[choiyounggi]

Knowledge flush — 4 claimed rows: 2 merged into existing pages, 1 new page, 1 anhinga plan-gap kept local-layer

Run 20261011-213417-11470 (the headless auto-flush child; the lock owner id arrived in DEV_LOOP_FLUSH_RUN_ID, so step 0's acquire with a fresh id reported held … 24s and the re-run with the inherited id reported already-owned). Claimed 4 rows — every row in the queue: 5e9bdfb102f22421 (qa), eec9adde8f6cf9c3 (infrastructure), cb99014ed0e30451 (qa), 73c466a4b014e5c2 (anhinga plan-gap). ~/.dev-loop/repo was clean before the reset (git status --porcelain empty; the previous branch's commit is PR #270), so nothing needed salvaging.

Verified best-practice

Row	Claim	Sources checked (fetched 2026-10-11)	How verified	Confidence
5e9bdfb1	A count taken from an agent's, worker's or script's own summary ("24 of 24 URLs returned 200") is recounted from the raw output that produced it, with the denominator named; when that output is gone or the figure cannot be rebuilt from it, publish the lines and say the count was not reproduced	https://docs.claude.com/en/docs/test-and-evaluate/strengthen-guardrails/reduce-hallucinations (HTTP 200; "Verify with citations: … If it can't find a quote, it must retract the claim.")	The row's own field evidence was recounted from the originating transcript (~/.claude/projects/…/2af9c97c-….jsonl): gh pr view 270 body line 21 reads "24 of 24 distinct source URLs returned HTTP 200" and "the 8 remaining URLs in the diff are … fixture URLs"; jq over every Bash tool_use found one command containing http_code (the loop over urls-recovered-20261011.txt) and zero Agent calls; its tool_result has 32 lines starting with a 3-digit code: 21 200, 11 000, 0 other; all 11 000 hosts are example.com/evil.test fixtures, and https://example.com/ + https://evilexample.com/ (both 200) are fixture rows on #270's host-disguise-fixtures-for-url-allowlists page (lines 40, 61, 69) — so the output supports 19 source URLs all 200, and neither 24 nor 8 is derivable from it; the first "24 of 24" in the transcript (line 424) sits inside the run's own INGEST_REPORT heredoc	verified (official doc + the field case recounted here)
eec9adde	Before an automated job's checkout main + reset --hard origin/main on a reused checkout, inspect the tree and preserve leftovers; new material: inputs leased with a TTL come back on their own (the processing does not), and a leftover draft with cited sources is re-verified by re-fetching its URLs, re-running lint and a sample of its reproductions before adoption	https://git-scm.com/docs/git-reset, https://git-scm.com/docs/git-apply, https://git-scm.com/docs/git-status (all HTTP 200; the page's three quotes re-read verbatim)	Same trigger and directive as the existing page (see Existing-layer check); the candidate's own git diff HEAD wording was not imported because the page's Instead-of row already corrects it (untracked files are left out). Field figures recounted here: ~/.dev-loop/salvage/knowledge-20261011-132856.patch holds 23 diff --git headers; git show --stat e43e647 → 23 files changed, 421 insertions, 95 deletions; "five reproductions re-run" is the 19:29 run's own report (PR #270 body line 28) and is attributed as such on the page	verified (merged as Edge rows + a field-case Sources line)
cb99014e	page.setContent is document.write() on the existing document, not a navigation, so an init script registered on the page is never evaluated for it; serve the fixture through page.route + page.goto. A goto('about:blank') before setContent keeps the script's globals but document.open() erases its listeners	https://playwright.dev/docs/api/class-page (HTTP 200; addInitScript "Whenever the page is navigated … after the document was created but before any of its scripts were run"; setContent "This method internally calls document.write(), inheriting all its specific characteristics and behaviors"), microsoft/playwright#13274 (via gh api; maintainer pavelfeldman, 2022-04-03: "page.setContent does not run the init scripts, because JS world is not re-created, only navigation such as page.goto re-creates it"), https://html.spec.whatwg.org/multipage/dynamic-markup-insertion.html (HTTP 200; document open steps: "For each shadow-including inclusive descendant node of document, erase all event listeners and handlers given node")	Local matrix ~/.dev-loop/scratch/flush-20261011-2134/probe-setcontent.js (playwright-core 1.63.0 from the originating worktree's node_modules, Chromium 153.0.8010.12 headless shell): A page.addInitScript + setContent → __probe undefined, inline script saw undefined, __events undefined; B + route/goto → 1, number, ["animationstart"]; C goto('about:blank') then setContent → 1, number, []; D context.addInitScript before newPage() then setContent → 1, number, []; E no init script + route/goto (control) → undefined throughout; F setContent twice → undefined. Known-good (B) and known-bad (A, E) were both observed, so the probe itself is proven	verified
73c466a4	anhinga docs/backends.md section-8 capability table (Korean text, lnpl outbox/relay rows, lnpl-postgres#8)	n/a — not researched as general practice	Fails the layer test: the directive names one project's document sections, modules and issues	n/a → local layer

