🛠 Project View

# BUILD: Equity Comp — Liquidity Event Calculator Suite (Phase 2.5) Created: 2026-05-11 Status: SEALED — pending risk-check ## Goal Add an 8th tab "Calculators" to the existing Equity Comp Dashboard containing 5 collapsible, interactive client-side calculators that support liquidity-event decision-making for the May 14 / July 28 early-release window — fully additive to the current comp dashboard, no edits to existing tabs. ## UX requirements (cross-cutting) - **Easy to execute**: every calculator should be usable in under 30 seconds with no documentation. Defaults pre-populated from `comp_state.json`. One-click "Reset to defaults" button per calculator. - **Helper text everywhere**: every input has a short "ⓘ" tooltip or inline `<small>` explainer. Every output has a one-line plain-English interpretation (e.g. "This is what hits your bank after taxes."). Every calculator has a top "What this answers" 1–2 sentence intro and a bottom "Caveats" note. - **No jargon without decode**: AMT, bargain element, strike, vested, sell-to-cover all get inline glossary the first time they appear in each calc. - **Visual feedback**: outputs animate/highlight when inputs change; red text when a constraint is violated (e.g. allocation sliders sum != 100%). - **Mobile-aware spacing** even though mobile card is deferred — sections should not break the desktop layout if viewed on a tablet. ## Scope ### In - New tab `tab-calculators` in `dashboards/comp_dashboard.html` (desktop only — mobile deferred) - 5 collapsible `<details>` sections inside that tab, each rendering one calculator: 1. Early Release / Liquidity Calculator 2. Proceeds Distribution / Allocation Engine 3. ISO Exercise Timeline Model 4. AMT Sensitivity Model 5. Long-Term Equity Tracker (read-only summary aggregating existing comp_state.json) - All calc logic in pure client-side JavaScript inline in the rendered HTML — no new Python compute - New `render_calculators_tab(s)` function in `scripts/comp_render.py` - Tab nav link added in dashboard header - Defaults pre-populated from `comp_state.json` (current shares, strike, price) but every input editable ### Out (not in this build) - Mobile card (`mobile/comp.html`) — deferred to Phase 2.6 - Pipeline wiring for A39 (`/comp-refresh` skill remains manual trigger) - Per-grant AMT (already in HIGH backlog as Phase 2 sub-item — leave there) - Persistence of calculator inputs beyond page reload (use `localStorage` only — no JSON sidecar yet) - Preset save/load for allocation sliders (deferred — sliders start at 0 each session) - Any change to `comp_compute.py` schema or `comp_state.json` writers - Any touch on shared CSVs (recommendation_events, control_group, sil_learnings, etc.) — isolation contract preserved ## Files affected ### Created - (none — all changes inline in existing files) ### Edited - `scripts/comp_render.py` — add `render_calculators_tab(s)` function (~250 lines new); add nav link entry; register tab in main render assembly - `dashboards/comp_dashboard.html` — regenerated as output of `comp_render. …[truncated at 3072B — open file for full content]

# BUILD: Session-start health check cap fix + full-report path surfacing Created: 2026-05-16 Status: SEALED — pending risk-check ## Goal Stop silently truncating the MDE health check at session start so the user/Claude make decisions on the **complete** warning list, not a clipped prefix. Surface the path to the full HTML report alongside the (now larger) inline preview so the user can always reach the unabridged version in one click. ## Scope **In:** - Bump the `[:4000]` slice on the MDE health check hook from 4000 → 12000 chars (`C:/Users/jjwat/.claude/settings.json` line 25). - After the truncated preview, print one line: `Full report: <path to health_check_report.html>`. - Verify the resulting hook payload still fits within Claude Code's per-hook context budget (no crash, no truncation on Claude's side). **Out:** - Smart-truncation logic (prioritize ERRORS first) — over-engineered for current scope; revisit if 12000 still clips. - Reformatting the health check report itself — separate concern. - Touching the digest hook (line 31) — digest is verified accurate at 22/22 sections; no cap needed. - Touching session_context.py (line 19) or the workspace CURRENT.md hook (line 15). ## Files affected - **EDIT** `C:/Users/jjwat/.claude/settings.json` — single line (25) — two changes inside the embedded one-line Python: 1. `s[:4000]` → `s[:12000]` 2. After `'MDE HEALTH CHECK:\\n'+s[:12000]`, append `+'\\n\\nFull report: C:/Users/jjwat/OneDrive/Claude Code_All/Claude Code Terminal/projects/investments/dashboards/health_check_report.html'` (verify path exists before sealing the string). No other files touched. ## Steps 1. **Verify full-report HTML path exists** — `Glob` for `dashboards/health_check_report.html` under investments. Acceptance: file exists and is mtime-current (regenerated by pipeline Step 31 or similar). If missing, abort and file follow-up to generate it before re-sealing. 2. **Read current settings.json line 25** to confirm the exact embedded Python string still matches the `[:4000]` pattern (no parallel-session edit since audit). Acceptance: line matches expected shape. 3. **Edit settings.json line 25** — bump cap + append path line. Acceptance: file written; manual diff shows only those two textual changes. 4. **Validate JSON** — `py -c "import json; json.load(open('C:/Users/jjwat/.claude/settings.json'))"`. Acceptance: exits 0, no parse error. (Critical — a broken settings.json breaks every future session.) 5. **Dry-fire the hook** — run the modified command line directly: `py -c "<command>"` from a shell. Acceptance: prints JSON with the SessionStart payload, no Python errors, payload includes the path line at the end, total length ≤ ~15KB. Confirms the change works before next real session uses it. 6. **Document** — append to `CHANGELOG.md` under today's entry; close the `[FIX]` task with `[x]` + Closes line. ## Rollback - **Failure modes:** (a) JSON parse error → next session fails to start hook → no digest/health at session-start (but Claude still works). (b) Hook timeout — unlikely, …[truncated at 3072B — open file for full content]

# BUILD — Pipeline run efficiency overhaul **Date filed:** 2026-04-27 (Session 109) **Tier:** HIGH **Class:** Performance / architecture (not data integrity — that's the parallel CRITICAL stdout-deadlock fix) **Linked task:** TASKS.md HIGH — `[INVESTIGATE] Pipeline run efficiency overhaul` **Linked to:** `references/BUILD_yfinance_timeout_guard_20260427.md` (Phase 0 stdout streaming fix is a hard dependency — without visibility, we can't measure efficiency) --- ## Symptom Pipeline runs are taking **90+ minutes** with new CSVs. Each session ships data but the wait is painful and discourages running. Multiple architectural decisions compound: - Strictly sequential step orchestration - Network steps interleaved with cheap compute (no batching) - No incremental / no-change short-circuit (full re-run every time) - No per-step timing logged (we're guessing where time goes) - Slow/optional/non-blocking steps gate everything downstream - Some steps (mobile deploy) are fire-and-forget candidates running synchronously at the end - Staleness guards exist on a few steps (Step 10, 10a) but not consistently applied --- ## Hard dependency ⚠ **`BUILD_yfinance_timeout_guard_20260427.md` Phase 0 must ship first.** Without real-time stdout streaming we cannot measure per-step duration accurately, and the 64KB pipe-deadlock will reappear inside any parallel-batch implementation. Phase 0 is also the cheapest single change with the biggest immediate quality-of-life win. --- ## Inefficiency catalogue | # | Inefficiency | Where | Fix direction | Phase | |---|---|---|---|---| | 1 | No per-step timing logged | run_pipeline.py top-level | Wrap each step in a wall-clock timer; append to `data/logs/PIPELINE_TIMINGS.csv`; surface top-3 slowest in session digest | A | | 2 | Step 6 dashboards run sequentially | watchlist → take_action → gate_alts → dir_stability → metric_insights → l1_dashboard → index | All independent — parallelize via `concurrent.futures.ProcessPoolExecutor` | B | | 3 | Network steps interleaved with cheap compute | 8.56 (net) → 8.57 (cheap) → 8.6 (cheap) → 9 (cheap) → 10 (net) → 10a (net) | Batch all network calls together near end; run concurrently | C | | 4 | yfinance one-batch-per-date in 8.56 | fetch_sold_prices.py:167 loop | Single bulk `yf.download()` over the date range; partition results client-side | C | | 5 | No incremental / no-change short-circuit | Phase 1 ingest, MPS, BF | If CSV mtime hasn't changed since last run, skip Phase 1 ingest. If MPS inputs unchanged, skip Phase 2. | E | | 6 | Mobile deploy (Step 12) runs synchronously at end | run_pipeline.py:660 | Move to fire-and-forget background subprocess (no `wait()`); pipeline completes immediately, deploy runs to completion in the background | C | | 7 | Staleness guards inconsistent | Step 10/10a have them; 8.56, 11, 11.5, 11.6, 12 do not | Apply staleness guard pattern to every "expensive but doesn't change often" step — check input mtime / output mtime, skip if output fresher than input AND output age < threshold | D | | 8 | Slow/optional steps gate …[truncated at 3072B — open file for full content]

