docs(claude-md): token diet — auto-loaded file 236 → 172 lines (Optimization PR B)#142
Merged
Conversation
PR B in the .md optimization sequence (Option D). CLAUDE.md is the file Claude Code auto-loads every session, so every line is "spent" in every turn — high-leverage trim target. CLAUDE.md (236 → 172 lines): - §Layout: skill inventory row (1 massive line) → 1 line + link to THIRD_PARTY_NOTICES.md (the canonical source for vendor / license breakdown). Future skill add/remove no longer needs to re-edit a multi-clause inline parenthetical. - §Commands: merged "After every workflow_dispatch" + "Connector- aware first-line check" into one 5-line block (was 14 lines split across two subsections that overlapped). - §Connectors: full §Multi-session audit pattern (28 lines) → 5-line reference + link to the moved-to-AGENTS.md version. Connector table descriptions tightened (Vercel / Supabase / Sentry / Gmail rows lost verbose justifications already implicit elsewhere). - §Conventions: "CLAUDE.md + AGENTS.md ship with every PR" bullet 12 → 6 lines (the rule is the rule — the four-paragraph justification was filler). - §Gotchas: Hypothesis property-based tests bullet 7 → 5 lines. - §Phase status: Epic #125 Item 3 + Karpathy LLM-Wiki detail (18 + 6 = 24 lines) → "Recently merged" + "Next deliverables" lists (12 lines). Detail lives in PHASE_STATUS.md and the skill's own SKILL.md. - §Companion files: added THIRD_PARTY_NOTICES.md; dropped claude-Creator.md self-reference (rarely useful in-session). AGENTS.md (344 → 386 lines): - New §"Multi-session audit pattern" hosts the full 4-step procedure + Section I forcing example (28 lines moved from CLAUDE.md). Non-Claude agents (Copilot / Cursor / Devin) read AGENTS.md so the pattern stays accessible cross-tool. - §"Phase + version state": added bullet documenting the PR B token diet + the multi-PR optimization sequence (PR A shipped #141 · PR B this one · PR C-G planned). CLAUDE.md + AGENTS.md edits ship per the lockstep rule. No code touched, no schema touched — pre-merge-prod-sim.yml won't trigger. Token impact estimate: CLAUDE.md is auto-loaded at session start + referenced from every system prompt assembly. -64 lines × ~12 tokens/line ≈ -750 tokens per session. Across many sessions per day this compounds. Next in sequence (TBD): PR C (AGENTS.md sync + dedup) · PR D (WORKFLOW.md archive) · PR E (SKILL.md restructure) · PR F (skill description audit ×38) · PR G (PHASE_STATUS.md restructure).
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
4 tasks
dackclup
added a commit
that referenced
this pull request
May 20, 2026
Third PR in the .md optimization sequence (Option D). PR B (#142) trimmed CLAUDE.md from 236 → 172 lines and moved the multi-session audit pattern out. PR C now closes the loop: AGENTS.md sections that duplicate CLAUDE.md become reference pointers; unique cross-tool content stays. AGENTS.md (386 → 342 lines): Dedup'd to reference CLAUDE.md as canonical: - §Tech stack — was 10 lines repeating CLAUDE.md §Stack with two extra deps. Now 6 lines: link to CLAUDE.md + note the two extras (pyarrow / yfinance) relevant for local build/test work. - §Commands — was 19 rows of full command table mirroring CLAUDE.md. Now 8 rows of cross-tool setup + dev-loop commands not in CLAUDE.md (install with extras, ruff --fix, single-module test, npm run dev, npm run lint). CLAUDE.md's verification ladder is the canonical command surface. - §Project structure — was 54-line tree with inline annotations duplicating CLAUDE.md §Gotchas (stale: "3 active vetoes" / "_EIGHT_K_DEFENSES_ENABLED = False until Phase 4"). Now 42 lines with file-purpose annotations only; bugs / drift live exclusively in CLAUDE.md §Gotchas. - §Phase + version state — was 46 lines duplicating CLAUDE.md §Phase status. Now 15 lines: reference to CLAUDE.md as canonical + cross-tool-specific bits only (production-verified-run baseline for local validation, open issue list, optimization PR sequence tracker). - §Companion