SCHEDULING.mdalready said "05:00 local", correct in both regimes. - A contradictory pointer-deletion condition. Backlog §1 said to delete the
CLAUDE.mdpointer when "§3 and §4" are empty;CLAUDE.mdsaid "§4 and §5". Both now read "every tier", which matters immediately — §4 and §5 emptied this pass while §3 gainedA13.
Added¶
A13in the backlog's §3 — the wrapper still logsRun completeon Stage B's exit status.C4makes the artifact assertion possible (every run now commits, so a matching commit in the delta is a reliable post-condition) but does not make it. Recorded rather than done: it changes unattended retry-path behaviour and deserves its own adversarial review. Sharpened by this change's round-2 review: the new Phase 5d guard fails closed only inside the skill — itsexit 1ends that bash block but does not makeclaude -pexit non-zero, so the wrapper still logsRun completeand nothing publishes. Pre-existing for every in-skill failure, but C4 adds a new way to reach it, which makesA13the guard's missing half rather than a tidy-up.- A digest entry explaining that empty sections are now expected on quiet days — so a future reader does not "fix" them or read them as a broken pipeline.
Housekeeping¶
CLAUDE.md's open-work note slimmed to a pointer (165 → 158 lines, back under the ~150-line guardrail's neighbourhood); its "do not re-spend budget on" list relocated to backlog §1, which is now its one home.RESUME.mdrewritten for the current state, and its duplicated workflow-resume lesson (recorded twice as items 3 and 6) merged.
[3.4.1] — 20260907 07:48¶
Follow-on to 3.4.0. The daily tracker run landed 21 minutes after step 11, appending 76
findings to 12 docs written against the pre-correction text. This reconciles the two.
Fixed¶
- Six defects found by a 15-agent merge-coherence review. Five were self-contained errors
inside one side's own new prose, not merge damage: the tracker compared Antigravity's maximum
nesting depth (10) to Claude Code's overridable default (3), and claimed its on-kill worktree
cleanup closes a gap
docs/03already documents as an automatic periodic sweep. One was ours — thedisableAllHooksexample was offered as a case of "a stricter lower-level value beats a managed one" while illustrating the exact opposite; the live exceptions table does not list it. - Three broken relative links in
plans/, found by a mechanical handoff audit. docs/24said "the removal test above" while that section sits below it.
Changed¶
- Four body joins, where the tracker restated something the sweep had just corrected. The
clearest:
docs/01appended one paragraph carrying two unrelated claims spliced mid-line, one of them Cobus Greyling restating the founding line the sweep had re-attributed (it is Addy Osmani's, not Boris Cherny's). A reader met the same idea twice with a company-scale detour between.docs/11held the only true duplicate-claim pair — the same cache rule stated fifty lines apart, split by an unrelated subsection. - Nine
LOOP_ENGINEERING.mdindex rows resynced. The tracker changed 12 docs and touched the index zero times. Three rows were amended by the synthesiser, which cut an editorializing clause, a figure carrying caveats an index row cannot hold, and an ", ever" the doc itself contradicts in the same list. - Two transcript-only artifacts written to disk before they were lost: the 47 numbered
UNVERIFIABLE claims (
KB_GAPS.mdinstructs a future pass to "drop entry #29, re-run #13/#33/#34/#44" — meaningless without the list) and an index of all 335 C10 findings, whose evidence pack called them "marked, not hidden" when the marking lived only in a workflow transcript. Both now linked from the files that reference them. - README's source table resynced again — the tracker added 7 sources (60 → 67) and
kb-structure-check.sh§5, added hours earlier in3.4.0, caught the drift on its first run after the change. The previous drift of that table went unnoticed for two releases.
Verified, not assumed¶
- Zero regressions across all 14 files both passes touched. History is linear
(
3b02afe's parent is0fd20b4), so the tracker wrote on top of the corrected text rather than beside it — the highest-risk failure mode was absent by construction, not by luck. The three highest-value removals were re-checked as still gone, including the settings-precedence block re-verified word-for-word against the live page. - Two findings were refuted, one of them a proposed deletion of a true claim — the
ultracodeconcurrency exemption is stated verbatim in the official subagents docs. Deleting it would have repeated theV13over-correction this KB has already paid for once. docs/index.mdwas left alone on purpose. It is a 125-line navigation tree (39 bare doc links, 9 descriptors), not a second summary index, so the "update both indexes" rule does not fire when no doc was added or renamed. The synthesiser checked this rather than complying with a wrong premise in its own instructions.
[3.4.1] — 20260907 07:01¶
Loop news run 2026-09-07 06:30 UTC. 76 new findings after dedup (24 of the 100 raw findings
were already integrated by 3.4.0's C7/C10 sweeps). PATCH: no new docs/*.md file, 12
existing docs updated.
Added¶
- docs/24 — token-variance/"startup tax" quantified corpus entry (70-fold spread, R²=0.99, The New Stack); ARC-AGI-3 Milestone Prize #1 harness-light counterpoint alongside the Astra harness-swing material; model-tiering-by-role additions (vibe-dev-plugin's write-denied verifier, hearting's cross-runtime manifest); omnigent's stalled-loop idempotency fix; Pi as a minimal-harness-by-design counterpoint; Anthropic's Model Hardware Standard as a named lab-equipment maker/checker deployment; Demystifying evals harness-definition citation.
- docs/21 — earliest academic source for "Graph Engineering" (35-author arXiv 2608.21156, predating the two cloud-engineering papers already cited); hanakoxbt's per-layer diagnostic test.
- docs/08 — omnigent's "approval valve" refinement to orchestrator-attached subagent cost budgets (silent hard block → ASK with persisted approval).
- docs/07 — cross-vendor subagent comparison (Google Antigravity: 10-level nesting depth, automatic git-worktree cleanup on subagent kill).
- docs/11 — per-turn model routing as a cache-invalidation anti-pattern (Uber Claude Code rollout example); OpenAI's internal research-acceleration telemetry (~$600/day median spend, 3.1:1 agent-to-human ratio, >50% of long tasks still needed human intervention).
- docs/17 — METR's own primary-source account of the OpenAI/Hugging Face incident (~700/1,200 agents coordinating via an unsanctioned message board, HMAC-exploitation scorer attack), closing the KB_GAPS item that flagged this as MindStudio-derivative-only.
- docs/26 — monday.com named factory deployment (Intake & Triage / Knowledge / Incident routines, "Brand Reviewer" maker/checker).
- docs/23 — ccteam cross-vendor fleet-coordination example (7 harnesses, live cost ledger, per-harness quotas).
- docs/32 — vibe-dev-plugin added as the 5th Reference Implementation entry.
- docs/01 — a16z/Acharya and Cobus Greyling framing quotes; Anthropic's "Claude Code guide for startups" as a first-party loop definition source.
- SOURCES.md — 7 new tracked repos (loop-js, InjayTseng, huvii174, milosCvetkovicDev,
andrewcigan/vibe-dev-plugin, dmlguq456/hearting, firstintent/ccteam) and 4 operational
source_updates (New Stack fetch method, X
from:query parenthesization, two GitHub search query refinements).
Changed¶
- docs/16, docs/04 — two bare repo-slug mentions converted to markdown links
(
kb-structure-check.sh§ 4c). - README.md — source-table counts corrected to match
SOURCES.md(github 23→30, total 60→67), closing the driftkb-structure-check.sh§ 5 flagged. - KB_GAPS.md — closed the "METR/OpenAI transcript-tampering report" gap (docs/17 now cites METR's own primary source directly).
[3.4.0] — 20260907 06:40¶
Backlog §8 step 11 — both large sweeps, C7 and C10 — plus H1, closed. This was the
biggest unclaimed content risk in the backlog and it is now measured rather than estimated.
MINOR on the same reasoning as 3.3.0: no new docs/*.md file, but new executable capability —
two more mechanical checks in scripts/kb-structure-check.sh.
The two sweeps found different diseases. C10 found the KB stale against a moving platform; C7 found the KB composing plausible content around a real citation instead of reading the cited source. The second is worse, because tracking upstream would never catch it.
Coverage, measured¶
| C10 — changelog | C7 — docs | |
|---|---|---|
| Scope | 3,774 of 3,774 previously-unswept bullets, 24/24 chunks, zero count mismatches | 4,025 previously-unverified lines across 23 docs, 34/34 units, none lost |
| Examined | 520 verified rows | 656 claims, 471 distinct URLs fetched |
| Landed | 29 edits, 18 docs | 150 edits, 25 files |
| Logged not written | 10 KB_GAPS entries |
11 KB_GAPS entries, 47 claims kept as UNVERIFIABLE |
Fixed¶
- Settings precedence was inverted in two docs.
docs/08anddocs/12both ordered sources "later overrides earlier" with managed policy first, making org-wide security policy the weakest input and CLI flags the strongest. The settings reference says the reverse verbatim. Found independently by both sweeps.docs/12now also separates hook entries (which merge) from settings values (highest wins) — a distinction the old text collapsed. docs/12's hook tables were wrong in three places: timeouts (command/http/mcp_tool are 600s, not 60/30/30),PreCompactcan block (v2.1.105), andPermissionRequestcannot block via exit 2 — so a guard written asexit 2there fails open. The last was missed by the sweep and found by its completeness critic.- Fabricated content, removed with the capture-date check every time. A blockquote in
docs/24absent from the cited article and from two Wayback snapshots predating the KB's own commit; a.claude/settings.jsonblock indocs/08whose keys the CLI does not recognise; adocs/34heartbeat schema with zero real fields; the unsourced ">80% of Anthropic engineers" quote indocs/01, whose only traceable origin is an X trending headline. docs/08's?wildcard — verified against Wayback snapshots from 2026-06-03, 07-03 and 08-03 before removal, so the claim is not a current-docs-only absence assertion.docs/26's OpenAI citation dated Jul 2026; the article's byline reads June 25, 2026 (openai.com 403s tocurl; the Wayback snapshot loads).- README's source-type table had drifted from
SOURCES.mdfor two releases (x7→9,html5→9, total 54→60).
Added¶
scripts/kb-structure-check.sh§5 and §6. §5 compares README's source-type table againstSOURCES.md; §6 enforces Part II's version-stamp promise, deriving Part II membership fromLOOP_ENGINEERING.mdrather than hardcoding it. Both promises were previously enforced by nothing, which is why H1 sat open for weeks. Both were proven by negative test — and §6's first regex required lowercase after--, silently missed every camelCase flag, and reported clean over a corpus it had undercounted. That is this repo's defining defect reproduced inside the check built to prevent it, caught only because the check was tested rather than trusted.docs/06gained a SKILL.md frontmatter reference — the doc that owns skills stated not one frontmatter field in 86 lines.- Two evidence packs in
plans/, so ~17M tokens of sweeping is never repeated.
Changed¶
- H1 closed, four docs by stamping and one by refusing to.
docs/05,06,19,35stamped.docs/31deliberately not: Claude Tag is a separate product with no CLI version gate — the official Slack page states no version requirement and the full changelog carries no Claude Tag entry, so a marker would assert a gate that does not exist. - The duplicated
## Session Watchdogsection existed verbatim indocs/25anddocs/33, withdocs/33cross-linkingdocs/25while restating it. Consolidated intodocs/25;docs/33keeps a pointer plus the one detail only it had. - The changelog sweep's own denominator was wrong and this repo recorded it four times. "2,076
bullets across 385 versions" — the real total was 5,132, and
Fixedalone was 2,740, which the same paragraph stated. Corrected in theplans/working documents in place and in the append-only digest by a new entry.
Known limitations, stated because a clean claim would be false¶
- C7's completeness critic returned INCOMPLETE, on a measured density inversion: r = -0.64
between unit size and scrutiny. Units over 160 unverified lines were checked at 10.8 claims per
100 lines; units under 100 lines at 26.5.
docs/24— largest doc, most findings, most high-severity — got the thinnest sampling, and a spot-check inside a range it declared clean found a real defect.34/34 units returnedis reporting, not coverage. - C10 applied a refutation cap: 245 of 335 actionable findings reached triage unrefuted, marked rather than hidden.
- Both are logged in
KB_GAPS.mdwith concrete re-cut plans, not left implicit.
[3.3.0] — 20260906 15:39¶
Backlog §8 step 10 — corpus hygiene, H4–H9, H11 and H12, all eight closed — plus H10, H13
and step 12, found already resolved and never struck off. Cut MINOR rather than PATCH on the
same reasoning as 3.2.0: no new docs/*.md file was created (this repo's table would say PATCH),
but a new executable capability shipped — scripts/kb-structure-check.sh, made a mandatory step of
integrate-loop-news's Phase 4c — and the unattended agent's Bash allowlist widened to run it. A
pipeline-contract change is a capability change the file-count table does not cover.
Added¶
scripts/kb-structure-check.sh— six mechanical structural checks (H11). Orphan check, heading check, duplicate-coverage check, and three bare-citation checks (baregithub.commentions, bare@handles, and a repo slug linked elsewhere in the KB appearing bare). Phase 4c has been mandatory after every run for ~20 runs and caught none ofH5,H7orH8, because it asked for a judgement rather than producing a list. It is a report, not a hard gate — exit 0 for findings, since several legitimate doc shapes trip the heuristics — but it aborts with exit 1 if it cannot run at all (bad repo root, no docs matched) rather than reporting clean. Phase 4c must fix or explicitly waive every hit in the digest.B_ARGSinscripts/run-loop-news.shgrants exactlyBash(bash scripts/kb-structure-check.sh*).KB_GAPSV19 (H12) — the co-resident session-leak security claim, which existed as one sentence in a digest's "deliberately excluded" list and was never logged. A sweep found no CVE, advisory or locatable post; two adjacent reports are recorded as explicitly not this claim. Nothing was added todocs/19ordocs/33— an unverified security claim in a public doc is worse than no claim.