# BUILD — Health CHECK 45 + CHECK 46 + Pre-Fix Q7 Rewrite (class-of-failure prevention layer) **Session:** 2026-04-23 Session 133 **Closes:** `[BUILD] Health CHECK 45` (TASKS.md line 362) + `[REVIEW + BUILD] Recurring class-of-failure` umbrella (TASKS.md line 276) **Related:** `references/BUG_REPORT_2026-04-23.md` (Prevention plan section) **Unblocks:** Gate Intelligence Layer P1 (TASKS.md line 360 — sequenced behind CHECK 45) --- ## Context — Why this is being built Four incidents in 25 days share one class of failure: **a shared CSV was modified (schema, row-type, or dtype) and the downstream consumer audit was incomplete.** Three caused data loss; the fourth (2026-04-23) would have if it had landed in a writer instead of a render script. | Date | Incident | Covered by existing check? | |---|---|---| | 2026-03-28 | FIELDNAMES schema drift | ✓ CHECK 18 / 26 / 35 | | 2026-04-18 | Writer-side rec_type scope (same-date overwrite) | ✓ CHECK 37 | | 2026-04-19 | Reader-side rec_type filter missing | ✗ **GAP → CHECK 46** | | 2026-04-23 | Dtype consumer cast (int on float) | ✗ **GAP → CHECK 45** | Originally proposed as 4 new checks (CHECK 43/44/45/46). Phase 1 exploration (this session) found CHECK 43 equivalent covered by CHECK 18/26/35 and CHECK 44 equivalent covered by CHECK 37. **Net new work: 2 checks, not 4.** Pre-Build Q7 as written relies on verbal confirmation ("confirm all writers have identical FIELDNAMES"). The 2026-04-23 postmortem demonstrates this class of process control has failed 4× in 25 days. Per user memory `feedback_user_never_verifies.md`: governance that asks the user to verify is worthless. Q7 is rewritten to require **Claude-self-executed grep output** as part of the Pre-Fix Assessment, not user-paste-proof. --- ## Scope (locked) 1. **CHECK 45** — shared-CSV consumer format/dtype audit. Extended scope: numeric dtype casts (`int()` on a field written as float) **and** string-format consumers (slice, split, regex on structured string fields). Per TASKS.md line 362 Session 132 close decision. 2. **CHECK 46** — reader rec_type filter audit. Flags any script that reads a rec_type-bearing shared CSV without an explicit rec_type filter. 3. **Pre-Fix Q7 rewrite** — drop verbal-confirmation wording; replace with Claude-self-executed grep protocol. 4. **Task closure** — mark CHECK 45 (line 362) and umbrella (line 276) `[x]` done. 5. **Governance housekeeping** — CHANGELOG entry with `**Closes:**`, PROJECT_MAP entry for this plan file, BUG_REPORT prevention-plan update. --- ## Risk assessment **Safety rating:** LOW. All additive or text-only. No shared CSV touched. No pipeline step added/removed. Fully reversible. **Downstream impact verified:** - `session_digest.py` — 0 references to Q7 or any Q-number text - 7 files reference "Q7" by number only (pass-by-number, not text-parsed) - Q7 number preserved; only wording changes - Gate Intelligence Layer P1 (line 360) is sequenced behind this build — will unblock on completion, no conflict - Pre-Build Q10 (write_safe, `BUILD_write_safety_hardening_2 …[truncated at 3072B — open file for full content]

# BUILD — Gate Identity System (GIS) — Naming + Propagation > ## ✅ Build Complete — Session 152 (2026-04-25) > > All 4 batches shipped end-to-end in a single session. Zero regressions on a Danger Zone CSV that produced **4 incidents in 25 days** prior. **Health: 0 errors / 47→27 warnings / 134→141 OK.** > > ### Post-build retrospective — what kept this safe > > **What the user did:** > - Capped scope upfront: "4 batches ideal, 6 tops if it makes it safer" — set discipline before any code touched. > - Demanded LESSONS_LEARNED.md review before /risk-check, not after a problem surfaced. > - Reviewed `--dry-run` output explicitly before `--execute` on Batch 4 (the only non-revertable action) — surfaced their own insight (v10 7x stricter than v9) during review. > - Raised clarifying questions mid-build (gate_hash semantics, dist-sensitivity tradeoffs, SHADOW contamination concern) — caught design issues early. > - Refused to bundle stages — accepted that Batch 4 was the high-risk action and gated it explicitly. > - Approved at each go/no-go gate explicitly. Never inferred-and-executed. > > **What the system did:** > - **/risk-check fired before each batch** — Q0–Q9 + Pre-Fix Steps 1–6 with concrete file:line refs, not boilerplate. > - **Schema add decoupled from writer population** — Batch 2 proved the FIELDNAMES contract held alone before Batch 3 introduced population logic. The two prior schema-mismatch incidents (2026-03-28, 2026-04-18) were both caused by bundling these. > - **Recon revised scope** — Batch 3 originally 4 writers. Code inspection classified 2 as read-modify-write (DictReader→DictWriter preserves new fields automatically) and proved it via simulation on all 2,128 actual rows BEFORE first live edit. Cut scope in half with mechanical evidence. > - **3-layer Batch 4 safety** — pre-write file snapshot + sha256 record + `--dry-run` default + atomic `write_csv_safe` + post-write asserts (row count unchanged, sha256 changed, expected gate_id values present, pre-tracking still blank). > - **Per-batch `.bak` rollback hooks** — project not under git, so file-copy snapshots replaced the planned `git revert` strategy. Hooks created BEFORE first edit each batch. > - **WIRING + PROJECT_MAP updated inline** — every new file (`gate_identity.py`, `migrate_rec_events_gate_id.py`) registered in the same batch that introduced it. CHECK 31 stayed clean; no orphans surfaced next session. > - **CHECKs 18/26/35/45 verified between every batch** — health check became the per-batch verification gate. Warnings dropped 47→27 across the build because WIRING fixes cleared a stack of CHECK 31 false positives en route. > - **Module-load collision assert** in `gate_alternatives.py` caught the alt9-vs-current rule-set match the moment the simplified hash shipped — converted from runtime mystery into compile-time exception. > - **Of 7 historical rec_events failure classes, 4 were structurally impossible for this build (no rec_type change, no overwrite scope change).** Recognized upfront, eliminated from the risk surface. > > **Replicabl …[truncated at 3072B — open file for full content]

# BUILD: Equity Comp — Liquidity Event Calculator Suite (Phase 2.5) Created: 2026-05-11 Status: SEALED — pending risk-check ## Goal Add an 8th tab "Calculators" to the existing Equity Comp Dashboard containing 5 collapsible, interactive client-side calculators that support liquidity-event decision-making for the May 14 / July 28 early-release window — fully additive to the current comp dashboard, no edits to existing tabs. ## UX requirements (cross-cutting) - **Easy to execute**: every calculator should be usable in under 30 seconds with no documentation. Defaults pre-populated from `comp_state.json`. One-click "Reset to defaults" button per calculator. - **Helper text everywhere**: every input has a short "ⓘ" tooltip or inline `<small>` explainer. Every output has a one-line plain-English interpretation (e.g. "This is what hits your bank after taxes."). Every calculator has a top "What this answers" 1–2 sentence intro and a bottom "Caveats" note. - **No jargon without decode**: AMT, bargain element, strike, vested, sell-to-cover all get inline glossary the first time they appear in each calc. - **Visual feedback**: outputs animate/highlight when inputs change; red text when a constraint is violated (e.g. allocation sliders sum != 100%). - **Mobile-aware spacing** even though mobile card is deferred — sections should not break the desktop layout if viewed on a tablet. ## Scope ### In - New tab `tab-calculators` in `dashboards/comp_dashboard.html` (desktop only — mobile deferred) - 5 collapsible `<details>` sections inside that tab, each rendering one calculator: 1. Early Release / Liquidity Calculator 2. Proceeds Distribution / Allocation Engine 3. ISO Exercise Timeline Model 4. AMT Sensitivity Model 5. Long-Term Equity Tracker (read-only summary aggregating existing comp_state.json) - All calc logic in pure client-side JavaScript inline in the rendered HTML — no new Python compute - New `render_calculators_tab(s)` function in `scripts/comp_render.py` - Tab nav link added in dashboard header - Defaults pre-populated from `comp_state.json` (current shares, strike, price) but every input editable ### Out (not in this build) - Mobile card (`mobile/comp.html`) — deferred to Phase 2.6 - Pipeline wiring for A39 (`/comp-refresh` skill remains manual trigger) - Per-grant AMT (already in HIGH backlog as Phase 2 sub-item — leave there) - Persistence of calculator inputs beyond page reload (use `localStorage` only — no JSON sidecar yet) - Preset save/load for allocation sliders (deferred — sliders start at 0 each session) - Any change to `comp_compute.py` schema or `comp_state.json` writers - Any touch on shared CSVs (recommendation_events, control_group, sil_learnings, etc.) — isolation contract preserved ## Files affected ### Created - (none — all changes inline in existing files) ### Edited - `scripts/comp_render.py` — add `render_calculators_tab(s)` function (~250 lines new); add nav link entry; register tab in main render assembly - `dashboards/comp_dashboard.html` — regenerated as output of `comp_render. …[truncated at 3072B — open file for full content]