Nothing was upgraded on the originating session's say-so: the two statements the queue rows carried that could not be recounted here ("five reproductions"; the harvester's note that its blog draft omitted the figure) are attributed to their report or left out.

Existing-layer check

Route: INDEX.md → qa ("evidence for completion claims", "verifying the quantitative claims in a document before publishing it") and infrastructure ("multi-agent orchestration … shared run state") → wiki/qa/index.md, wiki/infrastructure/index.md; wiki/testing/index.md read for the Playwright candidate (e2e / mocking categories).

Pages read: qa-process-completion-claims, qa-deliverables-quantitative-claims-in-a-published-document, infrastructure-agent-orchestration-resetting-a-reused-checkout-after-an-interrupted-run, qa-environments-browser-console-capture-gaps, testing-quality-markup-contract-assertions, testing-e2e-e2e-stability, testing-quality-harness-reverse-controls

wiki_search (k=5) per candidate, plus a whole-wiki grep for each candidate's key terms:

Row	Top-5 wiki_search hits	Whole-wiki grep	Result
5e9bdfb1	qa-process-completion-claims 0.721 (Instead-of row "Relay a subagent's 'task completed successfully'"), qa-deliverables-quantitative-claims-in-a-published-document 0.692, qa-document-verification-model-coupled-guidance-aging-detector 0.688, testing-quality-surviving-mutant-equivalence-triage 0.682, infrastructure-observability-logs-metrics-signals 0.680	`recount	restat
eec9adde	infrastructure-agent-orchestration-resetting-a-reused-checkout-after-an-interrupted-run 0.815 (trigger) and 0.786 (edge), infrastructure-agent-orchestration-worktree-isolated-workers 0.777, qa-process-scope-purity-checks 0.769, infrastructure-agent-orchestration-code-graph-as-orientation-layer 0.767	`reset --hard	--porcelain
cb99014e	testing-quality-markup-contract-assertions 0.752 (Playwright component tests → computed style; a different trigger), qa-environments-browser-console-capture-gaps 0.751 (trigger) and 0.743 (directive "Register on the page object before calling goto"), platforms-tools-version-keyed-artifact-cache 0.746, platforms-tools-unpacked-extension-source-reload 0.735	`setContent	addInitScript

Index rows: wiki/qa/index.md completion-claims row extended and one environments row added at the end of that table; wiki/infrastructure/index.md resetting row extended (maintenance invariant 1). Checks on this branch: node scripts/wiki-lint-prohibitions.js wiki → compliant: 84, violations: 0, info: 1; node scripts/wiki-structure-checks.js wiki → pages: 427, indexes: 13, findings: 0 (baseline on origin/main f7c865c: 84 / 0 / 1 and 426 / 13 / 0); body lines 66 / 63 / 68 (limit 120); no banned qualifier in any added directive (the one hit, "probably fixed", is the page's pre-existing quoted hedge-word example); every related: id resolves on main.

