knowledge: 3 insights — greedy-regex adjacent tokens in extractor fixtures (1 new page), subagent model-scoped 429 → per-invocation model override (+1 merge, 2 same-batch duplicates folded, Sonnet-limit drift fixed)

[choiyounggi]

Knowledge flush — 3 insight(s)

3 claimed candidates → 2 distinct insights: 1 new page, 1 merge into an existing page (2 duplicate candidates folded into it), 0 drops.

Queue id	Insight	Outcome
8c0693fbeb1dca30	Greedy regex merges adjacent tokens in extractor fixtures	New page testing-data-adjacent-tokens-in-extractor-fixtures
fe22774ea41d523f	Subagent 429 on a model-scoped (Fable) limit → per-invocation model override	Merged into infrastructure-agent-orchestration-usage-limit-paused-workers
2f27db59be18d9c2	Same insight as fe22774ea41d523f, from a second session	Duplicate within this batch; used as the second field observation on that merge

Verified best-practice

1. Adjacent tokens in fixture text for a greedy pattern extractor (8c0693fbeb1dca30) — confidence: verified

• Claim: a regex whose repeated character class admits separators (\d[\d\s.\-()]{6,18}\d) greedily merges two occurrences separated only by class-member characters into one span capped at the repeat's maximum, whose digits match neither original. So the test should assert the extractor's own spans first, and per-occurrence fixtures should use a separator outside the class.
• Sources checked:
  • https://docs.python.org/3/library/re.html — "The '*', '+', and '?' quantifiers are all greedy; they match as much text as possible"; re.finditer returns non-overlapping matches, scanned left-to-right.
  • https://developer.mozilla.org/en-US/docs/Web/JavaScript/Reference/Regular_expressions/Quantifier — "Quantifiers are greedy by default … until the maximum is reached"; lazy ? matches as few times as possible.
• How verified: reproduced with CPython re this session.
  • "- 1900068889\n- 1900068889" → one match '1900068889\n- 1900068' (20 chars = 1 + 18 + 1).
  • Pipe-separated → two clean matches.
  • Real-shaped footer "090 123 4567 028 3822 1234" → one merged match.
  • Also reproduced and added as edge cases:
    • Unbounded + merges both numbers (20 digits).
    • Lazy {6,18}? truncates 090 123 4567 to 090 123 4, and pipe-separated 1900068889 to 19000688.
• The session's field evidence (independent test-quality auditor caught it; pipe-delimiting fixed it) matches the reproduction.

2. Subagent failing on a model-scoped limit → per-invocation model override (fe22774ea41d523f + 2f27db59be18d9c2) — mechanism verified; the Fable-specific 429 text is field-tested

• Claim: when a subagent call fails at once with rate_limit/HTTP 429 naming one model (claude-fable-5-1), re-issue the identical call with model: "sonnet". The subagent definition's model is independent of the session's model, and the limit is scoped to that model.
• Sources checked:
  • https://code.claude.com/docs/en/sub-agents#choose-a-model (raw .md fetched). Resolution order: "1. The per-invocation model parameter 2. The subagent definition's model frontmatter 3. CLAUDE_CODE_SUBAGENT_MODEL 4. The main conversation's model". Before v2.1.251 the env var ranked first.
  • https://code.claude.com/docs/en/errors (raw .md fetched): "The session and weekly limits are shared across all models … The Opus and Sonnet limits each apply only to requests to that model family, so switching to a model outside the family with /model keeps you working". Also the Fable limit reached · continuing on Fable 5.1 uses usage credits … nothing was sent message for unattended sessions.
  • https://code.claude.com/docs/en/model-config#fable-and-usage-credits — Fable usage can bill to usage credits behind a consent prompt.
  • Local check: dev-loop 1.21.0 agents/test-quality-auditor.md and agents/integration-reviewer.md both declare model: fable. That explains why only the subagent failed while the session kept working.
• Honest gap: the docs do not print the exact You've reached your Fable limit … model sent to the API 429 text. That wording is cited only as a field observation (2 independent sessions on 2026-09-14, both of which succeeded after the override).
• The page stays verified because its directive rests on the documented resolution order and model-family scoping.

Existing-layer check

Pages read: infrastructure-agent-orchestration-usage-limit-paused-workers, testing-quality-tests-that-cannot-fail, testing-data-test-data-and-isolation

Also read: INDEX.md, wiki/testing/index.md (every quality/data/strategy load-when line), and the agent-orchestration rows of wiki/infrastructure/index.md. I also grepped wiki/ for greedy|regex|finditer|adjacent, rate.?limit|429, and model override|subagent.*model|fable.

• Insight 1: no existing page covers greedy-quantifier merging in test fixtures.
  • Grep hits were about regex gates, not extraction fixtures: source-text-wiring-assertions, completion-predicates, checks-that-cannot-pass.
  • → created new.
  • Related links added both ways with testing-data-test-data-and-isolation (fixture construction) and testing-quality-tests-that-cannot-fail (a test green for the wrong reason).
• Insight 2: usage-limit-paused-workers owns model-scoped limits and the /model switch → merged. Changes:
  • New step 7 (subagent per-invocation override).
  • 4 new edge-case rows: override rerun fails again; Fable consent in unattended sessions; CLAUDE_CODE_SUBAGENT_MODEL on CLI < v2.1.251; auditor model substitution + availableModels.
  • 1 new Instead-of row, and 2 new sources plus a field observation.
  • Index load-when extended.
• Drift corrected (not a contradiction): the page said only the Opus limit is model-scoped. The current errors doc also lists You've hit your Sonnet limit, scoped to the Sonnet family. I updated the marker row, the "only one worker stopped" edge row, and the Instead-of "Why" cell, and bumped last_verified to 2026-09-14.
• No conflicting directive found.