# BUILD: Gate Builder L1 dropdown completion (Session 172 second-instance fix) Created: 2026-04-29 Status: SEALED — pending risk-check ## Goal Apply Session 172's already-established `_L1_MPS_BY_NAME` name-fallback pattern to the gate-builder metrics dropdown render path (`window.GB.metrics` array, line ~2756 + ~2813) so SM L1 cols that have NR L1 equivalents (RSI, ROE) display real grade/mps/sep_pp instead of `?` / 0. ## Context — why this is "second-instance" Session 172 (2026-04-28) shipped name-fallback `_L1_MPS_BY_NAME` lookup at `gate_alternatives.py:858-862` for the alt **rule-row** rendering path. Smoke-verified live: "JSON 8/8 L1 rule rows mps>0; new silfixv2 Phase B variant picked up RSI as L1 anchor." That fix is real and unchanged. The gate-**builder dropdown** path (`window.GB.metrics` array at line ~2756 SM L1 branch + ~2813 NR L1 branch) is in the same file but a **separate render function** that was never wired to use `_L1_MPS_BY_NAME`. Empirical proof: live `dashboards/Gate_Alternatives_Dashboard.html` (last regenerated 2026-04-28 20:41, after Session 172 ship) shows: ``` "col":14,"name":"RSI","grade":"?","mps":0.0,"sep_pp":0.0 ``` Same bug class, same file, different function. This build closes the second instance. ## Scope **In:** - `gate_alternatives.py:~2748-2772` — SM L1 branch in metrics dropdown JSON builder - `gate_alternatives.py:~2806-2829` — NR L1 branch (same fix pattern) - Suffix-normalization helper (strip trailing `" (%)"` and whitespace before lowercase compare) — addresses Agent F showstopper for "Net Profit Margin" vs "Net Profit Margin (%)" - Direction-divergence WARN log if matched NR-L1 `winner_dir` ≠ SM-L1 `winner_dir` (rare, signals data vintage drift) - Reuse existing `_L1_MPS_BY_NAME` dict (loaded by Session 172 — not new infrastructure) - Extend CHECK 62 in `mde_health_check.py` to also audit `window.GB.metrics` array (currently audits scoring path only) - File a follow-up TASKS.md entry for the 15 SM L1 metrics that have NO NR L1 equivalent in `l1_metric_power_scores.csv` **Out (explicit):** - Re-scoring the 15 SM L1 metrics (D/E, ROIC, EV/EBITDA, FCF Yield, FCF Margin, NPM, P/C, P/C Multiple, P/FCF, P/S, Rev Growth, EPS, ROA, Int Coverage) — separate larger build, touches `mde_02_metric_power.py` + ingest - Schema changes to `l1_metric_power_scores.csv` - Changes to Session 172's already-shipped paths (lines 858-862, 1034, 1571-1572) — verified working, untouched - run_pipeline.py edits (none needed — Step 7.5 already runs gate_alternatives.py) ## Files affected - `scripts/gate_alternatives.py` (single file edited — A22 Danger Zone, single-writer, ~25 lines additive) - `scripts/mde_health_check.py` (CHECK 62 extended, ~5 lines additive) - `TASKS.md` (1 new follow-up task for the 15 unfixable SM L1s) - `CHANGELOG.md` (entry on ship) - `references/BUILD_gate_builder_l1_dropdown_completion_20260429.md` (this plan) ## Steps ### Phase 1 — Diagnose + Fix (REVERSIBLE) 1. **[Read-only]** Confirm `_L1_MPS_BY_NAME` is loaded and populated at module level (Session 172 s …[truncated at 3072B — open file for full content]

# BUILD: SIL Output Quality + Audit Trail Closure (two streams) Created: 2026-05-22 Status: SEALED — pending risk-check > **Supersedes:** `BUILD_sil_intent_vs_actual_closure_20260516_v2.md` (v2 + v2.1 patches). That plan is retired in favor of this scope after user direction 2026-05-22 dropped Phases 2, 3, 4-tuning, and 5-autonomous in favor of focused output-quality + audit-trail work. Predecessor plans `BUILD_sil_p2_enforcement_20260508.md` and `BUILD_sil_trusted_brain_20260423.md` remain SUPERSEDED (UNHOOK_LOG entries to be added in Stream 1 Phase A). ## Goal Make SIL's daily output decision-useful (Stream 2 — output quality) and make every new SIL deposit audit-trailable (Stream 1 — citation infrastructure). Ship in that order — output quality first so audit trail isn't "better filing on bad content." ## Scope **In scope:** **Stream 2 — Output quality pass (~1-2h, LOW risk):** - Fix 4-star-confidence boilerplate bug in SIL output (every ADD candidate insight text says "has 4-star confidence" regardless of actual stars — verified in `sil_grand_summary.json` 2026-05-22). - Rewrite ADD-candidate insight template to include real evidence (ctrl_d value + n, BF rank, MPS grade) instead of star-count restatement. - Investigate direction-flip threshold-clustering (14 warnings all firing at exactly 10 consecutive opposing runs — mass-trip suggests threshold needs raising or signal-strength differentiation). - Audit CSV-escaping in `sil_learnings.csv` writer — multi-line insight text with unescaped commas breaks readers parsing as plain CSV. **Stream 1 — Audit-trail infrastructure (~2.5h, LOW risk):** - Phase A: finish audit-trail plumbing (snapshot, ground-truth count, retroactive magnitude-guard unit test, deposit-writer wiring for evidence_refs/threshold_for_action/expiry_sessions, one-shot backfill, UNHOOK_LOG entries for 2 superseded predecessor plans, CHANGELOG + PROJECT_MAP). - Phase B: flip P2 WARN→BLOCK in `sil_charter.py` after ≥2 clean pipeline runs, new CHECK 73 enforces P2 BLOCK + rejects uncited deposits. - Housekeeping: update `sil_charter.py:240-267` AUTHORITY_RULES `_status` from "PLACEHOLDER — wired in Phase 4" to "DEFERRED — scoring diversity preserved per user 2026-05-22". **Out of scope (DROPPED from prior plan):** - Authority hierarchy refactor / AUTHORITY_RULES wiring into advisors (Phase 2 of prior v2.1) — user wants MPS/Cohen/Control/BF scoring diversity preserved. - `decision_lineage.csv` writer (Phase 3 of prior v2.1) — duplicates CHANGELOG without Phase 2 authority context. - Synthesis pattern tuning (Phase 4 of prior v2.1) — 76 `sil_synthesis` rows already deposited, patterns are firing. - Autonomous output badges (Phase 5 of prior v2.1) — digest already surfaces ADD candidates + warnings at `session_digest.py:920`. - Mobile SIL card decision — separate ~2-line edit, not gated on this plan. ## Files affected **Edited:** *Stream 2:* - `scripts/signal_intelligence.py` — fix 4-star boilerplate in `generate_ideas()` / synthesis-pattern deposit text (find the f-string template hard-coding …[truncated at 3072B — open file for full content]

(no additional detail)

(no additional detail)