Changed¶
- H4 — pinned the Boris Cherny "write loops" quote to one primary source.
docs/03anddocs/26reproduced the quote with differing wording and no citation;docs/32cited it but mis-called a secondary report (thenewstack.io, itself embedding a third-party tweet) "Primary source ... in his own words";docs/20paraphrases the same remarks in a table row with no citation — the backlog's original H4 row over-called that one a quote reproduction. All four now cite Cherny's own on-stage remarks at Acquired Unplugged, presented by WorkOS (YouTube, 2026-06-02), verified against the video's caption track at ~11:45.docs/03is an instance the backlog's four-location list missed;LOOP_ENGINEERING_NEWS.md's two captures are append-only history and are corrected by a new hand-authored entry rather than rewritten. - H5 —
docs/17now applies the taxonomy it declared. The intro named three root-cause categories and no row was labelled with one; the file had 0##headings across 37 rows and 13 inbound references, so nothing in it could be deep-linked. Rows regrouped under## Underspecification(5),## Capability errors(2) and## Harness errors(31), verified programmatically: all 37 original row lines survive byte-for-byte, the intro verbatim, all 17 URLs preserved. One row added — silent scheduler death, this repo's own eight-week outage, labelled as such. Anchors confirmed generated in the built site. - H6 — eight bare citations linked, and two completeness claims corrected.
openai/symphony(recorded, not rewritten — append-only), three Prithvi Rajasekaran attributions (docs/11,docs/13,docs/24), three indocs/17(Dilger, Ronacher ×2), and@roanbrasilindocs/32. Every URL fetched and every repo confirmed viagh apibefore linking. A five-angle manual sweep had reported the repo-slug angle clean; the new script found five slugs on its first run — of which three were real (jahwag/clem,thirai-classlab/hirai-method,ruvnet/ruflo) and two were already linked a few lines away. The mechanical check is a candidate generator, not a verdict. - H7 —
docs/21's one-way-orphan premise was already false, fixed incidentally by two unrelated commits on two different days. The original grep misled by matching plain-text mentions in the news/changelog symlinks while missing real inline links. Thedocs/13link was added anyway. - H8 —
docs/15substantiated in place rather than merged (user's decision). The stated duplication withdocs/36does not exist — they share no content, anddocs/36citesdocs/15by link as its Planning-phase mechanism. 26 → 79 lines, five headings, every claim verified against Anthropic's live permission-modes page.LOOP_ENGINEERING.mdanddocs/index.mdsynced. - H9 —
integrate-loop-newsPhase 0 now aborts on aschema != 1artifact. The wrapper already asserted it; the skill's abort list did not mention it, and the two can drift. The phase-collision half was checked and found already disambiguated at every live cross-file reference. - H10, H13 and step 12 struck off — resolved by the 20260905 decisions and the
v3.1.9backlog, and simply never marked. Tags and releases reconciled at 65/65;CLAUDE.mdnow says to count them rather than quote a number that goes stale on every release. - Corrected this pass's own output after an adversarial review. Ten defects, all self-inflicted:
a nullglob hole that made the new gate report a clean corpus it never read; an append-only
violation inside a dated run record; a fabricated "16 inbound references" count; a stale
H1list that this pass's ownH8work had falsified; and a waiver whose premise (1 of 27 entries, not ~25) did not survive measurement.
[3.2.0] — 20260906 13:26¶
Backlog §8 step 9 — C5, C9, C11, C12, all four closed. Cut MINOR rather than PATCH: no new
docs/*.md file was created (this repo's table would say PATCH), but fetch-loop-news's fetch
contract changed — the type: x procedure now mandates a second, unfiltered pass — which is a
capability change the file-count table does not cover.
Added¶
docs/16— native auto memory, the section it never had. The doc covered ten loop-external patterns and said nothing about the memory the harness writes for itself, on by default since v2.1.59. Adds what it is, a sourced reference table, the five properties that disqualify it as loop state (agent-curated, machine-local, invisible to subagents, unenforced, unversioned), three issue-tracker failure modes, and the rule: if losing it would break the loop, it goes in the repo. Also closes H1'sdocs/16item — zerov2.1.xmarkers to eight.docs/23— observable F0–F3 per-gate indicators, closing a gap open since Jul 2026. File- existence exit criteria per transition plus a runnablefleet-auditscore, from two files in cobusgreyling/fleet-engineering that nobody had opened — a repodocs/23already cited three times. Explicitly distinguished from this KB's own F0–F3 scale, which shares the four labels and nothing else.docs/11— the reach rule. A checker pays off "only where its reach covers the way the application actually fails" — not by codebase size or task horizon. A boot probe removes nearly all launch failures at ~35% of a full shell's token cost; a full shell costs 2.35× the no-tool baseline. (arXiv 2608.28795.)docs/30— catching an underspecified goal before it propagates. A halting Intent Agent, an EVPI-scored question budget, an industrial ambiguity classifier, and the counterintuitive timing result: goal clarification loses nearly all value after 10% of execution, and asking past mid-trajectory is worse than never asking.docs/04— cross-model pairing is asymmetric, and lineage diversity may not be what buys ensemble gain. Plus two-tier BLOCK/WARN verdicts: never escalate an inference to a hard failure — the split is what lets one verifier be both trustworthy and unattended.docs/28— Anthropic's own maintenance fleet, from a primary source the KB had missed: 388 PRs opened, 180 merged (46%) by named daily routines, a two-day instrument-then-decide verifier, and an outer loop that tunes the routines.docs/07— the/agentswizard removal (v2.1.198);docs/03— the worktree-cleanup loop;docs/24—/ultraplan's removal (v2.1.222) as the other half of the shrink bet: native features also depart.SOURCES.md§ Ruled-out source strategies — why LinkedIn person-search was measured and not adopted.plans/20260906_1259-c11-x-linkedin-baseline.md— the C11 evidence pack.
Changed¶
- Corrected a false claim the KB made in three places.
docs/24,docs/35anddocs/36said ClaudeWarp's/claude-warp-syncretires its components as Claude Code absorbs them. Measured against that project's own changelog: the mechanism is real and complete, has run twice across 65 Claude Code releases, and has retired zero.docs/36's wording is a quotation and stands untouched, with a measured counterweight beside it. The measurement is the better lesson — a retirement mechanism is cheap and retirement is rare. - First-party disclosure on all five docs citing ClaudeWarp (
11,22,24,29,35; canonical note in36). Same maintainer as this KB, figures self-reported — and it ships a skill whose job is to read this repo, so agreement between them is shared origin, never corroboration. That circularity is why ClaudeWarp's R0–R5 was deliberately not added todocs/04. fetch-loop-news— thetype: xprocedure now requires two passes, and the profile pass is unfiltered. The skill already told the run to visit the profile page but keyword-gated that pass too, so untracked vocabulary was invisible on both paths while the line's own parenthetical claimed otherwise. Measured on @bcherny: search 7 posts, profile 4, union 9, overlap 2.SOURCES.md—xstrategy rewritten to match;linkedincontent-row keywords broadened to includeharness engineering.docs/36—/code-reviewcurrency. The v2.1.215 claim that Claude no longer runs it on its own no longer holds: v2.1.246 says Claude "can also start it on its own". No bullet records where autostart was reinstated, so the gap is stated rather than inferred.KB_GAPS.md— three Active Gaps filled; cross-model reviewer pairing annotated as still open after a fourth retry, with a recommendation to stop searching.
Verification¶
Every version claim re-verified against the raw official changelog; every arXiv ID, title, date and
author against the arXiv API; all three GitHub issues via gh api (two are closed as stale, not
fixed). A source reachable only through a rewriting reader proxy was not cited — the publisher
returns 403 to a direct fetch. No binary-derived internals, absolute /Users/... paths or local
machine state were published: this repo is public.
[3.1.9] — 20260906 10:18¶
Changed¶
- Backfilled the 15 tagged versions that had no GitHub release —
v1.1.0,v1.2.0,v2.0.1,v2.0.2,v2.1.1,v2.1.2, andv2.3.1–v2.3.9. Each release body is that version's ownCHANGELOG.mdsection, verbatim, with a note explaining why the tag had no release. All created with--latest=false, solateststill points atv3.1.8. The repo now has 63 tags and 63 releases, 0 missing (it was 53/38 when the gap was logged).
This closes the backfill half of backlog item H10. The other half stands: the pipeline
still creates no releases, by the deliberate decision in CLAUDE.md § Releases — granting an
unattended agent gh release rights on a public repo is a wider permission than the convenience
is worth. The consequence is that every pipeline-cut version re-opens this gap and needs the same
manual backfill, which is the standing cost D2 has yet to settle.
[3.1.8] — 20260906 10:12¶
Closes every remaining P1 and P2 item in the C1b backlog, including the one the entry point
called the largest untouched risk: docs/11's model IDs and pricing.
Fixed¶
docs/11-cost-control.md— the KB's densest concentration of model IDs and prices, swept for the first time. ~60 claims checked against live sources. It came back the cleanest doc this project has audited, not the most defective: all 30 figures in the 6-row pricing table, all four model IDs, context windows, knowledge cutoffs and positioning quotes matched the live Models-overview and Pricing pages exactly, as did the cache-pricing, tokenizer and cache-invalidation blockquotes, the 6-row effort table, the subagent precedence order, and the release-note quotes for v2.1.243/.248/.251/.257. Four precision defects fixed:- the Opus 4.7
xhighdefault was stated as flat fact while citing the API models table.xhighis Claude Code's default; the API default for the same model ishigh. A loop split across Claude Code and the SDK silently runs at two depths and pays two bills. - the "roughly 30% more tokens" tokenizer figure was uncited and sat under a "(Same source.)" pointing at the wrong page. Now cited, with the caveat that page attaches and the 1.0–1.35x range from the Sonnet 5 announcement.
- the Explore-inheritance note omitted the Claude API's Opus cap — the cost-relevant half, and Claude-API-only.
- the model table's "API ID" column lists Haiku 4.5's alias, not its pinned snapshot ID.
- the Opus 4.7
docs/23— a falsified claim-of-absence, this repo's defining defect class. The KB claimed "Neither Bun's post nor Anthropic's names a token or dollar figure." Bun's post names both: "5.9 billion uncached input tokens, 690 million output tokens, and 72 billion cached input token reads — around $165,000 at API pricing." Rewritten around the real figures with the scope stated — that covers the whole 11-day port, not one 64-instance wave. Added the model attribution the backlog asked for ("a pre-release version of Claude Fable 5, a Mythos-class model") and a repricing table; Bun's own token counts on Fable 5's rate card give $165,500, 0.3% off its quoted figure. Which exposes the real lesson: the 72B cached reads are 44% of that bill, more than input and output combined.docs/04— implementer-to-reviewer ratio corrected to 1:2+, with the role separation the source states, now matchingdocs/23. The earlier pass had checked the other copy of the claim.docs/27— the uncited Uber figure, where citing it was necessary but not sufficient. The source says "per person per tool per month", covers Claude Code and Cursor, and sources itself to "reporting in the discourse". An anecdote cannot sit under a heading promising Real Cost Data; heading renamed, quote given exactly, provenance stated.docs/37— a named academic source with no link anywhere in the KB. Found by sweeping every bolded statistic indocs/for a nearby citation (37 hits, 1 real defect). Identified and verified as arXiv 2604.02460 (Tran & Kiela); every element of the KB's sentence was accurate. Now cited — with the caveat a Claude-focused KB owes it: the three model families tested are Qwen3, DeepSeek-R1-Distill-Llama and Gemini 2.5, no Claude model, and the paper reports artifacts in its own budget control.- Repo-wide blockquote audit — 414 blockquote lines across 113 blocks, plus 224 inline quoted
spans. Two defects, both in quoted third-party material:
docs/24truncated a sentence mid-clause: "...do not automatically gain teammate-grade stability." The source ends "...stability, accountability, or sustained judgment."docs/07joined two of Sabrina Ramonov's sentences with an em-dash, and carried a named attribution with no URL.
docs/04— the 93.4% paragraph re-read whole, so the corroboration sentence no longer asserts absolute non-overlap while the sentence before it admits 37 two-tool and 4 three-tool lines. Added the round-cap rationale: the cap is 3 because iterating "past ~3 rounds introduces bugs".
Added¶
CLAUDE.md§ Quoting a source that has since died — the policy question the backlog left open, now decided. A verbatim quote may stand on a dead source only when the citation says it is dead, dates it, and states the wording can no longer be re-verified; link a Wayback snapshot where one exists; never delete a quote merely because its source went away. Applied to the threeorobsonn/claude-harnessquotes; confirmed neither dead repo has a snapshot. ClosesKB_GAPSV10.CLAUDE.md§ Check the whole repo before calling a quote absent — a README-only fetch is not a search. Three counterexamples this pass hit:harness-books' quoted sentence lives in a chapter file with zero README hits;compound-reviewserves onlymaster; andv2.1.243has a git tag and aCHANGELOG.mdentry but no GitHub Release, so the releases API 404s on four perfectly good citations — theV13shape, avoided.- Four earned verification stamps in
docs/03, extending the section-2 convention to Routines, Skills, Connectors/Chrome and Sub-agents. Each was checked before it was stamped, at the CLI version actually installed (2.1.263). Section 6 is conceptual and correctly gets none; section 2's existing 2.1.261 stamp is left exactly as written rather than restamped. KB_GAPS.mdV17 (no primary source for the Uber figure) and V18 (docs/27may credit a repeater — a second explainx article attributes the Loop Contract to Developers Digest).
Changed¶
KB_GAPS.md— V16 settled:docs/04's ~85–90% reviewer non-overlap figure is verbatim in its cited source, traced viagit log -Sto198c302. V10 settled by the new policy.- The backlog and
CLAUDE.md's entry-point note now point at §8 step 9, and record two corrections the work produced:docs/04:257-260rests on a live source, not a dead one, and H6's "only these two bare citations" is disproved by thedocs/37find.
[3.1.7] — 20260906 09:34¶
Fixed¶
scripts/run-loop-news.env.examplewas corrupted in v3.1.6 and no longer parsed as shell. The section added in that release was written through an unquoted heredoc, so backticks around a command name were executed as command substitution and the command's output was injected into the file.bash -nfailed withsyntax error near unexpected token '('.
The live run-loop-news.env was unaffected (written with a quoted heredoc), so the tracker never
broke. But the template's own instruction is to copy it to run-loop-news.env, and the wrapper
sources that file — so anyone following the instruction would have broken their next run.
Rewritten without backticks and verified three ways: bash -n exits 0, zero residual command
output, and a copy-to-.env round trip sources cleanly.
Two lessons, both already rules here and both broken in the same minute: quote the heredoc
delimiter unless you specifically want expansion, and a gate's exit status must be what the
next command reads — the check that would have caught this was run as
bash -n file | head, which reports head's status, so it printed the error and still looked green.
[3.1.6] — 20260906 09:31¶
C1b round two: the coverage the first fact-check pass did not reach. 476 claims checked, 424
correct, 46 fixes applied, 90 of 90 cited URLs opened across docs/03, 04, 27, 29. Four
docs out-checked all fourteen of round 1 — the inverted claim density is fixed.
Fixed¶
- A second fabricated blockquote in
docs/04, presented as verbatim fromwquguru/harness-books; an authenticated code search returns zero hits for its key word. - Fabricated implementation detail —
SCOPE.mdand a "test/check index" that do not exist in the cited repo, describing a Stop-hook mechanism the source's own RFC explicitly rules out. - An unattributed verbatim quote in
docs/04that this KB cites correctly, with a link, in two other files. - Fugu role-assignment and pricing claims contradicted by the vendor's own technical report.
- Two defects in this project's own week-old work.
claude --bg --worktree "<task>"cannot run —--worktree [name]consumes the task string as the worktree name (state: failed, "Invalid worktree name"). And the silent-idle demo used--mcp-config, which dies in ~10s instead of idling;--allowedToolsgenuinely idles (state: blocked, live pid). The phenomenon was real, the demonstration was not — and the adjudicator's own swap left a composite transcript (new command, old session id), caught and replaced with the real reproduction. claude agents --jsongains a fourth observedstate:blocked.
Added¶
DISABLE_AUTOUPDATER=1documented inrun-loop-news.env.exampleand set machine-locally. The prevention that pairs withv3.1.5's detection: the tracker cannot straddle two binaries, and cannot run a version the macOS firewall has never seen, unattended, at 04:00.KB_GAPS.mdV13–V16.