files — refreshed to match CLAUDE.md's updated list (added THIRD_PARTY_NOTICES.md, dropped agent-Creator.md self-ref). Unique cross-tool content kept verbatim (no Claude-only-context equivalent exists): - §Testing (19 lines) — pytest @network marker + EDGAR_USER_AGENT requirement + where-to-put-tests guidance - §Code style (73 lines) — Python + TypeScript ✅ Good / ❌ Avoid examples; rationale why type hints + tabular-nums etc. - §Git workflow (29 lines) — branch naming + commit format + PR Draft↔Ready discipline + no-direct-main-push - §Boundaries (51 lines) — ✅ Always OK /⚠️ Ask first / 🚫 Never; GOLD content for non-Claude agents - §Security considerations (8 lines) - §Claude-Code-specific tooling (16 lines) — graceful-degradation note for Copilot / Cursor / Devin - §Multi-session audit pattern (30 lines) — moved here in PR B, full 4-step procedure + Section I forcing example CLAUDE.md (172 → 180 lines): - §Phase status "Recently merged" — added PR #142 (B token diet) - §Phase status — added "PR C in flight" note per lockstep convention Lockstep CLAUDE.md + AGENTS.md edit per the per-PR convention. Cumulative result for PR A + B + C on the agent-doc surface: - CLAUDE.md: 236 (pre-A) → 180 lines (today) = -24% - AGENTS.md: 344 (pre-A) → 342 lines = roughly flat, but the internal information density went up substantially (drift removed, duplication removed, multi-session pattern added) - Combined: 580 → 522 lines (-10%) with strictly more useful signal-to-token ratio Next in sequence: PR D (WORKFLOW.md archive Phase 0-3, 1732 → ~1450 lines) · PR E (SKILL.md restructure) · PR F (skill description audit ×38) · PR G (PHASE_STATUS.md "Current State" summary). Co-authored-by: Claude <noreply@anthropic.com>
9 tasks
dackclup
added a commit
that referenced
this pull request
May 20, 2026
… + top-10 movers) (#148) Epic #125 Item 3 PR 2 — closes the substantive remainder of Epic #125 after PR 1 (#140) shipped the workflow skeleton. Adds the per-ticker composite-score diff vs main, top-10 movers table, universe-size delta, and failure-path comment (PR 1 fell through silently on red checks). Baseline source: main's COMMITTED `frontend/public/data/` via `git show origin/main:...`, not a fresh re-run on main. The diff answers "did this PR change what production shows users?" — the right anchor is the last-committed main output, not a counterfactual. Free + no doubled EDGAR rate-limit pressure. Stale-baseline (>7 days) warning shows inline when the weekly cron hasn't run recently. Files: - `tools/pre_merge_diff.py` (~150 LOC, pure stdlib) - `tests/test_pre_merge_diff.py` (18 offline tests, no pandas/numpy) - `.github/workflows/pre-merge-prod-sim.yml` — adds 2 steps (fetch + diff), extends sticky comment, wraps post step with `if: !cancelled()` for failure-path comment - `CLAUDE.md` + `AGENTS.md` — lockstep update per Rule from PR #142 Co-authored-by: Claude <noreply@anthropic.com>
4 tasks
dackclup
added a commit
that referenced
this pull request
May 20, 2026
…loses #117) (#149) helper.py Section B was hard-failing on `non_reliance_filing` and `auditor_change` fires with "expected 0; flag broken?" — but PR #79 (Phase 4g, 2026-05-15) re-enabled both 8-K Tier-2 defenses by flipping `compute/scoring/tier2._EIGHT_K_DEFENSES_ENABLED = True`. Non-zero fires in the normal cohort band are EXPECTED post-4g, not bugs. Changes: - `section_b_tier2()` now takes `metadata` as a second parameter and replaces the hard-fail-on-any with a soft-band check against the academic cohort priors that calibrated each flag: * going_concern_disclosure — Mayew 2015: 1-3%; WARN > 5% * non_reliance_filing — Schroeder 2024: rare 4.02s; WARN > 2% * auditor_change — Cohen-Malloy-Nguyen 2020: 1-5%; WARN > 5% - Regression guard inverts: if `tier2_coverage_pct` ≤ 5% (proxy for `_EIGHT_K_DEFENSES_ENABLED = False` at compute time) and a flag still fires, that's the real bug — keeps the original "feature flag must hold" contract intact without flipping it backwards on healthy runs. - SKILL.md Section B description + Hard contract checks updated. - CLAUDE.md + AGENTS.md lockstep update per Rule from PR #142. Verification on current production data (commit `3da995dc`, 502 stocks): Section A-H run: 0 failures, 0 warnings (was: 2 failures pre-fix on the stale Section B expectations). Co-authored-by: Claude <noreply@anthropic.com>