# BUILD: Retire Sheets Finance — Phase A+ (PARKED — design pivot to standalone momentum_pull) Created: 2026-04-30 Status: 🅿 **PARKED** — design pivot mid-flight; resume requires user answers to Q1–Q3 below Predecessors: - `BUILD_beta_finviz_sheets_finance_retire_20260425.md` — v1 (Phase A only, Beta into finviz_pull) - `BUILD_beta_finviz_sheets_finance_retire_20260429.md` — v2 (Phase A+ scope, additive to finviz_pull) - `BUILD_beta_finviz_sheets_finance_retire_20260430.md` — v3 (additive plan, sealed + risk-checked GREEN/ULTRA-SAFE; not executed) --- ## Why parked User reviewed v3 (which extended `finviz_pull.py` additively, scored 0/ULTRA-SAFE) and pivoted the design 2026-04-30: > "needs to be its own run, just like finviz pull and ratio pull are their own, i want this to also be its own, can call them momentum pull maybe" Implication: instead of adding 3 columns to `finviz_pull.py`, build a **separate sibling script** (working name `momentum_pull.py`) — its own scrape run, its own checkpoint, its own output CSV, its own skill invocation. v3 plan invalidated as written. User also retracted the RSI-only / held-symbols-only quick-pull follow-up. Smoke-test gate retained. The pivot raises 3 open scope questions that **must be answered before resume**, plus several governance additions a standalone-script plan needs that the v3 additive plan didn't. --- ## Open questions (BLOCKING — must resolve to resume) ### Q1 — Scope mode for `momentum_pull.py` The new script would scrape the same Finviz snapshot pages that `finviz_pull.py` already loads (Beta, NPM, Sales Q/Q all live on the per-ticker snapshot). If both run on the full ~5,635-ticker universe, that's duplicate HTTP work for the same data. Three viable design modes: | Option | Behavior | Tradeoff | |---|---|---| | **A. Held-only default + `--symbols all` flag** | Default run = ~50–100 held tickers, ~2 min wall. Universe mode exists but rarely used. | **Recommended.** Effectively the RSI-only quick-pull idea repurposed for the momentum metric set. Real new capability (sub-2-min refresh anytime), minimal duplicate HTTP. | | **B. Universe pull, different cadence** | Full ~5,635 tickers but runs monthly (Beta is slow-moving) instead of weekly. | Reduces duplicate-scrape cost over time. Sales Q/Q updates quarterly anyway, so monthly cadence is fine. Loses the quick-refresh capability of A. | | **C. Naive sibling** | Full ~5,635 tickers, same weekly cadence as finviz_pull. | Cleanest mental model. But ~1h 53min of pure HTTP duplication per cycle + 2× Finviz request volume on overlap days. Pushed back on as wasteful. | **Default recommendation:** **Option A**. Decision deferred to user. ### Q2 — Script name Working name `momentum_pull.py`. Confirm or propose alternative (e.g., `quick_pull.py`, `held_pull.py`, `metrics_pull.py`). ### Q3 — Skill name Working name `/momentum-pull`, parallel to existing `/finviz-pull` and `/ratio-pull`. Confirm or propose alternative. --- ## Tentative scope (assuming Q1 = Option A) **In scope (when resumed):** - New file: `s …[truncated at 3072B — open file for full content]

# BUILD: Fix CHANGELOG `**Closes:**` enforcement — hook bypass + Check 3 BLOCK promotion Created: 2026-04-29 Status: SEALED — pending risk-check Supersedes (different scope): `references/BUILD_changelog_tasks_reconciliation_20260429.md` (PARKED — that plan was a fuzzy-matcher rebuild; /sweep surfaced Session 169 contemporaneous decision against fuzzy matching + 22% actual coverage vs 8.7% claimed. Premise rejected. This plan is the correct response: tune existing enforcement instead of building new detector.) --- ## Goal Close the bypass in `~/.claude/hooks/enforce_changelog_closes.py` AND wire `close_audit.py` Check 3's existing exit-code into a hard BLOCK at /close Phase 4. Tune existing infrastructure; do not build new fuzzy detection. --- ## Investigation findings (anchor for plan) - Hook lives since 2026-04-26 with bypass `if headers == 0: return True` — gives free pass when an Edit's diff doesn't contain a `### ` header. Multi-Edit composition during /close routinely bypasses. - Check 3 in `close_audit.py` already exits 1 when findings exist, but `~/.claude/skills/close/SKILL.md` Step 4 calls it without checking exit code. The "advisory" status is a missing wrapper, not missing logic. - 78% miss rate confirmed: 1,216 CHANGELOG entries / 267 `**Closes:**` lines = 22% coverage. - Session 174 (today) added Checks 6–10 + Step 0a/0b/4 split — orthogonal. Did not touch the hook, did not promote Check 3. - No prior incident postmortem on the bypass. UNHOOK_LOG.md has no entry. Bypass was BY DESIGN in original 2026-04-18 spec ("count headers in new content"); never documented as a failure mode. --- ## Scope ### In - Replace `enforce_changelog_closes.py` `check(text)` (diff-only) with `check_file()` — reads CHANGELOG.md after Edit, finds today's `## YYYY-MM-DD` block, counts `### ` headers vs `**Closes:**` substring matches. - Fail-closed default with documented escape: `try/except` wrapper; on crash → BLOCK + stderr "Hook crashed; fix or set HOOK_BYPASS=1 to skip." - Add 6 test fixture payloads (`references/TEST_enforce_changelog_closes.json`) + validation script — machine-verified per `feedback_user_never_verifies.md`. - `close/SKILL.md` Phase 4: wrap `close_audit.py --phase post` in exit-code handler. Reconcile prior text "audit can never block close" → "Phase 4 (Closes-line audit) is the only blocking gate; all other audit phases are advisory." - Phase 3 retro-fix: surface and retro-fix the 14 historical `[x]` tasks lacking Closes lines (TASKS.md L1273) BEFORE Phase 2 BLOCK ships — without this, first /close after eviction trips on legacy debt. - Round-trip end-to-end test: hook → Check 3 → `--scan-done` → eviction. ### Out - ~~New fuzzy CHANGELOG↔TASKS matcher~~ — Path A; rejected by /sweep. - ~~New `/reconcile-tasks` skill~~ — same. - ~~Auto-mark TASKS.md `[x]` from CHANGELOG~~ — TASKS.md L1275; orthogonal future build (`BUILD_tasks_md_simplification_20260426.md` Phase 2 owns). - ~~Backfill of all 949 historical entries lacking Closes~~ — separate one-shot project, Phase B+ candidate. --- ## Files aff …[truncated at 3072B — open file for full content]

# BUILD: CHANGELOG → TASKS.md Auto-Close Scanner **Created:** 2026-04-18 **Status:** APPROVED — ready to build **Task:** L872 HIGH — [BUILD] CHANGELOG → TASKS.md completion scanner **Risk:** LOW — additive only, suggestions-only output, no auto-close path --- ## Context The task list has 140+ open tasks requiring manual archaeology each session to close completed work. `task_hygiene.py` already has a `find_done_not_marked()` function that reads CHANGELOG, but it was disabled due to keyword-matching false positives — "building toward X" and "X is built" look identical to a keyword scanner. The root problem: tasks are closed manually by Claude after spotting the match. A task list that can't update itself becomes a permanent backlog. **Chosen approach:** Add a structured `**Closes:**` line to CHANGELOG entries at write-time. The scanner reads this explicit field — no NLP, no ambiguity, zero false positives. Machine control at write-time (Claude writes the reference) + machine detection at read-time (scanner surfaces the match). --- ## What We're Building ### Part 1 — CHANGELOG entry format extension (write-time) Add an optional `**Closes:**` line to CHANGELOG entry format. Claude writes it whenever a CHANGELOG entry resolves a TASKS.md item. **Format:** ``` **Closes:** [FIX] combo dashboard ctrl_n display and sort order ``` - Value = the task title text (copy from TASKS.md `**[TAG] title**`, strip `**` markers) - Multiple closes on one entry: one `**Closes:**` line per task, or comma-separated - Optional — not every CHANGELOG entry closes a task (pipeline runs, data pulls, etc.) - Enforce via CLAUDE.md standing rule (Part 3 below) ### Part 2 — Scanner in task_hygiene.py (read-time) Add `scan_changelog_closes(tasks_path, changelog_path)` function: 1. Parse CHANGELOG.md for all `**Closes:** <text>` lines (full file, not just top 200) 2. For each close reference, match against open tasks by: - Tag match: `[FIX]` in closes text → search `[FIX]` open tasks - Title overlap: 3+ specific words from closes text appear in task title (use `_specific_words()`) 3. Skip tasks already marked `[x]` — no re-surfacing of already-closed items 4. Return only `HIGH` confidence matches (tag + 3+ words). Drop MEDIUM/LOW — too noisy. 5. Output format (extend `--scan-done` flag): ``` SUSPECTED DONE (from CHANGELOG **Closes:** entries): [HIGH] [FIX] combo dashboard ctrl_n display — matched "combo dashboard ctrl_n" → Close this task? Run with --auto-close to mark done, or close manually. ``` **Key guardrails:** - Never auto-marks tasks done — suggestions only - HIGH confidence threshold only (tag + 3+ words) - Skips CRITICAL tier tasks entirely (too risky to suggest-close automatically) - Skips `[x]` tasks (no re-surfacing) - Run on demand: `py scripts/task_hygiene.py --scan-done` - Optionally wire into session digest (after validation — not at first build) ### Part 3 — CLAUDE.md rule (process enforcement) Add to `investments/CLAUDE.md` under Standing Rules (after the session-event triggers table): > **Standin …[truncated at 3072B — open file for full content]