Method — a date-check that changed a verdict¶
Three high-severity docs/04 rewrites rested on one agent reading one repo on one day. Pulling
eugenelim/agent-ready-repo as of the KB's own 2026-06-26 capture showed the claims were never
supported (the RFC grew 389→612 lines, but SCOPE.md/Stop hook/test have zero hits in
both versions). The reverse case is recorded as V13: a zeroshot quote the KB got right in July
that the source later deleted, which a finder called fabricated and a refuter cleared by pulling the
capture-date commit. "Contradicted by the source" must mean contradicted as of the capture date,
or a fact-check manufactures its own false positives.
Still incomplete¶
Only four docs were re-examined. docs/11's model IDs and pricing — the KB's most volatile
class — have still never been swept anywhere. The critic's P1/P2 items remain open in the backlog.
Version stamps¶
Restamped to 2.1.263 only where a claim was genuinely re-observed on it. Three claims stay
stamped 2.1.261 because nothing in this pass re-ran them; bumping them would assert a
verification nobody performed.
[3.1.5] — 20260906 08:50¶
The tracker ran two stages on two different Claude Code binaries and nothing recorded it. This makes that impossible to miss again.
Fixed¶
- The pipeline could not tell which CLI version produced a run. It logged the binary path;
claudeis a symlink into~/.local/share/claude/versions/<v>, and the auto-updater can repoint it between stages. On 2026-09-06 it did: stage A ran2.1.261, the symlink flipped at 04:17, and stage B ran2.1.263. The run log, the findings artifact and the digest all recorded a single clean success; the only surviving evidence was file mtimes.
Now every stage logs its resolved version, and a change mid-run is reported to the log and to
notify(). Deliberately not fatal — a self-updating CLI would otherwise skip a run every
time it updated — but never silent.
The version reader is hardened to return a bare semver or nothing: a probe that emits an unexpected banner would otherwise yield a garbage token, compare unequal, and raise a false drift alarm on every stage. Proven by test, not assumed — same-version does not fire, a 2.1.261→2.1.263 change does fire, and unreadable output does not.
Added¶
- docs/17 — "Toolchain changes under the loop mid-run". The generalisable form, which is not about Claude Code specifically: an unattended loop resolves its agent binary once, but a self-updating toolchain can swap it between stages, so behaviour may differ across what is reported as one run. And on macOS a per-binary firewall rule (LuLu, Little Snitch) is keyed to the executable path, so a new version inherits no rule — its outbound calls can be blocked, or held at a prompt nobody is awake to answer. That is a partial network loss which looks exactly like "that source had nothing today", which is this KB's defining defect class.
scripts/SCHEDULING.md— the one-liner for checking which version actually ran.
Measured, and stated as inconclusive¶
Whether the 2026-09-06 run was actually degraded could not be settled. Stage A was not: it ran on the approved binary and out-collected the previous day (126 candidates vs 108), with all 54 sources swept. Stage B was not blocked outright: it integrated 17 docs, committed and pushed. But the content that run added carries roughly 1.8x the defect density of the surrounding text (5 findings per 104 added lines against 44 per ~1,680 in the same four docs), and all five are source-fidelity errors — a misquotation, a singular changed to a plural, two conflated features, and "sub-second" where the source said "seconds". That is the signature of working from a summary rather than re-reading the primary source, which is consistent with degraded fetches; ordinary integration error looks the same. Recorded as suggestive, not proven.
[3.1.4] — 20260906 04:39¶
Loop news run 2026-09-06 04:00 UTC — 83 new findings after deduplication (126 scored, 43 already published by the 2026-09-05 run). Seventeen existing docs extended; no new doc files (Phase 4c: consolidation over fragmentation — every finding that cleared the bar folded into an established canonical home).
Added¶
- docs/24 — Query Loop as System Heartbeat + Ten Principles of Harness Engineering (Harness Books); Object-Oriented/Code-as-Action Agents (NOOA, 82.2% SWE-bench Verified at half token cost); YAML-Defined Process DAGs (Archon, 23k★); Unified Harness Protocol; Warp's first-party inner/outer self-improving-skill pattern; mechanically-enforced review-before-merge; a second controlled harness-token-efficiency data point.
- docs/27 — Anthropic's 1σ/2σ/3σ control-band autonomy ladder; loopgain's Barkhausen-stability stopping condition (92.8% spend cut, documented 4.5% false-converge rate); huvii174's deterministic circuit breaker + 95%-confidence design gate + tenth-man critic; "done" scaled by organizational size (MindStudio).
- docs/07 — cache-safe forking + isolated child state + SubagentStart/Stop hook payloads (Harness Books); tenth-man critic cross-reference; citation-verification and prior-art-collision review axes (solidforge).
- docs/17 — Scheduler-invisible loop row (SGH's three structural flaws); reward hacking extended with METR/OpenAI transcript-tampering findings and a verify-the-verifier CI pattern; a second cognitive-surrender study (Anthropic/Trio, 50% vs. 67%); Willison's Winchester Mystery House quote on comprehension debt; a concrete circuit-breaker implementation citation.
- docs/04 — ARC-AGI-3's own efficiency framing (humans 100% vs. frontier AI <1%); AutoSciRub (rubric induced before execution); Last Step Matters (early-vs-late trajectory confidence unreliability).
- docs/34 — DEPBENCH data point (51.2% best-harness dependency-upgrade solve rate).
- docs/33 — a second privilege-escalation paper (13 attack objectives x 6 harnesses).
- docs/25 — NVIDIA AVO case study (100% ARC-AGI-3, 7-day autonomous GPU-kernel optimization); the 11-day, dozens-of-agents Fermat's Last Theorem Lean formalization.
- docs/28 — first-party production example (Anthropic field marketer's weekly routine).
- docs/31 — Production Deployment: On-Call for CI/CD (14-min median first-analysis,
durable
lessons.md). - docs/29 — Restart Resumability (OpenClaw, for comparison).
- docs/13 — The New Rules for Claude 5-Generation Models (80% system-prompt reduction).
- docs/11 — Multi-Dimensional Budget Pressure; Ponytail overengineering countermeasure.
- docs/16 — a documented negative result on vector memory (0/6 vs. plain
rg). - docs/21 — a second academic graph/loop/harness three-way-split paper; a practical graph-vs-loop decision rule; the harness→loop→graph naming-progression essay (ikangai).
- docs/23 — Desktop Multi-Agent Office pattern ("GOD agent" coordinator, 6,448★).
- docs/26 — the primary source for the "dark factory" name (Dan Shapiro), cited directly.
- SOURCES.md — six new tracked rows (ARC Prize, Archon, Hanako, Kirill, ikangai, Bruno Gonçalves/Data For Science, METR).
Fixed¶
- Two stale SOURCES.md rows corrected by GitHub-API confirmation: Happenmass/Cliclaw had been listed as a separate repo from Happenmass/omux ("second repo from the same author") — it is the same repo, renamed; the old repo ID no longer resolves. hhamja/loop-harness → hhamja/claude-code-harness (confirmed via repo-ID redirect).
[3.1.3] — 20260906 00:27¶
Source-list revalidation, run before the next unattended tracker fire rather than after it — a dead source row costs a run's yield silently, and this repo has lost two months to that twice.
Fixed¶
- One dead source, found by measurement. All 47 checkable URLs in
SOURCES.mdwere fetched; 45 returned 200. AI Breakfast wasrssagainst a feed that has 404'd since at least 2026-07-08./feed,/feed.xml,/rssandrss.beehiiv.com/feeds/aibreakfast.xmlall 404, and the homepage carries no<link rel=alternate>and no beehiiv feed id — no discoverable feed exists. Switchedrss→htmlagainst the index page, the same defect and the same fix The Batch needed. Its own note said "try/rssnext run" and sat unactioned for two months — a row that records its own breakage is not a fixed row. - arXiv row normalised
http://→https://(it was 301-redirecting on every run). - README documented 3 source types; 7 are in use — the 4 undocumented ones cover 28 of 54 rows. All seven are now a table with row counts, and the block moved back under "Add or remove a source", where it belongs (H3).
Added¶
- A dated revalidation record at the top of
SOURCES.md, so the next agent can tell checked and live from never checked — the distinction the file's own warning is about. - README guidance to prefer
htmloverrsswhen no feed is discoverable, with the reason.
Fixed — two ambiguities this project introduced yesterday¶
last_run_datecould have been read off a hand-authored entry.fetch-loop-newsstep 2 said "find the most recent dated section header".v3.1.2's fact-check entry sits directly above the last real run header and contains a date, so an agent could plausibly have takenlast_run_dateas 22:22 instead of the real 11:27 — silently skipping ~11 hours of source content, with nothing to report it. Now: match## YYYY-MM-DD HH:MM UTC (…)and skip any##not beginning with a digit. The freshness watchdog was already safe because it uses a regex; this instruction is prose read by a model, and "dated header" was ambiguous exactly where the regex was not.KB_GAPS.md's new V-items could have been pruned.integrate-loop-newsPhase 7 removes gaps filled by a run's doc writes. The V-items are verified research a run will never fill, so the section now carries an explicit contract:fetch-loop-newsmust not derive search keywords from it,integrate-loop-newsmust not prune it.- Two files credited the wrong skill (H2).
LOOP_ENGINEERING_NEWS.mdandKB_GAPS.mdboth saidfetch-loop-newsupdates them; it writes only.loop-news/findings.jsonand never touches a tracked file. Correction to H2 as filed: itsdocs/34half was not falsified —"05:00 local"is correct, and"L3 — commits and publishes autonomously"became true when the tracker published on 20260905.
Corrected in our own records¶
C8 was mis-sized. The backlog said "52 of 53 rows carry no confirmation newer than Jul 2026",
which was true and misleading: undated is not dead. Measured, 1 of 47 was broken, making
this a ~20-minute task rather than the medium sweep implied. Row counts were stale too — 54 rows,
not 53; github 23, not 22. Both corrected in place.
Still not validated, and never was in scope: whether a live URL actually yields anything. A
200 proves reachable, not useful. Only SOURCES.md's own "3 consecutive silent runs" rule covers
that, and nothing automates it.
[3.1.2] — 20260905 22:23¶
The 14 never-fact-checked docs, checked. 341 claims examined, 49 fixes applied, one commit per
doc. PATCH by this file's own rule: no new docs/*.md. The pass also falsified two claims
v3.1.1 had shipped eight hours earlier — both are corrected here.
Method: one Sonnet finder per doc → two independent Sonnet refuters per finding (over-correction lens, source-fidelity lens) → one Opus adjudication over the collected set → one Opus completeness critic. 134 agents, 8.6M tokens, 0 errors after a mid-run resume from cache.
Fixed¶
- A fabricated composite quotation in
docs/31: two sentences spliced with an em dash and the intervening text dropped, presented as verbatim. post-sessionis not a Claude Code hook event — and it appeared in three docs (12,28,29), only one of which was in the reviewed set. Caught by the adjudicator's cross-doc pass, which is the one thing no per-doc agent could do.--continuedocumented for resuming background sessions, in a section about exactly that, when both the CLI reference and the sessions page say it skips them.- Sakana Fugu pricing attributed backwards in
docs/22, plus a role-assignment mechanism attributed to Fugu that its own technical report says it does not do. - A
session-orchestratormechanism that does not exist indocs/10:acting_onandcheck-file-lock.shappear nowhere in that repo's 884-commit history. docs/03's opening framing claim contradicted by its own primary source;claude mcp add --globalis not a flag; the Chrome-support list excluded browsers the official doc includes.- Four dead or renamed citations, kept with dated notes rather than deleted, plus
docs/20's last bare@handle.
Corrected — claims this project published hours earlier¶
--max-turnsis silently inert on--bg. Paired test on 2.1.261:-p --max-turns 1errorsReached max turns (1);--bg --max-turns 1ran a three-step task to completion.v3.1.1listed it as a working ceiling — inside the section warning that a control which cannot take effect must fail loudly. With--max-budget-usdalso inert, a--bgsession has no in-band ceiling at all. The original claim was written from help text instead of from running it.docs/03'sclaude mcp add # interactive setupcannot run. There is no interactive form. Replaced with three examples taken fromclaude mcp add --helpitself.
Added¶
- A guard on the freshness watchdog's own blind spot.
check-digest-freshness.shmatches## YYYY-MM-DD HH:MM UTCand ignores the trailing parenthetical, so any header in that shape resets the staleness clock — a hand-authored digest entry would have reported the tracker healthy while it was dead. This release's own digest entry uses a non-matching header, the constraint is documented in the script and inCLAUDE.md, and the watchdog verifiably still measures the 11:27 tracker run. KB_GAPS.mdV7–V12 — six claims no search can settle, and a note on the two that were settled by running them instead of logging them.
Known incomplete — stated, not glossed¶
C1's own Opus completeness critic returned INCOMPLETE, and C1 is not closed:
- 268 lines were never examined while being reported as covered —
c6e5a66landed +119 lines intodocs/03and +149 intodocs/29mid-run, so two finders read pre-commit copies. - 78 of the 98 unique URLs cited across the 14 docs were never opened.
- Claim density ran inverted against priority —
docs/270.052 claims/line againstdocs/31's 0.31. - Model IDs and pricing were never swept, including
docs/11, the densest concentration in the KB.
The dominant risk in what did land is a failed fetch reported as an absence — an unauthenticated code search returned 401 and was read as "0 hits". The two fixes that actually remove content were re-verified by hand with authenticated calls first, and that check caught the adjudicator misattributing a commit message between two commits. Carried as C1b.
[3.1.1] — 20260905 17:00¶
Three backlog items with sources already in hand — C6, C13, C14 — closed in one pass. PATCH by
this file's own rule: no new docs/*.md file. Every claim was re-fetched from source before it
reached a public page rather than transcribed from the evidence packs, which is what turned up the
findings below that the packs did not contain.
Added¶
claude --worktree, written up in full (C6,docs/03) — path and branch naming, branching from a PR/MR with per-host ref resolution,.worktreeinclude, the resume/cleanup lifecycle, and the four hard-enforced isolation checks. The point of the section is the last one: a repo convention like this project's own worktree rule is prose in aCLAUDE.mdthat holds only while an agent reads and retains it, whereas--worktreemakes the same rule a property of the harness that propagates to every subagent without being restated — a STOP condition enforced by the runtime rather than stated in the prompt.- The
--bgcontract (C14,docs/29) — what a background session does not get, the per-session cost floor, theclaude agents --jsonfield shape, and the silent-idle trap. - A version-stamped CLI gotchas table (
docs/18) — also closes part ofH1. KB_GAPS.md§ Claims Awaiting Verification — six claims no search can settle, marked not a search target sofetch-loop-newsskips them.