Lint results on this branch:

• node scripts/wiki-structure-checks.js wiki → pages: 277, indexes: 13, findings: 0.
• wiki-lint-prohibitions.js → 2 violations. Both predate this PR and sit outside wiki/ (plans/harvest-dedupe-processed/..., tests/fixtures/prohibitions/bad.md). Neither changed page is flagged.
• wiki-lint-model-era.js:
  • usage-limit-paused-workers is reported as model-coupled, no verified_model. It was already flagged that way on main, since the page mentions Claude/Opus and no wiki page carries verified_model.
  • I tried verified_model: claude-opus-5 (the model this flush ran on). The lint rejected it as "not in current set" because the script's DEFAULT_CURRENT is still ['opus-4', 'fable-5'].
  • I left the field off to match every other page, rather than picking a value just to pass. Refreshing DEFAULT_CURRENT is a separate owner decision.

Open-PR check

Listed open knowledge/* heads with gh pr list --repo choiyounggi/dev-loop --state open --search "head:knowledge/": #190, #189, #188, #187, #186, #185, #183, #182, #181, #180, #179. For each, I diffed its own changes with git diff origin/main...origin/<head> -- wiki/ and grepped the added lines for greedy|finditer|adjacent|regex merge|model-scoped|Fable limit|subagent model|per-invocation. No added lines matched.

• Insight 1 (8c0693fbeb1dca30) — verdict new.
  • No open head adds a regex/extractor-fixture page.
  • The closest is knowledge: 12 insights — workflow file that registers no triggers, workflow-authored PRs (create-PR setting, personal-repo ruleset, GITHUB_TOKEN checks, --auto), Firecrawl main-content vs contact details, testcontainers community namespace, pointer-attracted particle collapse, producer wire format in mocks (+5 merges incl. Homebrew clang SDKROOT contradiction reconciled) #187's backend/common/integrations/contact-details-from-scraped-pages.md, which covers where to read phone numbers (tel: links, header/footer vs main content), not how a regex tokenizes adjacent numbers. That's a different trigger, so no overlap.
• Insights 2/3 (fe22774ea41d523f, 2f27db59be18d9c2) — verdict new (merge into the main-branch page).
  • knowledge: 11 insights — login-expired panes, O_APPEND blackboard, numbered-protocol grafts, --ours on mixed conflicts, unittest floor false positives, SwiftUI sheet gating, silent hotkey registration, rolled-back-run assertions (+3 merges) #180 touches usage-limit-paused-workers.md, but only its related: line (it adds login-expiry-during-unattended-turns). No content overlap.
  • This PR leaves that related: line unchanged, so the two PRs should not conflict on that line.
  • 2f27db59be18d9c2 duplicates fe22774ea41d523f within this batch, so it is folded into the same merge rather than dropped.
• Merge-order note: knowledge: 21 insights — fork-PR secret gates, review-bot exit codes, dict-subclass copy loss, iterative DFS, citation sweep, plugin-prose identity leak, xcodegen regen loss, narration ordering tests, autouse shadowing, stub contracts, real-CLI spot checks (+8 merges, 1 fold onto #183) #188, knowledge: 12 insights — dropzone copy vs drop handlers, destination-in mask chaining, media-query inset reset, spatial clamp, env restore vs pop, synthetic-corpus floor, plan-claim recompute, multi-name command -v, alert suppression key, sibling validators (+2 folds) #181, and knowledge: 12 insights — deny rules under bypass, merged-tree gate, worker verify command, Kotlin daemon heap, extracted-method this (+7 merges) #179 also edit testing/data/test-data-and-isolation.md, and knowledge: 21 insights — fork-PR secret gates, review-bot exit codes, dict-subclass copy loss, iterative DFS, citation sweep, plugin-prose identity leak, xcodegen regen loss, narration ordering tests, autouse shadowing, stub contracts, real-CLI spot checks (+8 merges, 1 fold onto #183) #188/knowledge: 9 insights — fresh-context review, gate parser/executor split, coordinator under tmux, sun_path limit, REPL sentinel await, allowlist bypass reproduction, version-drift gate, mock ts at emit (+3 folds onto #180/#181, 1 stale drop) #183 edit tests-that-cannot-fail.md. This PR only appends one id to each page's related: list, so any conflict is a one-line related: union.

Routing decision

• Insight 1 → testing / data / new page wiki/testing/data/adjacent-tokens-in-extractor-fixtures.md.
  • The lesson is about constructing fixture input so each synthetic occurrence stays a distinct token. That is the data category ("tests need fixture data and you are choosing how to create it").
  • quality was the runner-up (a test that passes for the wrong reason). It is linked through related: instead, because the directive acts on the fixture, not the assertion strategy.
  • No new category needed.
• Insights 2/3 → infrastructure / agent-orchestration / existing page wiki/infrastructure/agent-orchestration/usage-limit-paused-workers.md.
  • That page already owns "which usage limit was hit decides the move", and a subagent is an orchestrated worker billed to the same seat.
  • platforms/processes was rejected: the fix is a model-routing decision inside the orchestrator, not an OS/CLI invocation difference.
  • No new category needed.