# BUILD: CHANGELOG ↔ TASKS reconciliation — recurring audit command Created: 2026-04-29 Status: SEALED — pending /risk-check + user approval Companion (deferred to Path B, separate build): session-close machine control that BLOCKS/WARNs when a CHANGELOG entry is added without a `**Closes:**` line for an open task it textually overlaps with. --- ## Goal Make CHANGELOG↔TASKS.md reconciliation a recurring, low-false-positive audit command so already-completed tasks stop accumulating in TASKS.md. User-observed symptom (2026-04-29 Session 173): cruise-candidate tasks that were "quite sure all also completed" — system never surfaced them. Hard data: 1,214 CHANGELOG entries, only 106 (8.7%) carry the `**Closes:**` field that `--scan-done` reads. The fuzzy scanner that would catch the other 91% (`find_done_not_marked` at `task_hygiene.py:558`) **already exists but is explicitly disabled** at line 826 due to false positives ("CHANGELOG vocabulary overlaps too much with open task titles"). Path A (this build) = re-enable + tighten the fuzzy scanner; ship it behind a tiered-confidence command. Path B (deferred) = forward-looking machine control — see Companion above. --- ## Scope ### In - Strengthen `find_done_not_marked()` — replace single-line scan with **paragraph-window scan** scoped to recent CHANGELOG entries, scored by: - count of `_specific_words` matches (existing — already strips `_DOMAIN_STOP`) - **rare-word weighting** — add a corpus-frequency penalty so high-recurrence words ("phase", "gate", "shadow") don't trigger; rare words ("CHECK 64", "STEP_OUTPUTS", "Phase 5a") count more - co-occurrence with done verb (`done|complete|built|closed|fixed|added|created|wired|live|shipped|ships`) - **3-tier confidence**: HIGH (≥4 rare-weighted matches + done verb in same entry), MEDIUM (≥3), LOW (≥2 + done verb within 3-line window) - Wire as new flag: `task_hygiene.py --reconcile` — prints HIGH-confidence by default; `--reconcile --include-medium` and `--reconcile --include-low` for wider sweeps - Output format: per-task block — task title, matching CHANGELOG entry header, confidence, 1-line evidence ("matched: STEP_OUTPUTS, CHECK 64, Phase 5a, banner"). Advisory only — never modifies TASKS.md. - Slash skill: `~/.claude/skills/reconcile-tasks/SKILL.md` — wraps the command; default = HIGH-confidence; documents `--include-medium` for periodic deeper sweeps. - Update `references/TASK_HYGIENE_LOGIC.md` — add a "Reconciliation" section describing the 3-tier model. ### Out — DEFERRED ENTIRELY (Path B, separate build) - ~~Session-close machine control that BLOCKS/WARNs on CHANGELOG entries missing `**Closes:**`~~ — needs its own /risk-check (touches close skill + governance flow). File as separate task. - ~~Auto-marking `[ ] → [x]` based on reconciliation matches~~ — duplicates Phase 2 of `BUILD_tasks_md_simplification_20260426.md`. Reconciliation is suggestion-only; that build owns the apply path. - ~~Wiring into session_digest at session start~~ — start as on-demand command. Promote to digest only after 2+ weeks o …[truncated at 3072B — open file for full content]

# BUILD: /projects skill — terminal print + mobile drill-down Created: 2026-04-29 Status: SEALED — pending risk-check --- ## Goal One terminal command (`/projects`) that prints all in-flight, parked, blocked, built-not-verified, and unscoped work in clean emoji-prefixed sections — so the user can see the full project board at a glance. Phase 2 adds a mobile dashboard with click-through drill-down to full build/project scope detail. --- ## Scope ### In scope (Phase 1 — terminal) - New skill: `~/.claude/skills/projects/SKILL.md` - New script: `projects/investments/scripts/print_project_view.py` - Reads (read-only): `TASKS.md`, `references/BUILD_*.md` filenames, `CHANGELOG.md` (last 30d), `CURRENT.md`, `PROJECT_MAP.md` - 7 categories with emoji headers + counts - Fuzzy filename match between TASKS.md text and `BUILD_*.md` filenames (first 4 keywords) - DEPRIORITIZED items appear but capped at **max 1 per category** to avoid old-build flooding - Sort within sections by tier (CRITICAL → HIGH → VERY HIGH → MEDIUM → LOW → VERY LOW → NEEDS REVIEW → DEPRIORITIZED) - "Needs scope" filter: CRITICAL / HIGH / VERY HIGH only - "Built not verified" filter: suppress items shipped >30d ago - PROJECT_MAP.md row registering the new script ### In scope (Phase 2 — mobile, deferred 2 weeks) - New script section in `generate_mobile.py::generate_projects_card` - New file: `mobile/projects.html` - Collapsible sections matching terminal categories - **Drill-down:** clicking an item opens its full detail — `BUILD_*.md` content rendered for plan-backed items, full task-text + tier + age for task-only items - Pipeline wiring deferred (manual refresh first; promote per `feedback_shadow_investments_surfacing.md` after 2 weeks reliability observation) ### Out of scope - No edits to `TASKS.md`, `CURRENT.md`, `CHANGELOG.md`, or any shared CSV - No pipeline wiring in Phase 1 - No desktop HTML dashboard - No auto-promote / auto-close logic — display only - No new CHECK in `mde_health_check.py` (read-only skill needs no validator) --- ## Files affected ### Created (Phase 1) - `C:/Users/jjwat/.claude/skills/projects/SKILL.md` — skill definition + invocation block - `C:/Users/jjwat/OneDrive/Claude Code_All/Claude Code Terminal/projects/investments/scripts/print_project_view.py` — printer (~200 lines) ### Edited (Phase 1) - `C:/Users/jjwat/OneDrive/Claude Code_All/Claude Code Terminal/projects/investments/PROJECT_MAP.md` — register new script + this BUILD plan ### Created (Phase 2 — separate session) - `C:/Users/jjwat/OneDrive/Claude Code_All/Claude Code Terminal/projects/investments/mobile/projects.html` - New function in `projects/investments/scripts/generate_mobile.py` --- ## Categories — final detection logic | # | Category | Emoji | Detection | Cap | |---|---|---|---|---| | 1 | In the works | 🚧 | `BUILD_*.md` exists AND CHANGELOG has partial-ship entry (e.g. "Phase N SHIPPED" but plan defines more phases), OR text appears in CURRENT.md Interrupted/Open block | none + 1 deprio | | 2 | Parked | ⏸ | `TASKS.md` PARKED section + plans flag …[truncated at 3072B — open file for full content]