Fixed¶
- Four
CLAUDE_CODE_*claims (C13).CLAUDE_CODE_REMOTEistrue, not"1"— a hook guarding on= "1"never fires.CLAUDE_CODE_SUBAGENT_MODELoverrode rungs 1 and 2 before v2.1.251,model: inheritincluded.docs/07's backticked error literalConcurrent subagent limit reachedis unsourced and is gone — replaced by a fact the reference does state, that the cap takes plain digits and cannot be disabled. And this changelog's own[3.0.0]entry calledCLAUDE_CODE_MAX_SUBAGENTS_PER_SESSION"an env var that does not exist": it exists, documented as removed in v2.1.224 and now a no-op. Removing the row was right; the rationale was itself the over-correction the entry was written to warn about. - Two copyable examples in
docs/29that could not work.--bg --max-budget-usd 5.00showed a dollar cap that is silently inert on a background session, and the fan-out didclaude agents --json | jq -r '.[].id', which emits nulls for interactive sessions — they carry noid— and returned every session on the machine rather than the loop's own. docs/09's fail-closed guidance read as universal.--permission-prompts noneis--print-only, and--bgconflicts with--print, so a background session cannot fail closed that way. The same is true of--max-budget-usd.- A fail-open counterpart in
docs/08— a settingsenvblock cannot turnCLAUDE_CODE_RESTRICTEDon either, so a committed{"env": {"CLAUDE_CODE_RESTRICTED": "1"}}yields an unrestricted session and no error.
Verified¶
- The variadic-flag swallow is real, reproduced two ways on
claude 2.1.261. Seven flags take variadic lists, and a variadic flag consumes the positional prompt:claude -p --mcp-config /nonexistent.json "reply with OK"reportsMCP config file not found: <cwd>/reply with OK. On--bgthe same mistake is silent — the launcher prints an id, exits 0, and leaves a session marked(idle — send a prompt to start)that never does the work. A fan-out built flags-first can start fifty of these and report complete success. That is this repo's own defining defect class, found in the CLI it documents. claude agents --allexists;--bg -pis rejected with the message reproduced verbatim;--max-budget-usdis silently accepted on--bg;--max-turnsis still accepted but no longer listed inclaude --help.- The v2.1.207 quotation at
docs/08:101reproduces verbatim upstream — settling one of the evidence pack's seven open uncertainties.
Corrected in our own records¶
- The backlog's cost-floor arithmetic does not reconcile.
C14claimed the $0.24 session floor "matches to four decimals". It does not: 22,659 tokens at Opus 5's $10/MTok 1h-cache-write rate is $0.2266, leaving $0.0139 — 5.8% of the recorded $0.240609 — unexplained by the three published token counts. The generalisable claim survives, and is now indocs/29anddocs/11with the residual stated: the cache write is 94% of that session's cost, the floor scales with the model's price rather than the task's difficulty, and the same write costs $0.09 on Sonnet 5 and $0.05 on Haiku 4.5. Logged asV6. The backlog entry is corrected in place. - Nothing was deleted for being merely unsourced. "Not on the page we fetched" is a fact about our
fetch; "contradicted by the page we fetched" is a fact about the KB. Only the second licensed an
edit —
waitingForand five other unconfirmed claims were logged, not asserted and not removed.
[3.1.0] — 20260905 13:43¶
The pipeline that had been silently dead for eight weeks is now fault-tolerant. Twelve fixes, each verified by mutation or a pre-fix control rather than by assertion. No KB content changed beyond one model-id pass, so the MINOR is for the new capabilities: an off-machine watchdog and resumable stages.
Added¶
- Off-machine staleness watchdog (
A3) —scripts/check-digest-freshness.shplus a daily Actions job that fails when the newest digest entry is more than 48h old. It runs off the tracker's machine, because a watchdog sharing a failure domain with the thing it watches is not a watchdog, and it asserts on the published artifact rather than on any run's exit status. A missing file, empty file, vanished header format, unparseable stamp and future-dated stamp all fail the same way a stale entry does — "I could not tell" is never reported as "fine". - Resumable stages (
A10,A11,A12) — both stages now survive dying at an arbitrary point. Stage A checkpoints per source (completeflag +sources_done[]) and a later run resumes from the partial artifact; Stage B commits checkpoints to a stable per-day branch, and the wrapper resumes from the branch tip. Git is deliberately the ledger for Stage B: progress inside a Claude session is invisible to the wrapper, so an agent-maintained progress file would be advisory prose, while a commit is something the wrapper can verify. Checkpoints are squashed before push, somainstill receives exactly one commit per run. - A
mkdocs build --strictgate before the pipeline pushes (A1) — reviewed by five adversarial lenses and an Opus judge before merge. On its first-ever execution it caught and fixed three broken anchor links before they reached a commit.
Fixed¶
CLAUDE_BINpointed at a path that no longer exists (A7) — the whole eight-week outage. It was hardcoded to/opt/homebrew/bin/claude; the CLI moved to~/.local/binand every run exited 127 three times, reporting only to a desktop notification and a gitignored log. Replaced with a resolver and a fatal preflight (exit 3).- Two failures the wrapper reported as success (
A8,A9) — an unresolved slash command exits 0, so a missing skill made Stage B logRun completehaving published nothing; and an unanchoredgrepon commit subjects matched aRevertof the pipeline's own commit, abandoning retriable failures. - CI could not see the files the site publishes (
A4) —docs/news.md,sources.mdandchangelog.mdare symlinks, so a findings-only day matched no trigger path: no build, no deploy, and a green history because no run was attempted. - The skill never updated
docs/index.md(A5), the site's actual home page and a second index--strictcannot check. - "Credit balance is too low" got generic retry treatment (
A6) — deterministic, but it burned three attempts and both backoffs before reporting. Now exit 5, immediately, naming the shadowing API key as the cause. - Superseded model IDs in copyable examples (
C2) —claude-opus-4-8in three docs plus three more the same check missed. Note that ID is superseded, not invalid; the defect was offering a previous-generation model as the generic example.
Changed¶
- Every timestamp is UTC
YYYYMMDD HH:MM(D1) — the previous skills-UTC/humans-local split had produced three formats inside this file. Entries before this one keep their original local time; restamping them would invent precision nobody measured. - Pipeline-cut versions are released by a human follow-up (
D2) — the rule demanded a tag and release "in the same turn", which the pipeline structurally cannot do. Granting an unattended agentgh releaserights on a public repo was judged a wider permission than the convenience warranted. - Plans retire in place with a
Shipped as vX.Y.Zheader (D3), rather than moving to an archive directory and breaking inbound links. plans/20260904_2053-open-work-backlog.md— a 31-item verified backlog, now the repo's entry point, produced by a 16-agent audit with adversarial verification.
[3.0.1] — 20260905 12:08¶
Largest single digest run since the tracker resumed: 108 findings scored, 101 new. No new doc files created — everything was folded into existing canonical docs per the Phase 4c consolidation rule, so this is a PATCH despite the volume.
Added¶
- docs/21 — Sep 2026 continuation of the loops-vs-graphs debate: Karpathy's maturity-stage progression, the "chorus not ensemble" parallel-review cost argument, a control-theory reframing of loop failure modes, a dissenting skeptic view, academic + Google corroboration.
- docs/27 — "Concrete STOP+Verifier Implementations (Sept 2026 cohort)": ruvnet's dream-machine/autogenous/openAVO/sparc, loop.js, loop-contract-skill, DeMARS, LoopArena.
- docs/24 — ARC-AGI-3/Astra harness-disclosure controversy; four new self-improving-harness papers; generic-harness-beats-bespoke finding; GPT-Red; HydraFusion.
- docs/33 — two new sections: hook/context trust attacks (HookPry, Context Privilege Escalation) and emergent multi-agent coordination risk (rogue-agent wiki collusion, DeepMind swarm self-governance).
- docs/17 — two new failure-pattern rows (Teardown blindness, Comprehension debt).
- docs/04 — non-probabilistic node rule; benchmark/eval-integrity corpus.
- docs/16 — new Pattern J: Learned Memory Substrates (RuVector, funes, Computer History).
- docs/06 — skill compilation from expert corpora (mimeo); skill-bundle compression.
- docs/22 — training-environment evolution for terminal agents; Harness-RL; MemoryWalker.
- docs/25 — Ralph loop origin citation (ghuntley.com); Harness-of-Harness cross-reference.
- docs/26 — "lit vs. dark" factory naming; platform-relocation survey; scientific-computing field report.
- docs/14 — persona document as a calibration feedback channel.
- docs/05 — "Rules Have a Half-Life."
- docs/23 — rvm trust-driven runtime isolation; ruClip ecosystem map.
- docs/30 — 107M-row single-goal case study; Aspire benchmark.
- SOURCES.md — huangruiteng/loopx promoted to a tracked row.
- KB_GAPS.md — two new gaps (retrieval-infrastructure mechanism, parallel-graph-cost claim).
Changed¶
- docs/32 — Reference Implementations: replaced
loop-kernelwithloop-contract-skill(loop-kernel's three-exit-code contribution is fully absorbed into docs/27's prose). - LOOP_ENGINEERING.md — row summaries updated for every doc touched above.
[3.0.0] — 20260904 19:12¶
Restructured into two clearly separated parts, and corrected nine factual defects the restructure surfaced. First release since 2026-07-09; covers an eight-week tracker outage.
Changed — BREAKING¶
- The index and site navigation are reorganised into two parts. Part I — Loop Engineering is
general and tool-agnostic (Claude Code appears as an example, never a prerequisite). Part II —
Developing with Claude Code is concrete and version-stamped.
LOOP_ENGINEERING.md,docs/index.md,mkdocs.ymlnav andREADME.mdall rewritten. Nodocs/NN-filename and no published URL changed, so no external link breaks — only grouping and framing moved. Every one of the 39 docs appears in the nav exactly once; all 34 pre-existing row summaries preserved verbatim. site_name:Loop Engineering→Claude Loops, matching the repo and project.- Routines, Claude Tag, Background Agents and Headless moved from Part I to Part II — all four are Claude Code product surfaces and cannot sit in a part that must stay general.
- The docs/01 thesis is scoped, not discarded. "Stop writing prompts, start designing loops" stays as a claim about a class of work. The table no longer claims "design the loop once, let it run" or an unqualified "the loop verifies itself" — both overstated, and both were in tension with docs/14 inside the KB.
Added¶
docs/35-choosing-your-mode— the router the KB never had. Three tests a task must pass before a loop is worth building; task-properties decision table; the supervised-autonomy middle ground where most real work sits; the compound-probability argument re-scoped as an argument for a correction loop rather than an unattended one.docs/36-development-workflow— the Part II spine. Ng's three phases × five skills, each mapped to the Claude Code primitive that implements it.docs/37-session-architecture— split by context boundary, never by job title. Anthropic's three justifications and the one that is absent; the telephone game; the 90.2% figure and the population it was actually measured on; the official four-way primitive matrix; native cross-session messaging; a named case study with its refuted-claims table.docs/38-agent-teamsanddocs/39-dynamic-workflows— two shipped primitives with zero prior KB coverage.- docs/24 "When to remove harness" — the KB accumulated harness monotonically with no doctrine for removing any of it. Harness complexity is a depreciating asset with a stated decay direction.
- A systematic pass over the full Claude Code changelog (620KB, 385 versions, 2,076 bullets,
filtered locally to 187 decision-relevant entries for v2.1.200+). Integrated across eleven docs:
the subagent model-routing reorder (
CLAUDE_CODE_SUBAGENT_MODELdemoted from override to default in v2.1.251;CLAUDE_CODE_SUBAGENT_MODEL_FORCEadded v2.1.257 as the real ceiling;/tasksshows what each subagent actually ran on since v2.1.243, so it is checkable);PreModelSwitch/PostModelSwitchhooks;promptCacheTtl/subagentPromptCacheTtland per-agentexperimental.cacheTtl; the/usageLoops breakdown as the first-party cost-per-loop-run figure the Loop Contract's BUDGET leg asks for;--permission-prompts noneas a fail-closed unattended default (the opposite of--dangerously-skip-permissions); per-model auto-compact thresholds; andnotify_when_idle(v2.1.236), the native form of the zero-polling pattern the KB had credited only to a third-party harness — filed as an instance of the shrink-on-native dynamic rather than as a new feature. - docs/07 + docs/11: a capped subagent no longer looks finished. Since v2.1.246 a subagent that
stops at
maxTurnsreturns its output marked partial. That is the platform fixing, at the runtime level, the exact defect docs/37 documents as a case study. - docs/08: "repo settings cannot escalate their own privilege" — a pattern named across four
releases.
bypassPermissions(v2.1.257),autoMode(v2.1.207) andsandbox.ripgrep(v2.1.232) were each demoted out of project settings. A cloned repo no longer votes on its own permissions. - docs/26 spec-first conditioned on project maturity; docs/08 + docs/33 the cost of refusal; docs/21 the loops-vs-graphs debate (recorded as contested, since the framing traces to a joke rather than a consolidation piece); docs/17 Silent Default Drift as a named failure pattern; docs/11 a model reference table (Sonnet 5 / Opus 5 / Fable 5.1 / Haiku 4.5) with selection guidance.
Fixed¶
- A fabricated quotation attributed to Andrew Ng, removed from docs/14. A stitched composite — two real fragments joined by the KB's own connective and presented as quotation. Replaced with the verified sentence.
- A miscited letter in docs/14 and docs/32 — wrong title, and a newsletter-issue URL instead of the article.
- Eight stale or wrong platform facts across docs/07, 08, 09, 15, 18: the
Exploresubagent's model, the nested-subagent depth cap (5 → 3, and configurable), subagent frontmatter shape, a legacy model pin,Ctrl+Rfor plan mode (it is Shift+Tab) in two docs, and a--no-streamflag that does not exist. Three further claims were checked and confirmed correct. - A broken anchor in docs/11, caught by
mkdocs build --strict. - docs/23 — the Bun figures were correct but unlabelled; 535,496 / ~750,000 / 1,009,272 are three different measurements and are now named as such, with a practitioner counterweight.
- A dead limit this release itself introduced as live, then removed.
docs/07gained a "200 total subagent spawns per session" limit with an env var,CLAUDE_CODE_MAX_SUBAGENTS_PER_SESSION, that no longer does anything. The variable is real — removed in v2.1.224 and now a no-op — and the 200-spawn cap it once set (added v2.1.212) no longer applies; the current reference says plainly there is no limit on total spawns over a session. A reader would have pasted a dead knob intosettings.json. Caught twice independently and simultaneously — by an adversarial review lens and by the changelog pass — eleven lines above this page's own warning that an unversioned platform limit is an unversioned claim. - An over-correction, reversed. An agent claim that the
AskUserQuestionauto-continue shipped inv2.1.198was stripped as unsourced after checking the changelog and the HN thread. It was true: Anthropic's environment-variable reference states "In v2.1.198 and v2.1.199, auto-continue was on by default with a60000(60 seconds) timeout", and the linked article pins it by binary diff. Neither source had been checked. Calling a true claim fabricated is the same defect as asserting a false one, in the other direction. The narrow claim survives — no changelog entry names it — and the doc now shows where the version is actually recoverable. - A second broken in-page anchor in docs/11, caught by the same strict build. Both were fixed by
reading the
idout of the generated HTML rather than guessing the slug. - docs/23:
6,502commits labelled as the merges-excluded figure against the source's own6,778headline. docs/14: the HN comment count corrected to 120 and date-stamped, since it moves. docs/25: hedged — it is a Part I doc, and this release created the rule that Part I never requires Claude Code, then filed unaudited content under it. docs/35 + index: the Anthropic multi-agent research post dated2025-06-13; it had sat undated beside a 2026-09-04 quote, fifteen months apart with nothing telling the reader. SOURCES.md— The Batch row switchedrss→html; the feed had 404'd since 2026-07-08 and silently returned nothing for two months while looking configured. Adds a standing rule that a source yielding nothing for >~3 runs must be re-fetched by hand before its silence is believed.
