diff --git a/plugins/linkedin-studio/docs/hardening/plan.md b/plugins/linkedin-studio/docs/hardening/plan.md index 6fc9fd1..24998c2 100644 --- a/plugins/linkedin-studio/docs/hardening/plan.md +++ b/plugins/linkedin-studio/docs/hardening/plan.md @@ -1,571 +1,123 @@ -# LinkedIn Studio — Command Hardening Plan +# LinkedIn Studio — Command Hardening Plan v2 (interactive quality gate) -> **Plan quality: B+** (87/100) — APPROVE_WITH_NOTES (post-revision; plan-critic REVISE -> blockers + majors addressed — see Revisions) +> **Supersedes the v1 autonomous plan** (recoverable in git history: commit `2f90880`). +> Motivated by the 2026-05-31 S2 fabrication incident: the v1 method (autonomous, prose-heavy +> 5-step per command, per-command Opus reviewer swarm, `/trekreview` gate) produced **confident +> fabrications** (Claude wrote SIMULATE/EVALUATE/HARDEN narratives — char counts, line numbers, +> before/after — for files it had not actually observed) and **burned context without value** +> (full-file re-reads + long logs + reviewer agents for what are usually 0–1-line edits). A large +> parallel tool-batch also destabilized the tool-result stream (out-of-order / duplicated / +> truncated returns). Nothing false shipped — the failed edits + empty greps + real reads caught +> it, and it was reverted — but the method had to change. > -> Generated by trekplan v2.0 on 2026-05-31 — `plan_version: 1.7` -> -> Source brief: `docs/hardening/brief.md`. Predecessor: `docs/remediation/` -> (S1–S17, complete, commit `2633d32`). Research: none (`research_status: skipped` -> — the 2026 algorithm bar was triangulated in remediation research 01–03 and is -> frozen in `references/algorithm-signals-reference.md`). +> **v2 is a slow, dialogic, human-in-the-loop gate: ONE command per session, the operator as the +> truth source, every mechanical claim tool-grounded.** -## Context +## Intent -The remediation made every claim honest, wired every orphan agent, rebuilt the -lint, and reconciled the algorithm bar — *structure, correctness, honesty*. It did -**not** exercise each command's workflow end-to-end and judge **the quality of what -it produces** against the command's own stated intention. A command can be -structurally correct (right frontmatter, right agent wired, lint-green) and still -under-deliver: a step that under-determines the next move, a question that yields a -weak answer, a prompt that produces generic output, a missing graceful-degradation -path. +Harden each of the 29 commands to its own intent — **one command per session** — with the +operator in the loop. Last quality gate before GUI/M0. Brief: `brief.md`. -This is a **hardening phase**: a per-command pass that simulates a realistic -invocation, judges the result against intention + the 2026 algorithm bar + the -content-quality rules, and tightens the command definition where it falls short. It -is the last quality gate **before** the GUI/M0 track. The operator tests the -commands **live** in parallel; this phase is the complementary intention-fidelity -pass from the definition side, converging through a field-notes inbox. +## The locked interactive method (identical every session) -Scope is deliberately bounded (locked forks): **intention-fidelity + prompt-quality -only** — no structural redesign, no new features, no GUI/M0, no re-litigation of the -algorithm bar. The command set stays 29/19/25/6. +Per command, 8 steps. **Claude STOPS and waits at each "agree" point — no autonomous run-ahead:** -## Architecture Diagram +1. **Agree intention** — show the command's own description + its journey role; agree in one line. +2. **Agree test-method** — the class predicates (below) + what the simulation must prove. +3. **Agree persona** — who invokes + with what input. +4. **Run the test** — tool-grounded mechanical facts (each with `file:line`) + a GROUNDED persona + simulation (a real invocation → concrete output, anchored to the shown file region). +5. **Talk** — the operator's judgment is the truth source. +6. **Agree improvements** — surgical, or "no change". +7. **Implement** — Edit → grep-confirm landed → show the confirmed diff. +8. **Move on** — run the end-of-session ritual; the next command is its own session. -```mermaid -graph TD - subgraph "Hardening pass — 5-step method per command, 8 journey-grouped sessions" - M["docs/hardening/brief.md
(method contract)"] --> LOG["docs/hardening/log.md
(per-command audit trail,
unique anchor per entry)"] - FN["field-notes.local.md
(operator live findings, optional)"] -.steers priority.-> LOG - REF["references/algorithm-signals-reference.md
+ content-quality rules
(the frozen bar)"] --> EVAL +### Two guards (non-negotiable) - subgraph "Per command" - INT["1 INTENT"] --> SIM["2 SIMULATE
(persona + CONCRETE output)"] - SIM --> EVAL["3 EVALUATE
4 axes (mechanical predicate per type)"] - EVAL --> HARD["4 HARDEN
(edit commands/<x>.md, surgical)"] - HARD --> VER["5 VERIFY
lint + before/after + counts"] - end +- **Read-and-show BEFORE simulate** — the simulation is anchored to observed text, shown to the operator. +- **Every mechanical claim from a tool** — char count, ✓/✗, line number: shown from a tool before stated. - VER --> CMD["commands/*.md (29)"] - CMD -. invokes .-> AG["agents/*.md (19)"] - VER --> GATE["per-session gate:
test-runner.sh (Failed:0) + node --test
+ /trekreview ALLOW — reviewer adjudicates