# BUILD — Type 2 Separation Pilot (separation_calc.py + mobile Build Lab integration) **Date:** 2026-04-25 **Session:** 153 **Sources:** - User feedback post Wave 3b Phase 2 deploy: questioned `lift_pct` semantics, identified Type 1 vs Type 2 separation distinction - Parallel Session 156 build plan: `references/BUILD_separation_pp_universal_20260425.md` (PARKED) - User multi-source insight: *"should have shadow + control + treatment + sold ticker — have everything — limiting our signal here"* **Relationship to parked Universal plan:** Pilot complements, does NOT duplicate. Pilot generates empirical evidence to un-park Universal by answering its 4 open questions through measurement. **Status:** Plan committed; /risk-check fired LOW; SIL deposit done; user "go" given. --- ## 1. Intent Build a reusable utility module for combo-level Type 2 separation (return-based sep_pp), wire it into the mobile Build Lab card as a pilot, and produce a methodology doc that anchors the Universal plan when it un-parks. **Type 1 (current BF):** held=1/sold=0 binary classifier lift over baseline. **Type 2 (this pilot adds at combo level):** avg actual return of combo passers − avg actual return of combo failers. Per-metric Type 2 already exists (`ctrl_sep_pp` in `control_lab_direction_signals.csv`). Combo-level Type 2 does not exist anywhere. ## 2. Why pilot, not full standardization - Universal plan has 4 open questions; pilot answers them empirically by running. - Pilot scope (mobile-only) is throwaway-cheap; data it generates is high-value signal. - Reusable module = leverage when Universal un-parks. - User explicit framing: *"more experiment, prototype and nothing to lose situation."* ## 3. Scope **IN scope:** - New `scripts/separation_calc.py` reusable utility module - New `references/SEPARATION_METHODOLOGY.md` explanation doc - Multi-source returns blending: FORMAL > SHADOW > control > holdings (priority order) - Per-source attribution surfaced (so user can sanity-check coverage) - Coverage_pct flag if <30% of combo passers have return data - Mobile Build Lab card displays `sep_pp_real` alongside existing `lift_pct` - Top 10 L2 combos + top 10 L1+L2 hybrid combos get the new field - CLI for standalone testing: `py scripts/separation_calc.py --combo "57,504"` **OUT of scope (intentional):** - Desktop combo dashboard updates (deferred to Universal plan when un-parked) - BF engine modification (`mde_04_brute_force.py` UNCHANGED) - run_pipeline.py wiring (deferred — JSON refreshes via existing pipeline call) - Column rename collisions (Universal Phase 4) ## 4. Source Files & Multi-Source Returns Strategy **Returns data sources (priority order):** 1. `data/processed/recommendation_events.csv` — FORMAL recs with `return_pct` populated 2. `data/processed/recommendation_events.csv` — SHADOW recs with `return_pct` populated (skip if PENDING per Session 153 CSL finding) 3. `data/processed/control_group.csv` — `excess_vs_spy` for ~500 tickers 4. `data/processed/holdings_vs_control.csv` — return for held positions **Dedupe by tic …[truncated at 3072B — open file for full content]

# BUILD — Universal Return-Separation-PP Sweep + Naming Standard **Created:** 2026-04-25 **Last Revised:** 2026-04-25 (Revision 3 — Phase 1b sort flip folded into Phase 1 after 2nd /risk-check; user approved combined ship) **Status:** EXECUTING — Phase 0 + Phase 1 (now including sort flip) approved + in-flight. Phases 1.5 / 2 / 3 / 3.5 DEFERRED — revisit after Phase 1 lands + 1 week observation. **Priority classification:** HIGH (data-clarity growth lever, not a CRITICAL data-integrity bug). **Owner:** Claude (under user direction) --- ## File Discoverability (per CLAUDE.md standing rule) 1. **Who needs this and when?** Claude at build time, on every session that touches BF / gate alts / control lab / dashboards / METRIC_FORMULAS until build is complete. 2. **How will they find it?** Linked from CURRENT.md "In Progress" row; referenced in TASKS.md task entry; this file lives in `references/` (governance-discoverable). 3. **Right location?** Yes — multi-phase build plan, references/BUILD_*.md is canonical per memory `feedback_build_plan_location.md`. 4. **Existing file covers this?** No. METRIC_FORMULAS.md and SYSTEM_STANDARDS.md touch metric quality but not return-separation methodology. 5. **Process or machine control?** Phase 0 produces a governance doc (process); Phases 1–3 add machine-enforced columns surfaced in pipeline output (machine control). Phase 4 (rename collisions) is deferred. --- ## North Star Tie-In This build serves the causal chain directly: **separation power IS the mechanism by which gate strength becomes returns** (per GOALS.md). Today the system measures separation at the per-metric L1/L2 level (`ctrl_sep_pp`) but **not at the combo or gate level** — meaning gate composition decisions are flying half-blind on the most important quality signal. Fixing this is North-Star-aligned. --- ## The Problem (one paragraph) The user's preferred metric — `return_separation_pp` (avg return of tickers above threshold MINUS avg return below threshold, direction-aware) — is computed today only in `control_lab_direction_signals.csv` as `ctrl_sep_pp`, per individual metric. It does **not exist at the combo level** in BF, gate alternatives, or anywhere else. Worse, two existing column names collide: BF's `return_spread_pct` is a held/sold *classifier lift* (not a return separation), and the registry's `distance_pct` is a *threshold-to-median weight* (also not a return separation). Both currently live in dashboards under names that imply they ARE return separation. This makes the system foggy on its most important signal. --- ## Definitions Lock (canonical — to be enshrined in METRIC_LANGUAGE_STANDARD.md) | Term | Definition | Where computed today | |---|---|---| | `win_rate` | Of tickers passing the threshold, % that beat baseline (precision). | BF `win_rate_pct`, control lab | | `win_rate_recall` *(optional)* | Of all winners, % that sit above threshold (recall, direction-aware). | NOT computed | | `return_separation_pp` ⭐ | Avg return(above threshold) − Avg return(below threshold), direction-aware. …[truncated at 3072B — open file for full content]

# BUILD: SIL — Intent-vs-Actual Closure Plan Created: 2026-05-16 Status: SEALED — pending risk-check > Supersedes (folds in): `BUILD_sil_p2_enforcement_20260508.md` (→ Phases 0+1) and `BUILD_sil_trusted_brain_20260423.md` Phases 0-4 (→ reframed as Phases 0-3 here). Phase 6 of trusted_brain (Decision Memory + Outcome Tracking) deferred until after this plan ships. --- ## Goal Bring the Signal Intelligence Layer from its current state ("scoring reporter running in parallel to decision logic") to its declared intent in `SIGNAL_INTELLIGENCE_LAYER.md` + `SIL_CHARTER.md` ("trusted brain that informs decisions, with audit-trailed evidence"). Six sequenced phases, each with a hard acceptance gate. ## Scope **In scope:** - Fix the `signal_intelligence.py:443-445` magnitude-guard chokepoint contaminating 11+ consumers. - Ship P2 (evidence_refs) enforcement end-to-end: writer wiring, historical backfill, WARN→BLOCK promotion. - Activate P6 authority hierarchy by refactoring `gate_advisor.py` and `direction_advisor.py` to consume `AUTHORITY_RULES` instead of hardcoded weights. - Wire P3 decision lineage: persist `decision_lineage.csv` populated by `capture_session_decisions()`. - Activate synthesis output (SYSTEM rows) + surface `sil_grand_summary.json` content in session digest + resolve mobile/dashboard discoverability of SIL. - Promote SIL from advisory → autonomous powers (auto-ADD candidates, direction-flip warnings, VALIDATED SIGNAL surfacing) with one-time user sign-off. **Out of scope:** - Gate Alternatives SIL-Driven Fix v1/v2 specifics (per user direction this session). - Phase 6 of trusted_brain plan (Decision Memory + Outcome Tracking) — revisit after this plan completes. - New SIL mobile/dashboard cards beyond re-promoting or retiring the existing one. - Phase 7 (reserved / not yet detailed in any doc). ## Files affected **Edited:** - `scripts/signal_intelligence.py` — magnitude guard (L443-445 region + import), writer wiring for evidence_refs/threshold_for_action/expiry_sessions, surface `capture_session_decisions()` output. - `scripts/sil_charter.py` — flip P2 WARN→BLOCK; promote AUTHORITY_RULES from PLACEHOLDER → ENFORCED. - `scripts/mde_gate_advisor.py` — refactor `compute_evidence_score()` (~L274) to consume AUTHORITY_RULES. - `scripts/direction_advisor.py` (or equivalent file — confirm name in Phase 2) — refactor to read SIL confidence per AUTHORITY_RULES at decision time, not display-only at L59-60. - `scripts/mde_health_check.py` — add CHECK NN (P6 advisor compliance: grep advisors for hardcoded weights → 0 hits), CHECK NN+1 (P3 decision_lineage.csv freshness + row schema), CHECK 51 (P4 synthesis tripwire). - `scripts/session_digest.py` — surface `sil_grand_summary.json` content as 🧠 SIL GRAND SUMMARY block. - `mobile/mobile_registry.py` — flip SIL entry `nav: True` OR retire with reason logged in UNHOOK_LOG.md. - `metric_discovery_engine/scripts/run_pipeline.py` — confirm SIL step ordering still valid; no new step expected. **Created:** - `scripts/backfill_sil_evidence_refs.py` — one-shot migratio …[truncated at 3072B — open file for full content]