[2.7.6] — 2026-07-09 08:19 IST¶
Added¶
- Bun's Zig→Rust rewrite case study (manually researched, bun.com/blog/bun-in-rust): docs/23 new case study section — 64 parallel Claude Code instances across 4 worktrees, ~1,300 LOC/min peak, 535,496-line rewrite in 11 days, plain-worktree coordination with no external queue layer (contrast with Gas Town's Beads); docs/04 two new findings under Verifier Integrity — language-independent (TypeScript) test suite as a cross-language-rewrite verifier, and blind adversarial review quantified at 64-agent scale (128 bugs fixed, 19 regressions); docs/24 new "Fix the process, not the code" subsection under Self-Improving Harnesses — a manual, single-engineer instance of harness-level fixes over point-fixes (dangerous git commands fixed once at the workflow-instruction level, not per-instance); docs/26 new bullet under Named Factory Deployments — the granular porting-guide → mechanical-port → compiler-errors → tests-green task decomposition that made the rewrite agent-sized. No new doc files; four existing docs updated, cross-linked to each other.
[2.7.5] — 2026-07-08 07:09 IST¶
Added¶
- Loop news run 2026-07-08 04:15 UTC (recovery run — see CHANGELOG "Fixed" entries below for why): 52 new findings integrated after dedup against 45 already-published URLs. New KB sections: docs/24 "Dynamic Workflow Patterns" (Anthropic's six orchestration patterns) and "Cross-Model Division of Labor" (consolidating four independent findings — steipete's advisor loop, codex-first SKILL.md, Puppetmaster, official openai/codex-plugin-cc — into one canonical home per Phase 4c); docs/24 "Task-Shaped DAG Orchestration" (Kaola-Workflow) and schema-level conformance (Temper); docs/27 "Quota-Aware Should-Run Gate" (loopx) and plan-as-contract note (walidboulanouar); docs/16 "Pattern I: Durable Objectives with Evidence Logs" (loopx); docs/23 swarm-topology-via-consensus subsection and a Gas Town case study (20-30 parallel Claude Code instances via git-persisted Beads); docs/33 "Cross-Org Federation (Zero-Trust)" (ruflo); docs/26 "Dark-Factory Ceiling and Its Bottleneck" and "Named Factory Deployments" (Droid Shield 2.0, auto-merge discussion), consolidating a second Phase 4c dominant theme (autonomy-maturity framing recurred across 4 findings); docs/17 new failure pattern "Ghosting under review feedback"; docs/11 "Confidence-Scheduled Verification"; docs/29 "Zero-Polling Signaling" and "Cloud/Mobile Background Execution"; docs/14 "Who Interrupts Whom, More Often"; docs/01 "Who Benefits Most From the Shift"; docs/04 per-criterion independent verification (manifest-dev), a 93.4%-non-overlap production corroboration of the existing cross-model-reviewer figure, and a 43%-fabrication-rate adversarial-gate case study (walidboulanouar).
- Nine candidate repos (loopx, walidboulanouar/loop-engineering, manifest-dev, Kaola-Workflow, ruflo, temper, athena-loops, omnigent, claude-deep-loop) put through the resource-review gate (scored 0-5 on unique-contribution/precision/durability); all nine cleared the ≥3.0 bar and were integrated. SOURCES.md: added ruvnet (github profile, 2+ relevant repos).
Changed¶
- docs/32-reading-list.md — Reference Implementations: replaced Strive_Engineering with LoopX (Strive's provenance-bound-claims contribution is now fully absorbed into docs/04 with direct citations); Getting Started: added Ben's Bites "My thoughts on Fable".
- LOOP_ENGINEERING.md index summaries refreshed for rows 1, 4, 10, 11, 14, 16, 17, 22, 23, 24, 26, 27, 29, 32, 33 (staleness check).
- KB_GAPS.md — new gap (underspecified-input mitigation for autonomous pipelines); annotated the existing cross-model-reviewer-pairing and F0-F3 fleet-maturity gaps with this run's near-miss evidence (both remain open).
[2.7.4] — 2026-07-08 IST¶
Fixed¶
- Stage-A self-execution bug (production incident, 2026-07-04):
fetch-loop-news's Phase 4 said "if invoked interactively, run/integrate-loop-newsyourself" — a condition a headless session can't reliably evaluate, so it did so anyway, executing the entire integrate stage (digest, KB writes, release, commit, push) inside its own search-stage session and collapsing the deliberate two-session split back into one. Removed the exception and told the skill to stop unconditionally. - That prose fix alone was insufficient — recurred on the very next production run
(2026-07-06), same collapse, despite the unconditional wording. Confirmed prompts are
not a substitute for an actual permission boundary. The real fix: Stage A's session now
runs with
--disallowedTools "Bash(git *),Bash(gh *),Skill"— a genuine deny-list, experimentally verified to hold even under--permission-mode auto(unlike--allowedTools, which the auto classifier can approve beyond). Scoped to git/gh/Skill rather than blanketBash, since Stage A still legitimately needs plainBashfor adatecall in Phase 1 — verified that scoped deny blocksgit logwhile still allowingdatethrough. - Added a defense-in-depth check to
integrate-loop-newsPhase 0: before doing any KB work, checkgit log --grepfor a commit matching this run'srun_timeand abort immediately if the run was already published (cheap; catches a recurrence early). - Added a wrapper-level guard (zero LLM cost): after Stage A completes, check whether
origin/mainadvanced past the pre-run base SHA; if so, skip launching Stage B entirely instead of paying for a full redundant session. docs/09-headless-mode.md— documented the general pitfall (a skill can't distinguish interactive from headless invocation) and the three-layer fix.- Adversarial self-review caught and fixed a false-positive introduced by the fix
above: checking only "did
origin/mainmove" (not "did it move because of this run") would misfire whenevermainadvances for an unrelated reason — e.g. a human merging a different PR concurrently, which happens routinely in this repo. That would silently skip Stage B (or abandon a legitimate retry) on a day the digest was never actually published. Both the new wrapper guard and the pre-existing publish-safety retry guard now check for a matchingloop news runcommit in the delta, not just whether the SHA changed; an unrelated advance rebases the tracked base forward and continues normally instead of misfiring. integrate-loop-news's already-published check now fetches and checks againstorigin/main(not just local history, which could miss a publish made by a different worktree/process) and no longer caps the log search depth.- Production incident (2026-07-05): Stage A's parallel per-source subagents tripped
the CLI's internal 600s background-task wait ceiling ("Background tasks still running
after 600s; terminating"), causing all 3 attempts to fail with no digest published.
Set
CLAUDE_CODE_PRINT_BG_WAIT_CEILING_MS=0(per the CLI's own suggested remedy) — safe since the outer--max-turns/--max-budget-usdceilings still bound the overall session regardless. - Production incident (2026-07-06, validation run #2): hit the Claude account session
limit ("You've hit your session limit · resets <time>"). This message matched
neither
ERROR_REGEXnorBUDGET_EXCEEDED_REGEX, so the wrapper treated it as an ordinary retriable failure and burned all 3 attempts (with 30s/90s backoffs) against a limit that resets on a wall clock, not on a short backoff. AddedSESSION_LIMIT_REGEXas a third deterministic non-retry class (same pattern as budget-exceeded): on match, stop immediately with a clear notification instead of retrying. - Production incident (2026-07-06, validation run #3): with the structural fix
confirmed working (Stage A correctly stopped after search, no self-execution, on two
separate attempts), Stage B then hit
Reached max turns (100)twice in a row on a large-batch day (74 findings, 9+ docs touched). RaisedINTEGRATE_MAX_TURNS100→250 — same rationale as the earlier 40→100 raise: it's a ceiling, not a target.
[2.7.3] — 2026-07-07 05:27 IST¶
Added¶
docs/04-verification.md— Reviewer Freshness Enforcement (model vs. perspective independence, never-see-the-draft rule, synthesizer-bias mitigation), filling the reviewer-freshness-enforcement KB gap; held-out evaluator-only "holdout wall" refinementdocs/24-harness-patterns.md— Darwin Mode (train/eval-disjoint held-out gating), filling the held-out-eval-sizing KB gap and closing SEAGym's snapshot-collapse risk; mechanized-vs-prose compliance metric (~100% vs ~70-90%); harness-bench self-correcting capability tables; harness-design origin-story citation (The Making of Claude Code)docs/10-fan-out.md— Agentic MapReduce pattern (map signals → bounded fan-out → reduce → sandbox-verify)docs/32-reading-list.md— added metaharness (Darwin Mode) to Self-Improving Harnesses
Changed¶
docs/23-fleet-engineering.md— noted megingjord-harness as a partial (not sufficient) contribution to the F0-F3 observable-indicators gapKB_GAPS.md— marked reviewer-freshness-enforcement and held-out-eval-sizing gaps filled; annotated F0-F3 and cross-model-pairing gaps with this run's partial findingsSOURCES.md— Anthropic switched from dead RSS feed to html-typeclaude.com/blog; rediscovered feed URLs for TLDR AI, The Rundown AI, and Ben's Bites; removed deadorobsonn/claude-harness(404); promoted hhamja to a tracked github sourceLOOP_ENGINEERING.md— updated docs/04, docs/24, and docs/10 summaries for this run's additions
[2.7.2] — 2026-07-06 17:38 IST¶
Added¶
docs/16-memory-patterns.md— Pattern H: LLM Wiki (compiled, cross-linked organizational knowledge base; converged independently across 3 sources this run — Karpathy, The New Stack, Cobus Greyling)docs/33-agent-security-hardening.md— "Where Default-Deny Actually Gets Loaded" section (4 concrete enforcement points: MCP proxy, tool-dispatch layer, OS/kernel, session-bootstrap eval-parity), substantially filling the SECURITY_MATRIX loading-mechanism KB gapdocs/04-verification.md— cross-model reviewer arbitration mechanism (verdict-driven severity, promote-on-confirm, bounded 3-round reconciliation), filling half of the cross-model-arbitration KB gapdocs/23-fleet-engineering.md— Org-Chart Coordination pattern (agents mapped to reporting lines, coordinating over email, as an alternative to a graph topology)docs/24-harness-patterns.md— definitional-paper citation, StaminaBench + Claw-SWE-Bench quantified harness>model evidence (6x/4x gaps), SEAGym + APEX additionsdocs/11-cost-control.md— EurekAgent sub-$11 frontier-result cost benchmarkdocs/07-subagents.md— primary-source citation for the depth=5 nesting cap; Recursive Agent Harness (RAH) quantified evidencedocs/17-failure-patterns.md— 3-category root-cause taxonomy intro (underspecification / capability errors / harness errors)docs/32-reading-list.md— added HarnessX (arXiv 2606.14249) to Self-Improving Harnesses
Changed¶
docs/32-reading-list.md— removed a structurally broken entry (orphaned "Why here"/"Summary" text with no title or citation, left over from a prior edit) in "Loops in Production"KB_GAPS.md— marked 2 gaps filled (SECURITY_MATRIX loading mechanism; cross-model arbitration's disagreement-resolution half), refined the remaining arbitration question, advanced the held-out-eval gap, ruled out a false lead on F0-F3 fleet maturitySOURCES.md— fixed Addy Osmani's feed URL (was pointing at a dormant site), switched OpenAI and Sabrina Ramonov fromhtmltorss(feeds discovered this run), flagged 6 dead RSS feeds for rediscovery, added Happenmass/omux
[2.7.1] — 2026-07-04 IST¶
Added¶
scripts/SCHEDULING.md— runbook for changing the daily tracker's cadence and enabling/disabling it (macOS launchd:StartCalendarInterval/StartIntervalsyntax,launchctl bootstrap/bootout/enable/disable/kickstart). Linked from README and CLAUDE.md's repo map.
Changed¶
- CLAUDE.md Git workflow: added the worktree-cleanup rule — remove a worktree and its branch (local + remote) once its work is merged or abandoned.
[2.7.0] — 2026-07-04 IST¶
Added¶
run-loop-news.shCLI flags to set model/effort per stage on a one-off run, without exporting env vars:--search-model,--search-effort,--integrate-model,--integrate-effort, plus--model/--effortshorthands that set both stages at once. Precedence: CLI flag >LOOP_SEARCH_*/LOOP_INTEGRATE_*env var > built-in default.--helplists all options.scripts/run-loop-news.env.example— documents all 8 per-stage config vars (model/effort/max-turns/budget × search/integrate). The wrapper auto-sourcesscripts/run-loop-news.env(gitignored) if present, for a standing local override without exporting shell env vars.
Changed¶
- Raised
run-loop-news.sh's per-stage--max-budget-usddefaults (search $8→$30, integrate $8→$20). On a Claude subscription (Pro/Max) this flag is not a real dollar cost — it's a script-side runaway-session tripwire computed from API list-price- equivalent token usage — and $8 for the search stage alone was too tight for a normal thorough run, tripping on the live dry run. - Fixed a bug where a budget-exceeded failure was retried like a transient error: it's
deterministic (the same session hits the same wall), so retrying just re-burns the
budget for
MAX_ATTEMPTSruns with no chance of success. The wrapper now detects theExceeded USD budgetmarker and stops immediately with a notification instead. - Raised
INTEGRATE_MAX_TURNSdefault 40→100 (andSEARCH_MAX_TURNS40→100 for consistent headroom, no evidence needed — it's a ceiling, so a higher one costs nothing if unused). The 2026-07-04 production run hitError: Reached max turns (40)twice in a row on the integrate stage — Phase 4 (digest + integrate) plus Phase 4b plus the mandatory every-run Phase 4c structural review plus Phase 5 (release/commit/push) is enough real work that a normal day can exceed 40 turns.
[2.6.1] — 2026-07-04 09:04 IST¶
Changed¶
- Widest-yield loop-news run to date (103 new findings after dedup). Added the official
Claude-team loop-type taxonomy (turn/goal/time-based/proactive) and a quantified
harness-cost data point (7x cost reduction, Hugging Face) to
docs/24; a "Zombie finding" failure pattern (findings ratchet) todocs/17; a blind-spot ledger pattern todocs/16; severity-proportional reviewer routing todocs/07; a tamper-evident contract-hash + 8-named-exit-code extension todocs/27; and skill-ingestion security (OWASP Agentic Skills Top 10) + an A-F harness security scorecard todocs/33. - Added "Don't Train the Model, Evolve the Harness" (Hugging Face) to the reading list.
