main
md 52 lines 7.02 KB
Rendered Raw
1 # Farnsworth — History
2
3 ## Core Context
4 - Owns editorial analysis of GitHub and adjacent press signals for the published weekly brief.
5 - Analysis output must stay useful to readers and structured enough for downstream automation.
6
7 ## Learnings
8 - The analyzer contract should remain a superset of the published weekly frontmatter plus the quality gate so one artifact serves editorial and pipeline needs.
9 - Weekly briefs work best as named macro trends supported by repo evidence, with repo links rendered from the crawl artifact's `url` field.
10 - Press and industry coverage are correlation signals, not material to repackage; the value is in explaining the gap between narrative and developer traction.
11 - Reader-facing renders need a cleanup pass that strips AI-only scaffolding before publication.
12 - The learning loop only matters when lessons are persisted and injected back into the next prompt through shared wisdom and skills.
13 - The squad reskill audit showed repeated charter and history scaffolding across agents; that boilerplate now lives in `minimal-agent-charter`, `agent-history-hygiene`, and `weekly-learning-loop`.
14 - The reskill pass also cut squad agent-doc footprint from 39568 to 12521 bytes, with every charter at or below the 1.5 KB target and the largest histories back under maintenance limits.
15 - 2026-05-21T12:33:16.507+02:00: Weekly analysis output must use a strong journalistic headline, explicitly state when no press data is available, and keep `Key References` complete so downstream publishing does not inherit placeholder artifacts.
16 - 2026-05-21T10:38:30Z: Scribe processed spawn manifest; decision on headline review appended to decisions.md and archived from inbox.
17 - 2026-05-25T11:56:08Z: W22 saw the most concentrated coordinated star-farming campaign in the crawl's history — a tight cluster of game-cheat, software-unlock, and AI-branded repos landing at 421–429 stars with zero forks in minutes. This is now a named pattern that should inform noise-filtering heuristics and lower editorial trust for new repos without fork activity.
18 - 2026-05-25T11:56:08Z: Press data was available this week (2026-W22-press-context.md) and surfaced a useful convergence: TechCrunch's article on inflated AI ARR metrics ran the same week GitHub new_repos was experiencing coordinated star inflation. Cross-referencing press themes against developer noise patterns can surface non-obvious editorial angles.
19 - 2026-05-25T11:56:08Z: The agent-skills ecosystem transition from hint (W21, chrisbanes/skills, vercel-labs/zero) to confirmed category (W22, multiple independent skill packaging repos alongside 140k+ ⭐ trending anchors) reinforces the wisdom pattern: clustered movement across independent teams is a stronger signal than any single repo launch.
20 - 2026-05-25T11:56:08Z: stars_tracked and repos_featured remain estimated due to lack of automated summation; these fields should be computed by the pipeline and injected into the prompt rather than hand-calculated each week — a recurring quality gap worth escalating as a skill or pipeline improvement candidate.
21 - 2026-06-01T12:16:50Z: W23's dominant noise pattern shifted from star-clustering (W22) to fork-inflation — polymarket bot repos landing with fork counts 10-20x their star count and keyword-stuffed descriptions. Same coordinated intent, new metric vector. Noise classification heuristics should track fork/star ratio as a manipulation signal alongside creation-timestamp clustering.
22 - 2026-06-01T12:16:50Z: Agent skills crossed a cultural boundary in W23 with the first major Chinese-platform design skills (Xiaohongshu, WeChat) achieving 1k-2k+ stars. Skills economy is now multilingual; editorial framing should treat geographic and linguistic expansion of skills as a trend signal distinct from English-developer workflow skills.
23 - 2026-06-01T12:16:50Z: GitHub Copilot token-billing backlash (TechCrunch, 2026-05-30) produced a clear same-week developer response in self-hosted AI workspace and alternative-routing repos — strongest press-developer convergence of W23. Monitoring for platform billing/policy changes as a leading indicator of self-hosting repo spikes is a reusable editorial pattern.
24 - 2026-06-01: Added hindsight validation that backfills Signal/Noise/Gaps repo calls from legacy summaries, writes a markdown scorecard for reskill, and treats frontmatter `predictions` as the forward-looking source of truth for cleaner future scoring.
25 - 2026-06-05T16:26:00+00:00: Multi-source press should feed weekly analysis as a source-aware compact correlation artifact, not as raw article dumps; preserve citations while tiering weak category/fuzzy matches separately from strong correlations.
26
27 ## 2026-06-05 LLM input strategy assessment
28
29 - Evaluated 4 options: raw dumps, merged/summarized, staged source-specific, compact correlation.
30 - Finding: New multi-source crawl produces ~74k tokens GitHub raw + ~11.4k tokens external-news; old was ~7k TechCrunch press context.
31 - Recommendation: Use compact press-context artifact (Option 4) as the only external-news input to weekly analysis.
32 - Press context design: ranked press items (5-10 with URL, source, date, relevance, why-it-matters), ranked repo/news correlations (5-10 strong), weak-correlation bucket (category-only), divergences (3-6), complete citations.
33 - Token budget: <= 8k token-estimate.
34 - Quality gate additions: check 3-5 article links, no raw dumps, distinguish strong vs. weak correlations, caveat on source errors.
35 - Decision recorded in .squad/decisions.md.
36
37 ## Round 2026-06-01T15:40
38
39 ### Issue #38: Hindsight Validation
40 - PR #227 opened
41 - Implemented `predictions` frontmatter schema: `{repo, direction, confidence}`
42 - Validator generates scorecards: `.squad/reskill/scorecards/YYYY-WNN.md` (human) and `data/metrics/scorecards/YYYY-WNN-scorecard.json` (machine)
43 - 548 tests passing
44 - Ready for merge
45 - 2026-06-05T17:42:56Z: Analysis map/reduce should be justified by context pressure and claim quality, not crawl speed; mappers should emit cited claim ledgers while a single reducer preserves global thesis, dedupe, citations, and the analysis gate contract.
46 - 2026-06-05T20:57:09Z: Best analysis decomposition candidate is signal-type claim-ledger map/reduce with deterministic preflight compaction, optional hierarchical repo/source sub-mappers only above budget triggers, one reducer/final writer for global thesis and voice, and critic passes as QA rather than competing article writers.
47
48 ## 2026-06-06 — Issue #258 map/reduce dry-run scaffolding
49 - Added deterministic/local `scripts/map_reduce_dry_run.py` candidate-only analysis path with `analysis_map_v1` mapper ledgers for new repos, trending repos, press correlations, and prior continuity.
50 - Reducer emits `analysis_editorial_plan_v1`, rejected-claims and contradictions sidecars, a single reader-facing candidate, and a QA comparison report that blocks publish eligibility by design.
51 - Workflow wiring exposes `analysis_path=map-reduce-dry-run` only for `dry-run`/`candidate-only`; current crawl/analyze/publish behavior remains default and promotion/deploy paths remain skipped.
52 - Validation: `pytest -q` passed (667 tests, 2 subtests).