# BUILD: SIL — Intent-vs-Actual Closure Plan (v2.1 — post-/sweep re-validation) Created: 2026-05-16 (v2) · Patched: 2026-05-22 (v2.1) Status: SEALED v2.1 — un-parked 2026-05-22, pending /risk-check before Phase 0 first Edit > **v2.1 patch notes (2026-05-22 — post-/sweep re-validation after un-park):** > - **[Δ1] CHECK number collision fixed:** CHECK 51/52/53/54/55 were ALL already claimed (Gate Alts 51-53, Control Lab 54, GDE 55). Renumbered to **73/74/75/76/77** (next free range, CHECK 72 added 2026-05-22). Mapping: P4 synthesis tripwire → **73**, P2 BLOCK enforcement → **74**, P6 hardcoded-weights audit → **75**, P3 decision_lineage freshness → **76**, P5 autonomous output guard → **77**. > - **[Δ2] Phase 0 partial-ship recorded:** > - Step 0.1+0.2 magnitude guard ✅ **SHIPPED 2026-05-21** at `signal_intelligence.py:451` (load_rec_events) + `:500` (load_shadow_rec_events), import at `:80`. Note: actual line numbers shifted from plan's original `L443-445` to `L451`. > - Step 0.6 (5 bare `os.replace()` → `_atomic_replace()`) ✅ **SHIPPED** — grep for `os.replace(` in `signal_intelligence.py` returns 0 hits. CHECK 42 passes. > - Remaining Phase 0 work: 0.0a snapshot, 0.0b ground-truth count, 0.2b unit test (new), 0.3+0.4 evidence_refs deposit writer + backfill, 0.7 UNHOOK_LOG entries, 0.8 CHANGELOG, 0.9 PROJECT_MAP. > - **[Δ3] New Step 0.2b:** create `scripts/test_sil_magnitude_guard.py` (~10 lines) to satisfy Step 0.1+0.2 acceptance retroactively. 1 fixture row with `return_pct="0.0"` → assert row skipped. Smoke runs <2s. > - **[Δ4] Line-ref drift refreshed:** > - `mde_gate_advisor.py` hardcoded weights: plan said `L70-72` → actual `L62-65` (MPS_WEIGHT=40, COHENS_D_WEIGHT=20, CONTROL_WEIGHT=20, BF_WEIGHT=20). > - All other line refs (synthesize_learnings L1330-1450, capture_session_decisions L1475+, AUTHORITY_RULES L240-267) verified intact. > - **[Δ5] Governance hooks added:** > - Phase 0 Step 0.0a first Edit gated on `/risk-check` pass (explicit). > - Per-phase: visible Pre-Fix checklist (Steps 1–6) before first Edit; CURRENT.md live update after each ship; `py scripts/refresh_command_center.py` after every CHANGELOG edit. > - Phase 4.4: DASHBOARD_REGISTRY entry required for `session_digest 🧠 SIL GRAND SUMMARY` block. > - AUTHORITY_RULES (Phase 2) clarified as data-driven const: Q1 user input → edit `sil_charter.py:240-267` only, no consumer code edits required. > > **v2 deltas (vs v1 sealed earlier 2026-05-16):** Folded in all 🔴 BLOCKING + 🟡 SHOULD-FIX gaps surfaced by /sweep. Phase 0 gains pre-mutation snapshot, ground-truth re-measurement of P2 violation count, migration of 5 bare `os.replace()` calls, and explicit UNHOOK_LOG entries. Phase 3 gains an explicit "wire capture_session_decisions to disk" step + new schema module. CHECK numbers assigned (51, 52, 53, 54, 55 — superseded by v2.1 to 73-77). All file paths corrected (`mde_direction_advisor.py`). Unit tests added per phase. CHANGELOG action made explicit per phase acceptance signal. AUTOMATION.md A43 + AUDIT_REGISTRY ent …[truncated at 3072B — open file for full content]

# BUILD: SIL Phase 1 — P2 Enforcement (evidence_refs end-to-end) **Created:** 2026-05-08 **Status:** SEALED — pending risk-check **Supersedes:** `BUILD_sil_p2_enforcement_20260424.md` (2026-04-24, never executed — incorporated wholesale below with 2026-05-08 deltas marked) **Tier:** HIGH **Estimated effort:** ~2.5h shipping (single session, must not split) + 2 days observation before P2→BLOCK **Risk:** LOW (additive fields, supersede semantics already exist, atomic write covered) --- ## Goal Close the SIL P2 prerequisite gap: every auto-deposited row in `sil_learnings.csv` must carry `evidence_refs` (and the other 5 Phase-1 fields) at write time, so SIL output becomes auditable and can later drive autonomous decisioning. One sentence: **make CHECK 50 P2 stop firing on every active row by populating evidence_refs at write time and back-filling the 114 historical rows that lack it.** --- ## Why this build exists (carried from 2026-04-24) CHECK 50 P2 currently fires on **100% of active SIL rows** (114/114 in the 2026-04-24 baseline; today's /inspect run surfaced 10 fresh warnings, but the underlying violation count is unchanged). A health check that fires on every row is broken signal — it tells the user/Claude to "remember to ignore it." **Root cause:** `signal_intelligence.py:deposit_sil_learnings()` (line ~914) builds every auto-deposit row with only the original 7 fields and skips the 6 Phase-1 columns added in Session 147 — including `evidence_refs`. The schema has the column; the writer ignores it. Charter at `sil_charter.py:315–332` correctly flags the violation. **Decision context (2026-04-24, reaffirmed 2026-05-08):** User wants SIL output to drive more systems beyond pinned alts. P2 enforcement is the prerequisite — without `evidence_refs` populated, every SIL-driven decision is unauditable. Memory `project_sil_autonomous_direction.md` records this as the gating precondition for any new autonomous wiring. **2026-05-08 additions:** Phase 3 Mobile cards + SIL deposit injection (TASKS.md HIGH line 845) is explicitly DEFERRED waiting for this — every day of delay extends the block. --- ## Scope (in / out) ### In scope - **Piece 1:** Writer wiring in `signal_intelligence.py` — populate all 6 Phase-1 fields at every `new_rows.append(row)` site in `deposit_sil_learnings()` - **Piece 2:** One-shot backfill script `migrate_sil_p2_backfill.py` for the 114 historical rows - **Piece 3:** P2 WARN→BLOCK promotion in `sil_charter.py:145` after ≥2 clean pipeline runs - P5 fields (`threshold_for_action`, `expiry_sessions`) populated as a **side effect** of Piece 1 (no extra work — they're written together) ### Out of scope (deferred) - **P5 promotion to BLOCK** — separate decision after observation; today's plan only populates the fields, does not flip P5 severity - P1 enforcement (no dashboard restatements) — separate task; needs `is_dashboard_template` regex tested against current rows first - Any specific SIL output → autonomous wiring — that's the *next* decision after this build clears the prerequisite - An …[truncated at 3072B — open file for full content]

# BUILD: Sub-Project Framework v3a — Plug-and-Play + Sync + Unified Mobile Hub (v2) **Created:** 2026-05-10 **Status:** ⏸ PARKED — 2026-05-10. /sweep + /risk-check surfaced strong evidence that framework is premature: B2 dormant 37 days; 2 prior framework plans (2026-04-19, 2026-05-09) never shipped; cost-benefit NET NEGATIVE if PE is sole consumer (-13 hr at sub-project #2; payoff at sub-project #3-4); Phase 9 mobile registry not actually additive (Hard Rule 1 violated); v3a alone ships scaffold without UX (v3b is 2-week prereq for "PE-focused session"); 20-25 hr realistic effort vs 6-8 hr for PE one-off path. **Revisit trigger:** when sub-project #2 actually arrives (concrete user intent, not aspirational projection). Replaced by `BUILD_pe_dashboard_oneoff_20260510.md` for the immediate PE work. Original sealed pending status preserved below for the eventual revisit. **Original status:** SEALED — pending /risk-check **Supersedes:** `BUILD_subproject_framework_v3a_20260510.md` (v1, same session) — folds in 5 user-context additions: v3b scope clarification, /refresh-template skill (evergreen template), new-system-detection mode B, mobile hub section config, additivity reinforcement **Companion task split:** v3a (this plan — framework + cloning + sync + mobile hub) | v3b (separate future plan — /main + /<slug> switcher commands + project-aware skill refactor across 31 skills + 5 hooks) **Related stale doc:** `DECISION_pe_sub_folder_scope_20260510.md` — framework decisions superseded; PE-domain notes carry forward to a future `BUILD_pe_subproject_spinup_*` plan --- ## Goal Build a parent-owned framework that lets the investments project clone, govern, and continuously sync sub-projects with mirrored infrastructure — and do so additively, with zero interference to existing systems — so any new sub-project gets investments-grade governance out of the box, the master template stays evergreen, and existing sub-projects can both detect drift AND detect newly-added canonical systems they don't yet have. **Non-goal (deferred to v3b):** Custom slash-commands for switching sub-project context (`/main`, `/<slug>`), `.active_project` state file, refactor of 31 existing skills + 5 global hooks to be project-aware. This v3a plan delivers the **framework + cloning + sync + mobile hub**; v3b delivers **navigation between projects in a session**. Independent; either can ship first; v3a unblocks PE so it's prioritized. --- ## Guiding Constraint — Additivity (HARD RULES, user requirement 2026-05-10) **This build must be very low risk, very additive, functional, and evergreen. It must NOT interfere with current systems.** Four hard rules every phase must satisfy: 1. **No modification of existing logic in any current script.** Edits to existing files (run_pipeline.py, mobile_registry.py, mde_health_check.py, CLAUDE.md, settings.json) are restricted to APPENDING new code/sections. Existing functions, classes, steps, CHECKs, standing rules, and hook entries stay byte-identical. 2. **Failure of any new component must not affect an …[truncated at 3072B — open file for full content]