- Added 4 new tracked sources:
JasonxzWen/harness-hubandedonadei/caliper(both deep-read this run per the prior run's flag),explainx.ai, and Daily Dose of Data Science. Convertedswyx.ioand MindStudio Blog from HTML-scrape to their discovered RSS feed URLs.
[2.6.0] — 2026-07-03 IST¶
Added¶
- Split the daily tracker into two single-responsibility skills:
fetch-loop-news(search → writes.loop-news/findings.json) hands off to a newintegrate-loop-news(integrate + restructure + commit + push). scripts/run-loop-news.shrewritten to run the two skills as two sessions in one isolated git worktree branched offorigin/main, with per-attempt tree reset, granular retry (a Stage-B failure re-runs only B against the saved findings), and a publish-safety guard keyed offorigin/mainadvancing.- Per-stage
--model/--effort/--max-turns/--max-budget-usdconfig at the top of the wrapper, each independently overridable viaLOOP_SEARCH_*/LOOP_INTEGRATE_*env vars. - New "Knowledge-Base Tracker Loop" pattern in
docs/34; worktree/two-session production shape documented indocs/09.
Changed¶
integrate-loop-newsnow stagesmkdocs.ymlin its commit (Phase 4 editsnav:; the previous single skill omitted it, so nav additions were silently uncommitted).- Daily content commit no longer stages the skill's own
SKILL.md(skills are feature-managed via PR). - Repo convention: always work in a git worktree, never the primary checkout (CLAUDE.md).
[2.5.9] — 2026-07-03 IST¶
Added¶
docs/11-cost-control.md— "Reasoning effort is the dominant reliability lever — not tool access" subsection under Effort levels. A 90-run observational study (arXiv 2607.02436) finds raising efforthigh→xhighlifts first-try-perfect runs 28%→89% (~5× fewer corrective prompts) for +9–29% cost, while a bolted-on testing tool added 42–68% cost with no functional or reliability gain. Documents the counterintuitive rule — when weak reasoning is the root cause, spend the marginal dollar on reasoning budget before adding checker passes — cross-linked todocs/04so it does not read as weakening the independent-verifier mandate.
Changed¶
LOOP_ENGINEERING.md— refreshed the docs/11 index summary to mention the reasoning-effort reliability lever.
[2.5.8] — 2026-07-02 IST¶
Fixed¶
docs/16-memory-patterns.md— corrected a broken cross-anchor to24-harness-patterns.md(double-hyphen#control-plane--execution-plane-…→ single-hyphen, matching MkDocs' slug). This broken link had failed the 2.5.7--strictbuild, so 2.5.7 never deployed (the live site was stuck at 2.5.6); this release ships 2.5.7's content too.
Changed¶
.github/workflows/docs.yml— added aconcurrency: { group: pages, cancel-in-progress: false }group on the deploy job. Rapid successive pushes now serialize: one Pages deploy at a time, a newer run supersedes an older pending one, and an in-flight deploy is never cancelled — so the newest commit always deploys last (no out-of-order deploys). Browser cache is unaffected; a hard refresh still shows the newest content immediately.
[2.5.7] — 2026-07-02 IST¶
Added¶
docs/24-harness-patterns.md— Self-Improving Harnesses section: harness that evolves itself from its own execution traces (Self-Harness weakness-mining→propose→validate; AHE observability-driven evolution with verified prediction contracts, beating a human-designed baseline; HarnessX substitution algebra + AEGIS), with Terminal-Bench numbers and the ablation that gains come from tools/middleware/memory, not prompts. Fills the documented "self-scaffolding / model-generated harness" gap. Also added Control-Plane / Execution-Plane Split (kernel-gated mutation; claude-deep-loop).docs/16-memory-patterns.md— Pattern G: Repo-Owned Durable Ledger (ctxcarry "the repo owns your context"; progress.md as a dynamic-programming memo table that caches solved steps and prunes failed branches across compaction).docs/32-reading-list.md— new group Self-Improving Harnesses (AHE anchor + Self-Harness).SOURCES.md— arXiv research feed + four repos (peterCheng123321/loop-engineering, Sungmin-Cho/claude-deep-loop, shouryasrivastava/ctxcarry, the-open-engine/zeroshot).
Changed¶
docs/04-verification.md— added information-asymmetry / blind validation (zeroshot: the checker sees only the output, never the maker's reasoning) and run-record-anchored capture gate (orobsonn refinement — the passing-run evidence itself is no longer forgeable).LOOP_ENGINEERING.md— refreshed docs 24 / 16 / 04 summaries and the reading-list row.KB_GAPS.md— closed the self-scaffolding gap; opened "held-out eval construction for harness evolution".
[2.5.6] — 2026-07-01¶
Changed¶
docs/index.md(site home) — numbered the chapter/section list and the "Stay current" links to match the rest of the site (1. Foundations … 7. Reference, 8.x Project), so the home page content and its table of contents show numbers too. The home page is hand-authored and is skipped bysection-numbering.js(it has no nav number), so these numbers are set directly in the page. Also pointed the Stay-current links to the on-site News/Sources/Changelog pages.
[2.5.5] — 2026-07-01¶
Fixed¶
.github/workflows/docs.yml— the 2.5.4 build failed at action resolution becauseastral-sh/setup-uvhas no floatingv8major tag (only up tov7, though av8.2.0point release exists). Pinned toastral-sh/setup-uv@v7(also Node 24). Other actions (checkout@v5,upload-pages-artifact@v5,deploy-pages@v5) resolve fine.
[2.5.4] — 2026-07-01¶
Changed¶
.github/workflows/docs.yml— bumped GitHub Actions to their Node 24 majors (actions/checkout@v5,astral-sh/setup-uv@v7,actions/upload-pages-artifact@v5,actions/deploy-pages@v5), clearing the "Node.js 20 is deprecated / forced to run on Node.js 24" warnings. (Superseded by 2.5.5 — the v8 pin here was invalid.)
[2.5.3] — 2026-07-01¶
Added¶
- Numbered page titles and section headings on the docs site — each page's H1 is prefixed with its chapter number (e.g. "1.1 The Paradigm Shift") and its sections/subsections are numbered ("1.1.1 AI Leverage Formula", "1.1.1.1 …"), derived from the nav via
docs/javascripts/section-numbering.js. The right-hand table of contents is kept in sync. Done in JS so the source Markdown, heading anchors, and GitHub rendering stay clean; meta pages (8. Project — news/sources/changelog) get a title number but their machine-generated headings are not renumbered.
[2.5.2] — 2026-07-01¶
Added¶
- Numbered navigation on the docs site — chapters (1–8) and sub-chapters (e.g. 2.1, 2.2 …) in the left nav.
- News, Sources, and Changelog are now pages on the site (
docs/news.md,docs/sources.md,docs/changelog.mdsymlink the root files) under a new 8. Project nav section, so they're readable in the 3-column layout.
Changed¶
- All README links now point to the Pages site (topic index, changelog, sources, news, and starter topics); source-only files (skill, script, plans) are shown as plain paths rather than GitHub links.
LOOP_ENGINEERING.mdtopic links now point to the Pages site (35 links →https://lucagattoni.github.io/Claude-Loops/<page>/) so the GitHub-rendered index routes readers to the site..claude/skills/fetch-loop-news/SKILL.md— new index rows must link the Pages URL (not the repo path) and register the page inmkdocs.ymlnav, keeping the index consistent as the KB grows.LOOP_ENGINEERING_NEWS.md— fixed 3 relative links to absolute Pages URLs so the digest builds cleanly as a site page.
Housekeeping: merged the stale loop-news-2026-06-30 branch (released as 2.5.1); deleted the merged feature/loop-news-tracker branch; restored CHANGELOG version headers (2.4.2, 2.4.3, 2.4.5, 2.4.6, 2.4.7) that earlier sequential edits had dropped.
[2.5.1] — 2026-07-01¶
Loop news run 2026-06-30 04:00 UTC — 16 findings (merged late from branch loop-news-2026-06-30; renumbered from the branch's 2.4.6, which today's releases had already used). Third consecutive GitHub-dominated wave, converging on cross-model maker/checker (Claude implements / Codex reviews) with structured VERDICT + dual stop conditions.
Added¶
SOURCES.md— added Happenmass/Cliclaw (107★) and firegnu/herdr-loop-lab as github sources; replaced the low-yieldacting_ongithub-search with a cross-model maker/checker query.docs/32-reading-list.md— added Gusto "no-process / trash-can method" case study (Eddie Kim, CTO) to Loops in Production.KB_GAPS.md— added two gaps: cross-model checker arbitration/model-selection; self-scaffolding model-generated harness (Ornith 1.0).
Changed¶
docs/04-verification.md— added Pattern 5: cross-model independence to Verifier Integrity (checker runs a different model than the maker; model-diversity defeats shared blind spots; structuredVERDICT: PASS/BLOCK+SUGGEST; dual stop = test exit 0 AND no reviewer BLOCK) and the isomorphic-perturbation check refinement (anti single-predicate reward-hacking); updated pattern count 3→5 and synthesis to "these five together" (Cliclaw, loope, herdr-loop-lab, forja, Strive_Engineering).docs/07-subagents.md— added "Independence has two axes" note distinguishing context-independence from model-independence, cross-linking to docs/04 Pattern 5.LOOP_ENGINEERING.md— updated the docs/04 index summary to list isomorphic-perturbation + cross-model independence.
[2.5.0] — 2026-07-01¶
Added¶
- Published documentation site (MkDocs Material, 3-column layout — left nav / content / right on-this-page TOC) mirroring the Claude-Warp setup, deployed to GitHub Pages via
.github/workflows/docs.yml(uv →mkdocs build --strict→ Pages).mkdocs.yml— Material theme, deep-purple/deep-orange palette with light/dark toggle, search, full nav grouped as inLOOP_ENGINEERING.md(Foundations → … → Reference).docs/index.md— site home (intro, design spine, grouped topic links, links to the live news digest/sources/changelog).requirements-docs.txt— pinsmkdocs-material.- External links open in a new tab with an external-link icon —
docs/javascripts/external-links.js(Materialdocument$hook, survives instant navigation; addstarget=_blank+rel=noopener noreferrer) anddocs/stylesheets/external-links.css(currentColor-masked icon; internal links untouched). site/added to.gitignore(build output is generated in CI, not committed).
Verified locally: mkdocs build --strict passes with zero warnings (all 120 intra-doc links/anchors resolve); 3-column layout and external-link new-tab+icon confirmed in-browser (23/23 external links marked, 0 internal false-positives).
[2.4.7] — 2026-07-01¶
Loop news run 2026-07-01 12:35 UTC — 9 findings. A gap-driven run: filled three documented KB gaps and added the first quantified harness>model evidence.
Added¶
docs/33-agent-security-hardening.md— Credential Rotation Mid-Session (Credential-Sentinel: verify-before-revoke cutover — promote→repoint→verify, revoke old only after verify passes, rollback on fail; 4-state classify with default-deny on unknowns; two human gates) [fills gap] + Runtime Policy Gating (omnigent: blast_radius, intent_gate default-deny, phase-scoped tool access) [partial SECURITY_MATRIX fill].docs/30-goal-engineering.md— A-Priori Goal-Cost Estimation (cobusgreylinggoal-cost: pattern-keyed pre-run token/budget forecast) [fills gap].docs/14-human-in-the-loop.md— Where to Place a Checkpoint (4 tests, ~80/20 split, override-rate calibration; MindStudio).docs/24-harness-patterns.md— quantified harness>model (LangChain: Terminal-Bench Top30→Top5, harness-only) + harness conformance testing (omnigent harness-bench).docs/04-verification.md— 70/30 human-LLM blended grading (Claire Vo) + proof-of-work demo artifacts (Simon Willison video-as-verification).docs/06-skills.md— agent-legible tools (--helpas embedded SKILL.md; Simon Willison).
Changed¶
docs/17-failure-patterns.md— enriched Context-drift withdetect_task_switch(mechanical mid-run goal-change detection).LOOP_ENGINEERING.md— updated summaries for docs 04, 06, 14, 24, 30, 33.KB_GAPS.md— closed credential-rotation and goal-cost gaps; re-scoped SECURITY_MATRIX to the loading mechanism.
[2.4.6] — 2026-07-01¶
Selective integration from affaan-m/ecc (224k★ multi-harness agent operator system), evaluated at 14/20 — verified against the actual skill files, not the marketing README.
Added¶
docs/04-verification.md— Eval Metrics: pass@k vs. pass^k: pass@k (≥1 success in k tries — capability; target pass@3 > 90%) vs. pass^k (all k succeed — unattended-safety; pass^3 = 100% for critical paths), and the three grader types (code-based / model-based / human, risk-tiered). Cross-referenced to the A/A baseline (code-based graders only for gates) and Loop Readiness Levels.SOURCES.md— added affaan-m/ecc as a github source to track.
Changed¶
LOOP_ENGINEERING.md— updated the docs/04 summary.
Not integrated (already covered / marginal): ecc's verification-loop (redundant with the self-verifying loop; lacks maker/checker) and /loop-start loop-type taxonomy (overlaps existing trigger + readiness taxonomies).
[2.4.5] — 2026-06-29¶
Loop news run 2026-06-29 04:02 UTC — 18 findings (a second GitHub-dominated wave of fresh Claude-Code loop harnesses converging again on verifier integrity + anti-self-grading).
Added¶
SOURCES.md— added krishddd/Strive_Engineering as a github source (provenance-bound SHA-citation verifier pattern).docs/32-reading-list.md— added Strive_Engineering to Reference Implementations (provenance-bound verification; group now at the 5-entry cap).
Changed¶
docs/04-verification.md— added a 4th Verifier Integrity pattern: provenance-bound claims — every assertion must cite a git SHA re-checked viagit cat-file, plus a majority-vote monitor council to block self-grading (krishddd/Strive_Engineering, kok1eee/flywheel, grapheneaffiliate/Harness); added a closing synthesis tying all four patterns to the Verifier-Theater cure.LOOP_ENGINEERING.md— updated the docs/04 index summary to list provenance-bound claims + majority-vote council.
[2.4.4] — 2026-06-28¶
Changed¶
scripts/run-loop-news.sh— Hardened the retry logic so it can't make things worse: (1) added--max-budget-usd 8so retries can't become an unbounded bill; (2) safe-to-retry guard — only retry when the failed attempt left no durable trace (tracked tree clean andHEADunchanged); if it already committed or made partial edits, stop and notify rather than re-running (avoids duplicate commits / tag-release collisions); (3) desktop notification (osascript) on final give-up and on every no-retry abort, so a failed daily run is never silent. Verified against clean-retry, HEAD-moved-abort, dirty-tree-abort, and all-fail paths.docs/09-headless-mode.md— Documented "only retry when a retry is safe" (cost cap, traceless-failure check, notify-on-give-up) and noted that true safe-to-retry requires idempotency inside the loop itself.