Open-PR check

Open knowledge/* heads from gh pr list --repo choiyounggi/dev-loop --state open --json number,headRefName filtered on the head prefix: #265 knowledge/choiyounggi-20261009-222812, #266 knowledge/choiyounggi-20261010-021455, #268 knowledge/choiyounggi-20261011-002014, #269 knowledge/choiyounggi-20261011-022522, #270 knowledge/choiyounggi-20261011-193253 (all OPEN, none merged; 6 open PRs in total). Each head was fetched and the added lines of git diff origin/main...origin/<head> -- wiki/ were scanned for every candidate's terms (recount|restat|raw output|denominator|pass count|…, reset --hard|porcelain|salvage|abandoned|uncommitted|…, setContent|addInitScript|init script|route.fulfill|page.route|…), and each matching line was read.

Row	Overlapping open PR	Verdict
5e9bdfb1	#268 extends quantitative-claims for a count edited inside one document (a breakdown keeps the old total) — a different situation; #266's done-criteria page matches only on "restated"; no head touches completion-claims	new (merged into a page no open PR edits)
eec9adde	#265's review-diff-base page and scope-purity-checks mention uncommitted files for other reasons; no head edits the resetting page; #265/#266 edit wiki/infrastructure/index.md at lines 35/49 and 30, not the resetting row (line 28)	new (merge)
cb99014e	#270's new proving-nothing-moves-under-reduced-motion page serves its fixture "through page.route + page.goto" and says an init script "re-runs on each navigation", but never states that setContent skips the init script, nor the listener erasure — it is the field origin of this candidate, not a carrier of its lesson. No other head mentions setContent/addInitScript	new (a related: link to #270's page is deferred: its id does not exist on main yet)
73c466a4	no open PR body mentions anhinga or t1-driver-capability-contract	n/a (local layer)

Nothing folded, nothing dropped as a pending duplicate. git merge-tree --write-tree of a temporary commit of this tree against each of the five heads: every wiki page and both indexes merge cleanly; the only conflict with each head is log.md, the append-only log every knowledge PR conflicts on with every other.

Routing decision

Row	Layer	Domain / category	Page	Kind
5e9bdfb1	bundled	qa / process	wiki/qa/process/completion-claims.md	merge (the candidate's domain hint was qa; the page owns "what evidence a completion claim requires" and already has the "relay a subagent's report" row this generalizes)
eec9adde	bundled	infrastructure / agent-orchestration	wiki/infrastructure/agent-orchestration/resetting-a-reused-checkout-after-an-interrupted-run.md	merge
cb99014e	bundled	qa / environments	wiki/qa/environments/init-scripts-skipped-by-setcontent.md	new page (candidate hint qa; the category already frames "reading a browser tool's measurement as a QA verdict"; testing/e2e was considered and rejected because the page is about a probe's load path, not test structure)
73c466a4	local (anhinga)	—	see Local-layer candidates	excluded, retired

No new category.

Local-layer candidates

Row	Project	Target
73c466a4 — Planning t1-driver-capability-contract: the docs/backends.md contract table	anhinga (/Users/choeyeong-gi/Desktop/workspace/linkly/anhinga)	wiki-local/backend/common/api-design/t1-driver-capability-contract-table.md — run wiki-ingest inside that project. Its research notes (JDBC DatabaseMetaData.supportsXxx, PEP 544 runtime_checkable, the transactional-outbox relay, the IETF Idempotency-Key draft) carry a general kernel — a driver contract self-reports optional capabilities through predicates, and a conformance kit reports declared-unsupported rather than skip — that no bundled page owns (the row's own semantic search found none); it was not harvested as a candidate and is noted here as a gap, not ingested

Notes for the reviewer

• Step 0 of the skill generates a fresh run id; inside the auto-flush child the lock is already held under the inherited DEV_LOOP_FLUSH_RUN_ID, so the fresh-id acquire reports held (the bundled page inherited-lock-ownership-in-a-spawned-session covers this). This run re-acquired with the inherited id and releases with it.
• Separately: PR knowledge: reduced-motion start counts, full-page overflow width, server-spawned agent CLI env/cwd, + 10 recovered (TZ zone loop, dispatch-time calls, host-disguise URLs, console spies, favicon 404, Vitest substring filters) #270's body line "24 of 24 distinct source URLs returned HTTP 200" is the figure this flush's first candidate recounts as 19 source URLs (all 200) plus 13 fixtures. That body is not edited here.

🤖 Generated with Claude Code