# BUILD: Sub-Project Navigation v2 — Write-Guard + Minimal Scaffold (Option C) Created: 2026-05-09 Status: SEALED — pending risk-check Supersedes: `BUILD_subproject_navigation_20260509.md` (v1 — 42/100 SAFE but premise half-wrong per /sweep) --- ## What changed from v1 v1 tried to do four things at once: navigation skills, registry/state, hook generalization, write-guard, scaffold, governance cloning. /sweep surfaced 4 BLOCKING gaps: 1. `.active_project` is unused by 31/38 skills + 5 global hooks → write safety not actually delivered 2. session_context.py SessionStart fires before `/b2` can run → cold-start always wrong context 3. Cloning investments governance imports 70–85% rule noise into PE/home/work-items domains 4. Existing 2026-04-19 v1 plan at TASKS.md:1301 (≈80% same scope) was never executed **v2 reframes around the actual user need:** "prevent overlap and overwrite of wrong CHANGELOG/CURRENT/TASKS in sub-projects." The real safety is at write-time, not at navigation-time. v2 ships a write-guard hook + minimal scaffold + filesystem-discovery `/projects-help`. No registry, no state file, no skill refactors, no governance clone. ~3 hours work, near-zero risk. User accepted Option C 2026-05-09 with a known-incomplete caveat: "this is not close to where I need it to be — add a task to get this better." That follow-up task is added to TASKS.md alongside this plan. --- ## Goal Deliver cross-project file-write safety + discovery of sub-projects with the smallest possible footprint. Existing B2 keyword UX preserved unchanged. Future, larger nav refactor tracked as a separate task. --- ## Scope **In:** - 1 PreToolUse hook: `~/.claude/hooks/subproject_write_guard.py` - 2 skills: `~/.claude/skills/new-subproject/SKILL.md`, `~/.claude/skills/projects-help/SKILL.md` - Minimal 3-file template at `~/.claude/skills/new-subproject/template/`: `CURRENT.md`, `TASKS.md`, `CHANGELOG.md` only - `~/.claude/settings.json` — add PreToolUse entry - `restore_claude_setup.py` — register the new hook + 2 skills (single edit, single phase) - CHANGELOG entries per phase ship **Out:** - `/main`, `/b2`-style explicit nav slash-commands (existing B2 keyword hook works fine) - `.active_project` state file - `~/.claude/project_registry.json` - `project_session_detector.py` (b2_session_detector.py stays) - Refactor of any existing skill or hook - Governance cloning (no GOALS/SYSTEM_STANDARDS/CLAUDE.md/PROJECT_MAP scaffolded) - Workspace-level rollup files - Mobile registry changes --- ## Files affected ### Created - `~/.claude/hooks/subproject_write_guard.py` — PreToolUse hook (~40 LOC) - `~/.claude/skills/new-subproject/SKILL.md` - `~/.claude/skills/new-subproject/template/CURRENT.md` - `~/.claude/skills/new-subproject/template/TASKS.md` - `~/.claude/skills/new-subproject/template/CHANGELOG.md` - `~/.claude/skills/projects-help/SKILL.md` - This plan file ### Edited - `~/.claude/settings.json` — add PreToolUse hook entry - `restore_claude_setup.py` — single addition: hook constant + skill scaffold writes - `investments/CHANGEL …[truncated at 3072B — open file for full content]

# BUILD — TASKS.md Self-Maintaining Closure Loop **Date:** 2026-04-26 **Author:** Session 170 (long-term plan track) **Status:** Draft v3 — final design after governance + history + code audit. Awaiting user approval + /risk-check before Phase 1. **Companion to:** parallel-session tactical work (170c/f/h/i/j) — already shipped, complementary, out of scope here. --- ## 1. Context — the value frame TASKS.md should be a **healthy, self-updating machine** — when work completes, the file reflects it automatically, with no manual `[ ] → [x]`, no manual eviction, no per-task cleanup. Every other property (cap math, tier flow, funnel stages) is supportive plumbing. **This plan exists to close the completion loop.** ### Design constraints (user-stated, non-negotiable) 1. **Long funnel is fine — even good.** 9 sections, more if needed. Don't simplify section count. 2. **DEPRIORITIZED is a parking lot, not a graveyard.** Items there are reviewable + refillable. **Never deleted, never auto-evicted.** 3. **Only `[x]` (truly completed) work gets evicted.** No other state ever leaves the active surface. 4. **`/task-hygiene`, `/buckets`, `/task-promote`, `/close`, `/align` all keep working.** Zero command rewiring. 5. **The matcher (Session 169 exact-prefix logic) is not touched.** It works. ### What's already shipped today (parallel sessions, OUT OF SCOPE here) | Session | Ship | Effect | |---|---|---| | 170 | `enforce_changelog_closes.py` hook | Blocks CHANGELOG writes without `**Closes:**` line | | 170c | `--archive-stale-review N=21` ran | 27 NEEDS REVIEW → DEPRIORITIZED | | 170f | `task_hygiene.py` UTF-8 stdout wrapper | Fixed UnicodeEncodeError crash | | 170h | `_DEMOTION_CASCADE → DEPRIORITIZED` extension | 100 items demoted; **all tiers within cap** | | 170i | `--refill-review N` flag | Pulls items DEPRIORITIZED → NEEDS REVIEW | | 170j | `close_audit.py` Check 3 fix (multi-session capture + closure-note strip) | False positives killed | **Current `/buckets` state — ZERO overcap.** CRITICAL 1/2 · VERY HIGH 4/5 · HIGH 6/8 · MEDIUM 14/15 · LOW 10/12 · VERY LOW 10/10 · NEEDS REVIEW 40/40 · DEPRIORITIZED 174 · BLOCKED 2. ### The remaining gap Closure flow is still half-manual: `--scan-done` is advisory; user manually edits `[ ] → [x]`; closed `[x]` items accumulate in tier sections forever. **That's what this plan fixes.** --- ## 2. Architecture review (governance + history + code, audited via 3 parallel Explore agents) The 9-section funnel is **load-bearing** per Session 116's Task CRM Overhaul: - **NEEDS REVIEW** = entry inbox (cap 40, drains via `--archive-stale-review` + cascade). - **DEPRIORITIZED** = uncapped overflow + parking lot. Refillable via `--refill-review N` (170i). - **CRITICAL** = protected (never auto-demoted) per explicit user decision after a prior cascade incident. - **`_DEMOTION_CASCADE`** terminates at DEPRIORITIZED (170h). - **`MIN_PROMO_SLOTS=3`** dead-band requires multi-tier granularity. 7+ scripts hardcode tier name literals (`task_hygiene.py`, `buckets.py`, `refresh_command_center.py`, …[truncated at 3072B — open file for full content]

(no additional detail)

(no additional detail)

(no additional detail)

(no additional detail)

(no additional detail)

(no additional detail)

(no additional detail)

(no additional detail)

(no additional detail)

(no additional detail)

(no additional detail)

(no additional detail)

(no additional detail)

(no additional detail)

(no additional detail)

(no additional detail)

(no additional detail)

(no additional detail)

(no additional detail)

(no additional detail)

(no additional detail)