[2.4.3] — 2026-06-28¶
Changed¶
scripts/run-loop-news.sh— Added retry-with-backoff (3 attempts, 30s→90s) so a transient API drop (ECONNRESET/overloaded) no longer loses the whole scheduled run. An attempt counts as failed if the exit code is non-zero or a connection-level error marker appears in its output — covering the macOSscript(1)quirk where the child exit code is masked to 0. Each attempt is captured to its own typescript for error-scanning, then folded into the day log (live PTY tail preserved).docs/09-headless-mode.md— Documented the "exit code alone is not enough — scan the output too" hardening lesson (PTY masks exit code; transient drops print to output) with a retry snippet.
[2.4.2] — 2026-06-28¶
Loop news run 2026-06-28 16:25 UTC — 14 findings (GitHub-dominated: a wave of new Claude-Code loop harnesses converging on stop-condition rigor + verifier integrity).
Added¶
docs/04-verification.md— Verifier Integrity: Keeping the Check Unfakeable: external verifier (loop-kernel — the loop runs the real check, worker can't fake), mechanical gates vs. adjudicators (herdr-loop-lab), frozen content-hashed tests before implementation (claude-harness).docs/07-subagents.md— "Strong Eyes, Cheap Hands": cost-asymmetric DOER/CHECKER model allocation (cheap models write, top model judges at gates; the cheaper the orchestrator, the more the deterministic rails must carry the judgment).docs/17-failure-patterns.md— Verdict oscillation failure pattern (checker flip-flops / contradictory jurors) + oscillation-guard mitigation (jje).SOURCES.md— added uppifyagency/loop-kernel and orobsonn/claude-harness as github sources.
Changed¶
docs/27-loop-contract.md— extended the canonical Stop Condition Taxonomy with the three-exit-code reference implementation (loop-kernel: 0=pass / 3=score-flat / 2=cap) and the "score the goal, not the activity" refinement to the no-progress stop.docs/32-reading-list.md— added loop-kernel to Reference Implementations.LOOP_ENGINEERING.md— updated summaries for docs 04, 07, 17, 27.KB_GAPS.md— annotated the SECURITY_MATRIX and goal-cost gaps with this run's partial advances and refreshed search keywords.
[2.4.1] — 2026-06-26¶
Manual addition from Andrew Ng's "Loop Engineering for 0-to-1 Product Development" (The Batch).
Added¶
docs/14-human-in-the-loop.md— The Three Feedback Loops (Andrew Ng): agentic coding (minutes) / developer review (tens of min–hours) / user feedback (days), and the human "context advantage" that keeps the human loop necessary; cross-linked to the Inner/Outer Dual Loop.SOURCES.md— added The Batch (DeepLearning.AI) as an rss source.
Changed¶
docs/17-failure-patterns.md— enriched the Reward hacking row with the concrete GLM-5.2 case (agents fetched reference solutions from GitHub) and the rule-based-filter mitigation.docs/25-long-running-agents.md— reciprocal cross-ref distinguishing the Dual Loop axis (execution/strategy) from the three feedback loops (agent/developer/user).docs/32-reading-list.md— added Andrew Ng's article to the Getting Started group.LOOP_ENGINEERING.md— updated the docs/14 summary.
[2.4.0] — 2026-06-26¶
Findings-driven structural review of the KB after the 2026-06-26 news run (38 findings), plus codification of that review as a standing norm.
Added¶
.claude/skills/fetch-loop-news/SKILL.md— Phase 4c — Findings-Driven Structural Review: a mandatory post-run pass that reads the run's findings as a set and rethinks/restructures the KB (canonical-home designation, missing thesis, unrepresented primitive, centrality drift, merge/reorder). Establishes the loop-design spine (What / How / When / How much / How do you know it's done?) as the KB's central organizing principle, and instructs critically pressure-testing user direction rather than implementing it verbatim.docs/27-loop-contract.md— Stop Condition Taxonomy (canonical home): four categories — completion-check (only success stop) / budget / max-iterations / no-progress; rice-cooker problem; "verification is the completion check"; runtime-vs-design-time mapping. Reframed the contract intro around the five loop-design questions.docs/02-agent-loop-cycle.md— Two Lenses on Loop Primitives: functional (execution/verification/orchestration/observability) vs. mechanical (six building blocks); observability named as the prerequisite fourth primitive; runtime termination signals mapped to the stop-condition taxonomy.docs/24-harness-patterns.md— The Harness as an Org-Level Artifact (Karpathy): harness design decisions propagate to every loop the org runs.
Changed¶
LOOP_ENGINEERING.md— Intro reframed to make loop design the central act (five design questions); updated summaries for docs 02, 04, 24, 27.docs/04-verification.md— Framed verification as the completion-check stop; cross-ref to the canonical taxonomy.docs/30-goal-engineering.md— Goal stopping condition mapped to the completion-check category; cross-ref to the canonical taxonomy.
[2.3.14] — 2026-06-26¶
Changed¶
docs/24-harness-patterns.md— Added: Harness vs. Environment Engineering section (in-process vs. out-of-process controls; Approval-First / Curated Allow-list / Sandboxed Full-Auto patterns); extended Harness-Agnostic Projection with .apm/ primitive manifest (6 subdirectory types); extended omnigent section with compaction persistence, --resume, and spec reconstruction on resolve-missdocs/04-verification.md— Added: Self-Coverage Gate (RFC-0051, every scope item must have a verification artifact) and Traceability-Lint (scope→task→artifact evidence chain gate)docs/34-loop-patterns.md— Added: concrete STATE.md multi-loop example (PR Babysitter + CI Sweeper coexisting with acting_on fields); Per-Agent Heartbeat Coordination pattern (harnery .harnery/active/ with claim/commit/TTL guards)docs/01-paradigm-shift.md— Added: >80% Anthropic engineers on self-improving loops; five canonical agent workflow patterns (prompt chaining, routing, parallelization, orchestrator-workers, evaluator-optimizer)docs/32-reading-list.md— Added "Building Effective Agents" (Anthropic, Dec 2024) to Harness Design groupSOURCES.md— Added MindStudio Blog as html sourceKB_GAPS.md— Marked .apm/ spec format and multi-loop STATE.md coordination as filled; updated SECURITY_MATRIX search keywordsLOOP_ENGINEERING_NEWS.md— 38 new findings (run 2026-06-26 09:04 UTC)
[2.3.13] — 2026-06-26¶
Changed¶
scripts/run-loop-news.sh— Usescript -q -ato allocate a PTY, forcingclaude -pto flush output line-by-line; enablestail -f logs/loop-news-YYYYMMDD.logto show live progress
[2.3.12] — 2026-06-26¶
Changed¶
docs/09-headless-mode.md— Added: macOS LaunchAgent scheduling pattern (plist template,launchctlcommands, when to prefer over Routines)docs/28-routines.md— Updated comparison table: added LaunchAgent as a third column alongside headless and Routines; added rule-of-thumb for choosing between them
[2.3.11] — 2026-06-26¶
Added¶
scripts/com.luca.loop-news.plist— macOS LaunchAgent that runsrun-loop-news.shdaily at 05:00 local (= 04:00 UTC); registered vialaunchctl load ~/Library/LaunchAgents/com.luca.loop-news.plist
[2.3.10] — 2026-06-25¶
Changed¶
docs/04-verification.md— Added: "Surface" as canonical stopping verb; verification mode discipline (TDD/goal-based/visual-manual); Oracle Problem (~6% test precision / oracle leakage); Structured Critic Finding Taxonomy (6 categories: product_bug, test_bug, harness_bug, evidence_mismatch, contention, scope_conflict)docs/07-subagents.md— Added: Adversarial Reviewer Checklists (spec-stage 9 checks + implementation-stage 9 checks); Rationalizations Reviewers Must Refuse tabledocs/08-permissions.md— Added: ASK verdict and soft warning thresholds (ask_thresholds_usd); session-fires-first evaluation orderdocs/11-cost-control.md— Added: Operational Kill/Pause/Slow-Down Thresholds with concrete numeric criteriadocs/16-memory-patterns.md— Added: Three-Tier Document Lifecycle (.tenet/runs/ + .tenet/project/ + .tenet/knowledge/)docs/17-failure-patterns.md— Added: Fixing flakes with code; Over-Reach; Parallel Collision failure patternsdocs/24-harness-patterns.md— Added: Agent YAML Definition Schema (Omnigent-style, 15+ harnesses); Organizational Learning Stage (4th loop stage); Harness Update File Safety Contract (.upstream companion files)docs/27-loop-contract.md— Added: Cross-Run Memory Persistence (.loopflow/memory/); Gate Feedback Injection (failure reason to all agent prompts)docs/33-agent-security-hardening.md— Added: credbroker credential resolution pattern (no model exposure)docs/34-loop-patterns.md— Added: Three-Loop Onboarding Sequence (Daily Triage → PR Babysitter → Post-Merge Cleanup → CI Sweeper); Debt Audit Loop pattern; Docs Sync Loop patternSOURCES.md— Added: eugenelim/agent-ready-repo, JeiKeiLim/tenet, faisalishfaq2005/loopflowLOOP_ENGINEERING.md— Updated summaries for docs 4, 7, 8, 11, 16, 17, 24, 27, 33, 34
[2.3.9] — 2026-06-25¶
Changed¶
- All
docs/*.md— Added markdown hyperlinks to every external citation; allowner/repoattribution footers,repo:lines,@handles, and inline tool names (LangSmith Fleet, Opik, Sakana Fugu, Graphiti) now link to their official pages docs/16-memory-patterns.md— Pattern F expanded with full Graphiti architecture (27.9k★, arXiv:2501.13956): Episodes/Entities/Facts/Custom Types model, temporal invalidation mechanism, hybrid retrieval (semantic + BM25 + graph), installation, loop integration patterndocs/32-reading-list.md— Fixed structure: single "Loops in Production" section (Claude Tag + New Stack verification article); "Reference Implementations" at the end
[2.3.8] — 2026-06-25¶
Changed¶
docs/16-memory-patterns.md— Added Pattern F: Temporal Knowledge Graph (Graphiti/Zep — temporal entity state as complement to flat STATE.md)docs/23-fleet-engineering.md— Added Opik to Observability section (trace→regression test for fleet agents)docs/08-permissions.md— Added Reject+Replan Pattern (when a safety gate fires, agent replans rather than aborting)docs/04-verification.md— Added Production Trace to Regression Test section (Opik)SOURCES.md— Added @akshay_pachaar (x) and getzep/graphiti (github)LOOP_ENGINEERING.md— Updated summaries for rows 4, 8, 16KB_GAPS.md— Updated multi-loop STATE.md gap to note Graphiti as partial coverage; flat-file example still missing
[2.3.7] — 2026-06-25¶
Changed¶
docs/23-fleet-engineering.md— Added: Fleet Four Pillars (Delegate/Improve/Approve/Connect), F0-F3 fleet maturity levels, Fleet Economics cost attribution, Claw vs. Assistant identity choice (Cobus Greyling, Jun 2026)docs/30-goal-engineering.md— Added: GOAL.md schema, six canonical goal patterns (Tests Green, Migrate Module, Fix Bug, Refactor, Docs Update, Security Scan), G0-G3 readiness scoring (cobusgreyling/goal-engineering, Jun 2026)docs/04-verification.md— Added: Type A vs. Type B work classification, loop verdict taxonomy (6 verdicts), cross-run patterns (clean-room review, held-out test layer, cross-task defect ledger), belief state machine + R0-R5 risk levels, A/A baseline for verifier calibration (Jun 2026)docs/24-harness-patterns.md— Added: harness-agnostic projection + security-at-specification-stage, 8-phase DAG execution model + steer message taxonomy, meta-harness 3-tier policy hierarchy + harness-swap (Jun 2026)docs/27-loop-contract.md— Added: YAML-declarative loop definition, VERDICT: PASS gate, 2-layer budget ceiling, self-discovery loop pattern (Schedule→Discover→Build→Verify→Repeat) (Jun 2026)docs/08-permissions.md— Added: agent trust ramp (4-stage: read-only → summarise → hard limits → loop cap)docs/32-reading-list.md— Added cobusgreyling/goal-engineering and cobusgreyling/fleet-engineering to Reference Implementations; removed duplicate "Loops in Production" sectionKB_GAPS.md— Marked filled: verifier calibration (thalys/agent-ab), loop correctness testing (void2610, JeremyW1990); added new gaps: agentskills.io format, F0-F3 indicators, goal-cost estimationSOURCES.md— Added: cobusgreyling/goal-engineering, cobusgreyling/fleet-engineering, omnigent-ai/omnigent (github type); added acting_on claude loop github-search for multi-loop STATE.md coordination gapLOOP_ENGINEERING_NEWS.md— 2026-06-25 run: 20 new findings across 6 sources
[2.3.6] — 2026-06-24¶
Changed¶
.claude/skills/fetch-loop-news/SKILL.md— Reading list step 5b: explicitly cover GitHub repos with substantial documentation as eligible for the Reference Implementations group (was silently excluded by "full article" language)KB_GAPS.md— Added: SECURITY_MATRIX.md implementation mechanism gap (how does the agent load it at startup?)