dackclup
added a commit
that referenced
this pull request
May 20, 2026
…label clarification (#151) Closes Phase 0 of foundation reconciliation roadmap (epic #150). Adds honesty surface for analytical claims the rest of the docs glossed over. docs/METHODOLOGY.md — new §"Known limitations" section covering: - Survivorship bias (Wikipedia current SP500, not point-in-time) - Score semantics (percentile rank, not absolute quality) - Pillar correlation (Quality + Profitability ROE double-count) - extreme_*_estimate as method-applicability, not manipulation - Pillar weight rationale (empirical, not academic-derived) - Top-decile vetoes fire on top 10% by construction - Known calibration drift cross-refs (#11, #16, Phase 4.5d, #130) frontend/components/PillarRadarChart.tsx — sub-header now reads "0-100 percentile rank against current S&P 500 (sector-relative for Quality/Value/Growth/Profitability)" instead of generic "against the universe". CLAUDE.md + AGENTS.md — lockstep update per Rule from PR #142. Co-authored-by: Claude <noreply@anthropic.com>
This was referenced May 20, 2026
dackclup
added a commit
that referenced
this pull request
May 20, 2026
…17 → 27 (#154) The CLAUDE.md `## Phase status` headline claimed "defense layer 17" (7 vetoes + 10 annotates). The 2026-05-20 quarterly audit on issue #130 discovered the actual emit surface is 27 boolean flags: - 7 active vetoes (rank-suppressors) — unchanged - 10 annotate flags (the documented set) — unchanged - 5 method-applicability flags (extreme_<dcf,ev_ebitda,pb,pe,rim>_estimate — currently mis-aggregated into `manipulation_index`, scheduled for semantic split in epic #150 Phase 2) - 5 additional informational flags (cross_source_disagreement, late_filing_notification, manipulation_triple_flag, rem_suspect, restatement_history) Phase 1.2 closure: this PR updates the headline summary in CLAUDE.md to reflect reality (17 declared veto+annotate flags PLUS 10 additional flags), links to the audit comment as canonical evidence, and notes that Phase 2 of epic #150 will reorganize the taxonomy properly. No compute logic change. No schema change. The 10 additional flags already emit in production — this PR just acknowledges them in the agent-facing summary docs. Also reframes the "in flight" section to reflect Phase 0 + 1.3 closed, 1.2 in progress, 1.4-1.6 + 2-3 remaining. CLAUDE.md + AGENTS.md lockstep update per Rule from PR #142. Co-authored-by: Claude <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
PR B in the
.mdoptimization sequence (Option D — 7-PR overhaul). PR A shipped via #141; this is the CLAUDE.md token diet.CLAUDE.md is the file Claude Code auto-loads every session, so every line is "spent" in every turn. The bloated parts were:
.claude/skills/row (one massive line listing every vendor / license bucket)Numbers
Net token impact: ~-64 × ~12 tokens/line ≈ -750 tokens per session auto-loaded. Compounds across all sessions per day.
What moved where
.claude/skills/karpathy-llm-wiki/SKILL.md(already there)Section-by-section trim
Risk notes
Rules 1-18(SKILL.md actually has Rule 18 — confirmed) ·§"Multi-session audit pattern"(anchor exists in AGENTS.md L345).What this PR does NOT touch
compute/,frontend/,tests/)pre-merge-prod-sim.ymlwon't trigger (path filter excludes.md)Next in sequence (NOT this PR)
PR C (AGENTS.md sync + dedup) · PR D (WORKFLOW.md archive Phase 0-3) · PR E (SKILL.md restructure) · PR F (skill description audit ×38) · PR G (PHASE_STATUS.md "Current State" summary).
Test plan
pre-merge-prod-sim.ymldoes not triggerhttps://claude.ai/code/session_015649aRyi2bvciQYZVNACd2
Generated by Claude Code