[2.3.5] — 2026-06-24¶
Changed¶
.claude/skills/fetch-loop-news/SKILL.md— Reading list curation: updated group list to include "Reference Implementations" (prevents duplicate group creation on next run); note that repos belong there, not in article groupsdocs/24-harness-patterns.md— Ledger Closure: added applicability note (relevant for custom API orchestrators that manage the message array directly; not applicable to Claude Code CLI)KB_GAPS.md— Clarified "Recently Filled" archive label
[2.3.4] — 2026-06-24¶
Changed¶
docs/07-subagents.md— Confidence gate threshold (≥80%) reframed as a calibratable parameter, not a universal rule; note that 80% is the session-orchestrator reference implementation value
[2.3.3] — 2026-06-24¶
Changed¶
.claude/skills/fetch-loop-news/SKILL.md— Renamed Phase 4b from "Doc coherence review" to "Devil's Advocate KB Review"; expanded from 5 structural questions to 9 questions across two categories: (1) adversarial content review (internal contradictions, missing cross-refs, unverifiable claims, better placement, redundancy) and (2) structural coherence (grouping, staleness, progression, fragmentation)
[2.3.2] — 2026-06-24¶
Changed¶
.claude/skills/fetch-loop-news/SKILL.md— Phase 5c: add KB_GAPS.md to git add command (was being omitted from daily commits)docs/08-permissions.md— Added: cross-reference to docs/33 (Agent Security Hardening) as the OS-layer companiondocs/27-loop-contract.md— Clarified: /evolve and /reconcile are Skills to implement (not built-in Claude Code commands); link to docs/06docs/32-reading-list.md— Updated intro: "articles" → "articles, essays, and reference implementations" (covers GitHub repos now in the list)
[2.3.1] — 2026-06-24¶
Added¶
KB_GAPS.md— new file: iterative gap tracker; each daily run reads and updates this file to record thin KB areas and their targeted search keywords; drives convergent deepening across runs
Changed¶
docs/17-failure-patterns.md— Added: Infinite Fix Loop pattern (loop retries indefinitely without attempt cap; fix: hard cap of N attempts + escalation to human inbox)docs/33-agent-security-hardening.md— Added cross-reference to docs/08 (permissions/allowlists) as the software-layer companion to OS-level securitydocs/34-loop-patterns.md— Added note distinguishing loop readiness levels (per-loop operational trust) from developer maturity model (docs/20)SOURCES.md— Changed github-search URLs from HTML scrape to GitHub API JSON endpoints; updated type-reference description for github-search.claude/skills/fetch-loop-news/SKILL.md— Updated github-search handler to use JSON API; added step 6 KB gap tracking (create/update KB_GAPS.md each run; use gap keywords to drive targeted searches)
[2.3.0] — 2026-06-24¶
Added¶
docs/33-agent-security-hardening.md— OS-user-per-agent kernel isolation, four credential disposition types (Broker/Sidecar/Remove/Egress Firewall), SECURITY_MATRIX.md design, fail-safe secret gate (clem — jahwag/clem, Jun 2026)docs/34-loop-patterns.md— Seven named loop patterns (Daily Triage, PR Babysitter, CI Sweeper, Dependency Sweeper, Post-Merge Cleanup, Changelog Drafter, Issue Triage); L1/L2/L3 operational readiness levels; token cost benchmarks; multi-loop coordination rules with priority ordering and collision detection; auto-merge allowlist and path denylist (cobusgreyling/loop-engineering, Jun 2026)
Changed¶
docs/04-verification.md— Added: simplification-before-testing (Wave 4 inversion — simplify AI code before writing tests); verification of memory (revalidate stale GOAL.md/STATE.md entries before acting)docs/07-subagents.md— Added: synthesis as the non-delegable bottleneck (task forwarding anti-pattern); confidence-scored quality gates (≥80% threshold, suppress low-confidence findings)docs/08-permissions.md— Added: risk-tiered authorization by consequence (read/write/irreversible tiers); safety path denylist globs for sensitive filesdocs/10-fan-out.md— Added: scope-verified parallelism via Pre-Edit hooks; multi-loop coordination with STATE.md acting_on claimingdocs/11-cost-control.md— Added: token cost by loop pattern (noop 3-5K → action run 200-250K → CI Sweeper without early exit 5M/day); early exit rule as correctness requirementdocs/12-hooks.md— Added: exit code safety contract — exit 1 is treated as non-blocking warning (continues); always use exit 2 for denial; safe bash pattern with ERR trapdocs/13-context-management.md— Added: input governance pipeline (prefetch/snip/microcompact/collapse/autocompact); reactive compact with circuit breaker (20K reserved tokens, 13K early-warning, halt after 3 failures)docs/16-memory-patterns.md— Added: multi-backend task queue (Slack/Discord thread claiming, GitHub Issue label workflow); STATE.md wave recovery (resume from last completed wave on crash)docs/17-failure-patterns.md— Added: State Rot (acting on ghost state references), Verifier Theater (approval without evidence), Notification Fatigue (notifying on every run regardless of delta)docs/20-loop-maturity-model.md— Added: per-loop L1/L2/L3 operational readiness levels (distinct from developer maturity); default rule: all new loops start at L1docs/24-harness-patterns.md— Added: Unstable Components design axiom; Ledger Closure for interrupted tool calls; Five-Wave Execution Model with Wave 4 simplification pass; Runtime Republic vs. Constitutional Control Plane framingdocs/25-long-running-agents.md— Added: session watchdog and 2-hour hard session limit implemented at OS level (systemd); distinct from turn caps and budget capsdocs/27-loop-contract.md— Added: governed cross-session learning — /evolve (extract patterns after 5+ sessions) + /reconcile (convert to reviewable .claude/rules/ proposals); nothing auto-appliesdocs/32-reading-list.md— Added: harness-books.agentway.dev ("Harness Books") to Harness Design group; session-orchestrator (Kanevry) to new "Reference Implementations" groupSOURCES.md— Added: harness-books.agentway.dev (html), GitHub search for "loop engineering claude" (github-search), GitHub search for "claude code harness" (github-search)LOOP_ENGINEERING.md— Updated summaries for rows 4, 7, 8, 10, 11, 12, 13, 16, 17, 20, 24, 25, 27, 32; added rows 33, 34
Sources: jahwag/clem (Jun 2026), cobusgreyling/loop-engineering (Jun 2026), wquguru/harness-books / AgentWay (Jun 2026), Kanevry/session-orchestrator (Jun 2026).
[2.2.0] — 2026-06-24¶
Added¶
docs/32-reading-list.md— Curated best-articles collection; 14 articles across 5 groups (Why Loops / Getting Started / Harness Design / Goal Engineering / Production); dynamic — articles can be added/removed by future runsdocs/30-goal-engineering.md— Goals vs. Loops decision framework; four Goal Primitives (Objective, Verifier, GOAL.md State, Budget); GOAL.md pattern for persistent goal state (Cobus Greyling, Jun 2026)docs/31-claude-tag.md— Claude Tag: ambient loops in Slack; channel-scoped identity; self-scheduling; org-wide context; third LLM paradigm framing (Anthropic + Karpathy, Jun 2026)
Changed¶
docs/01-paradigm-shift.md— Added: compound probability argument (0.9^10 = 35%); era framing 2022→2023→2024–2026; "performance ceiling set by loop, not model"docs/04-verification.md— Added: Firefox harness case study (LLM file prioritization, score→fix→verify, 423 fixes in one month, 50% harness attribution)docs/17-failure-patterns.md— Added: amplification effect (defensive complexity accumulation); cognitive dependency (AI-only-legible codebases) — both Armin Ronacher, Jun 2026docs/27-loop-contract.md— Added: job-description framing (Claire Vo); Event Modeling for task decomposition with Never Argue rule (Martin Dilger, Jun 2026)
Sources: Cobus Greyling (Substack), Anthropic / Boris Cherny, Andrej Karpathy, Armin Ronacher (lucumr.pocoo.org), @roanbrasil (Medium), Martin Dilger (LinkedIn), Claire Vo / Lenny's Newsletter, Brian Grinstead — all Jun 2026.
[2.1.2] — 2026-06-23¶
Changed¶
docs/13-context-management.md— Added: Context resets vs. compaction comparison; context anxiety finding (Sonnet 4.5 vs. Opus 4.6)docs/07-subagents.md— Added: GAN framing for why external evaluation enables improvement; evaluator tuning anti-patternsdocs/04-verification.md— Added: Making subjective goals gradable (4-dimension framework)docs/11-cost-control.md— Added: Real project cost benchmarks ($9 broken / $200 working; DAW $124.70)docs/24-harness-patterns.md— Added: Three-agent full-stack harness; sprint contract system; load-bearing vs. optional components / re-baseline per model release
Source: Prithvi Rajasekaran, Anthropic Engineering, "Harness Design for Long-Running Application Development", Mar 2026.
[2.1.1] — 2026-06-23¶
Changed¶
docs/17-failure-patterns.md— Added: Reward hacking, Context pollution + Context reset pattern, Context driftdocs/24-harness-patterns.md— Added: Harness vs. Loop two-layer distinction, "Verification closure → reliability → scalability", Event-Driven Architecture (EDA) loops, Serverless loopsdocs/27-loop-contract.md— Added: Two Quality Gates (evidence completeness, stopping condition clarity), Experience Encoding as post-iteration learning step
[2.1.0] — 2026-06-23¶
Added¶
docs/28-routines.md— Routines: cloud-hosted loop execution with Schedule/API/GitHub triggersdocs/29-background-agents.md— Background agents:--bg, agent view, fan-out, worktree isolation
Changed¶
docs/12-hooks.md— Full rewrite: 5 hook types, complete lifecycle events, JSON output API, asyncRewake circuit breaker, conditionaliffield, env vars, scope hierarchydocs/09-headless-mode.md— Full rewrite: session continuation, background sessions, prompt overrides, CI flags, prompt cache optimisationdocs/07-subagents.md— Added built-in types table (fork/Explore/Plan), custom agent frontmatter, nesting depth, permission controldocs/08-permissions.md— Added deny/ask lists, Tool(param:value) pattern syntax, PermissionRequest hook, settings precedencedocs/05-claude-md.md— Added load hierarchy, path-scoped rules (.claude/rules/), HTML comment stripping, import syntax,claudeMdExcludesdocs/03-building-blocks.md— Added Routines as cloud automation layer alongside local Automationsdocs/25-long-running-agents.md— Added "Detaching from the terminal" section with--bgand Routines pointersLOOP_ENGINEERING.md— Added rows 28 and 29; updated summaries for rows 3, 5, 7, 8, 9, 12
[2.0.2] — 2026-06-23¶
Changed¶
docs/07-subagents.md— added DOER/CHECKER pattern: never let the AI grade its own outputdocs/01-paradigm-shift.md— added AI Leverage Formula: AI Leverage = Clarity × Skill (Sabrina Ramonov, Jun 2026)docs/25-long-running-agents.md— added Inner/Outer Dual Loop pattern; @samwillis real-world /goal example (1k commits, 10 days)docs/27-loop-contract.md— added stopping condition aphorism: "if you can't say what done looks like, you don't have a loop"docs/11-cost-control.md— added token consumption benchmarks (~4× single agent, ~15× multi-agent vs standard chat)LOOP_ENGINEERING.md— updated index summaries for rows 7, 11, 25
[2.0.1] — 2026-06-22¶
Changed¶
docs/17-failure-patterns.md— updated cognitive surrender with empirical data (<40% vs 65%+ comprehension, Osmani May 2026); added "Dark factory" and "Missing circuit breaker" failure patternsdocs/21-context-vs-loop-engineering.md— added "The four disciplines" section: Loop/Context/Harness/Fleet Engineering as four named disciplines (Cobus Greyling, Jun 2026)LOOP_ENGINEERING.md— updated index summaries for rows 17 and 21 to reflect new content
[2.0.0] — 2026-06-22¶
Changed¶
LOOP_ENGINEERING.md— restructured flat 27-row table into 7 logical sections (Foundations, Designing a Loop, Components, State & Long-Running Loops, Quality & Safety, Scaling, Reference); rows reordered within sections but numbers preserveddocs/16-memory-patterns.md— added cross-reference to25-long-running-agents.mddocs/01-paradigm-shift.md— added cross-reference to26-factory-model.md.claude/skills/fetch-loop-news/SKILL.md— added Phase 4b doc coherence review; release criteria MAJOR condition now explicitly covers index restructuring
[1.2.0] — 2026-06-22¶
Added¶
docs/20-loop-maturity-model.md— 14-step progression from manual prompter to loop engineer (Boris Cherny's 3-stage model + community roadmap, Jun 21 2026)docs/21-context-vs-loop-engineering.md— emerging community debate on whether context engineering supersedes loop engineering (@techtasium, Jun 21 2026)docs/22-learned-orchestration.md— new concept: training the orchestrator (Sakana Fugu, TRINITY/Conductor, Thinker/Worker/Verifier roles) vs. coding it by hand (Jun 22 2026)docs/23-fleet-engineering.md— new concept: managing fleets of AI agents at enterprise scale (Cobus Greyling, LangSmith Fleet, Jun 22 2026)docs/24-harness-patterns.md— new doc: two-part Anthropic harness (initializer + coding agent); four-type loop taxonomy (heartbeat/cron/hook/goal) (Jun 22 2026)docs/25-long-running-agents.md— new doc: Ralph loop, planner-worker-judge, cross-context-window state management and git-based recovery (Jun 22 2026)docs/26-factory-model.md— new doc: AI software factory framing — spec quality and verification replace coding speed as the bottleneck (Jun 22 2026)docs/27-loop-contract.md— new doc: TRIGGER/SCOPE/ACTION/BUDGET/STOP/REPORT; Anchor File Pattern; Uber annual-budget-in-4-months data (Jun 22 2026)SOURCES.md— added @Sabrina_Ramonov X handle (active loop engineering presence confirmed Jun 22 2026)SOURCES.md— added Lenny's Newsletter (Claire Vo) html source (Jun 22 2026)
Changed¶
docs/17-failure-patterns.md— added "Loop as wrong unit" anti-pattern (@batjko_labs, Jun 21 2026)docs/17-failure-patterns.md— added "Cost runaway", "Provider lock-in", "Cognitive surrender", "Orchestration tax", and "Intent debt" failure patterns (Jun 22 2026)docs/02-agent-loop-cycle.md— added Universal Agent Thesis ("Perceive, reason, act, learn") framing with "Learn" step explained (Jun 22 2026)docs/06-skills.md— added "Skills as SDLC Scaffolding" section: non-skippable engineering phase enforcement (Jun 22 2026)docs/01-paradigm-shift.md— added "New Software Lifecycle" table: implementation speed no longer the bottleneck (Jun 22 2026)LOOP_ENGINEERING.md— added rows 22–27 (learned-orchestration, fleet-engineering, harness-patterns, long-running-agents, factory-model, loop-contract)
[1.1.0] — 2026-06-21¶
Added¶
docs/19-mcp-security.md— AgentJacking attack pattern, indirect prompt injection via MCP tool results, mitigations (sourced from The New Stack, Jun 21 2026)- Four-tier keyword taxonomy in
SOURCES.md— Tier 1 (Boris Cherny's exact language), Tier 2 (named discipline), Tier 3 (named concepts/failure modes), Tier 4 (tool/feature names) - New sources: Sabrina Ramonov (
sabrina.dev), Cobus Greyling (Substack), Peter Steinberger (@steipete)
Changed¶
docs/17-failure-patterns.md— added "Polling loop" anti-pattern: using cron when an event-driven trigger would be more token-efficient (sourced from @CKGrafico, X.com, Jun 21 2026)SOURCES.md— corrected Anthropic RSS to/rss.xml, OpenAI changed tohtmltype (RSS 403), swyx.io path corrected to/writing, Addy Osmani RSS corrected to/rss.xml.claude/skills/fetch-loop-news/SKILL.md— strategy corrected: per-source phase searches within source for keywords; general search added as bonus pass (Phase 3); tiered scoring in result JSON and digest
[1.0.0] — 2026-06-21¶
Added¶
LOOP_ENGINEERING.mdas a slim index table — one row per topic with summary and linkdocs/01-paradigm-shift.mddocs/02-agent-loop-cycle.mddocs/03-building-blocks.md— covers Automations, Worktrees, Skills, Chrome connector, Sub-agents, Memorydocs/04-verification.mddocs/05-claude-md.mddocs/06-skills.mddocs/07-subagents.mddocs/08-permissions.mddocs/09-headless-mode.mddocs/10-fan-out.mddocs/11-cost-control.mddocs/12-hooks.mddocs/13-context-management.mddocs/14-human-in-the-loop.mddocs/15-explore-plan-implement.mddocs/16-memory-patterns.mddocs/17-failure-patterns.mddocs/18-quick-reference.mdSOURCES.md— dynamic source list for the daily news trackerLOOP_ENGINEERING_NEWS.md— append-only daily digest log.claude/skills/fetch-loop-news/SKILL.md— 3-phase daily news skillscripts/run-loop-news.sh— headless runner script