EVERY hardened command's before/after"] - GATE --> PUSH["commit own files → push (Forgejo)"] - end -``` +### Anti-drift through dialogue -## Codebase Analysis +The per-command dialogue IS the drift-reduction mechanism: Claude **proposes** at each agree-point +(intention / test-method / persona / improvement) and **waits** for the operator to confirm or +correct before proceeding. Frequent grounding checkpoints are what keep Claude from confabulating — +never race past an agree-point. -- **Tech stack:** Markdown command/agent definitions (Claude Code plugin); Node.js - (.mjs) hooks + helpers (zero deps); TypeScript analytics CLI (`scripts/analytics/`, - `tsx`); bash structural lint (`scripts/test-runner.sh`, `set -e`, dynamic PASS/FAIL - counters, `EXPECT_*` count-guards); `node:test`. -- **Key patterns:** commands invoke agents via `subagent_type: linkedin-studio:` - (namespaced) — sometimes as a literal frontmatter-style line, sometimes via a - `--type → subagent_type` table (e.g. `headless-review.md`); content commands - auto-copy to clipboard via `clipboard-helper.mjs`; PreToolUse content-quality gate - (`hooks/prompts/content-quality-gate.md`); progressive onboarding (score hidden < 3 - posts, voice guardian < 5 samples). -- **Relevant files (verified, 216 source files):** - - `commands/*.md` — 29 surfaces (inventory confirmed). Agent wiring per command: - - *with `subagent_type:` line:* ab-test→content-optimizer · analyze/report→analytics-interpreter · - batch/pipeline→content-planner,trend-spotter · calendar→post-feedback-monitor · - carousel/quick/react→differentiation-checker · firsthour→engagement-coach,post-feedback-monitor · - newsletter→content-repurposer,editorial-reviewer,fact-checker (+ persona/voice/headless deeper) · - outreach→network-builder · post→content-optimizer,differentiation-checker · - setup→voice-trainer · strategy→strategy-advisor · video→differentiation-checker,video-scripter - - *invokes via a `--type → subagent_type` table (not a bare line):* `headless-review` - → content-reviewer/language-reviewer/fact-reviewer/persona-reviewer (CONFIRMED present, - `headless-review.md:77-79,141-142` — wiring is correct, just a different syntactic form). - - *no agent (prose/CLI/routing — verify intention + a type-specific mechanical predicate, not wiring):* - audit, competitive, create, first-post, import, linkedin, measure, monetize, multiplatform, - onboarding, pivot, profile. - - `references/algorithm-signals-reference.md` — the single source of truth (the bar). - - `references/*.md` (25), `skills/linkedin-studio/SKILL.md`, - `hooks/prompts/content-quality-gate.md`, `scripts/test-runner.sh`, - `hooks/scripts/state-updater.mjs` (`recordFirstHourPlan`, `recordOutreachContact`), - `scripts/analytics/` (`parseOptionalCount`). -- **Reusable code:** the hardening method (brief §"Hardening method"); the four - *mechanically-checkable* content-quality rules (hook 110–140 chars, length bands, - no body links, no buzzwords); the S17 disposition record as the log model. -- **External tech (researched):** none — the bar is frozen; no `/trekresearch`. -- **Recent git activity:** S17 (`2633d32`) closed remediation; plugin v4.1.0, stable, sync 0/0. +### Execution discipline (the S1+S2 incident lessons) -## Implementation Plan +- **ONE command per session.** Never hold multiple command files in context at once. +- **Never batch dependent tool calls in parallel.** Read→Edit→grep→log is strictly sequential. + Large parallel batches destabilize the tool-result stream. +- **Guard greps:** `grep`/`grep -c` exit non-zero on zero matches → append `|| echo NONE` so a + batch sibling isn't torn down. +- **Never write a HARDEN/after claim before the edit is grep-confirmed landed.** +- **If the tool layer degrades** (slow / truncated / empty / replayed returns): PAUSE and resume + in a fresh session. Do not push through a degraded window. -> **Unit of work = one command** (operator: "én og én kommando"). For tractable -> execution + per-session checkpoints, the 29 commands are grouped into **8 -> journey-sessions**; each session-step applies the 5-step method to every command it -> lists and records the result in `docs/hardening/log.md`. -> -> **Log-entry anchor convention (closes plan-critic blockers #1/#2 — substring -> false-positives):** every `log.md` entry begins with a UNIQUE, anchored header of -> the exact form `### /linkedin: — `. Coverage and -> per-step checks grep the anchored header (`^### /linkedin:`), never the bare -> word — so "measure" no longer collides with the "Measure journey" label, nor "post" -> with "post-publish". -> -> The per-command procedure for **every** command in **every** step: -> -> 1. **INTENT** — distil the promise (frontmatter `description` + journey role + -> quality rules + algorithm signals it must honor) → one paragraph under the -> command's anchored `log.md` header. -> 2. **SIMULATE** — pick a concrete persona (default: the plugin's ICP — a solo -> AI-advisor creator, ~1048 followers, "Validation", 5 pillars; PLUS a fresh adopter -> with no voice samples/analytics for graceful-degradation + progressive-onboarding -> paths). Walk the workflow exactly as written; answer its questions; **produce the -> CONCRETE output the command would generate from those exact mock inputs** (the -> actual draft/report/plan/routing-decision text — reproducible from the inputs, NOT -> a vague paraphrase). Log every friction/dead-end/under-determined step. This -> concrete **before** output is the artifact the oracle adjudicates. -> 3. **EVALUATE** on 4 axes → pass/gap each: -> - (a) **intention fidelity** — does it deliver the description's promise? (judgment) -> - (b) **algorithm bar** — consistent with `algorithm-signals-reference.md`? (cite) -> - (c) **MECHANICAL predicate (hard, per command — never "N/A → judgment")**: -> - *emits a post/hook* (post, quick, react, carousel, video, batch, pipeline, first-post): -> hook 110–140 chars · length band (1,200–1,800 std / 150–500 quick) · `grep` no body -> link · `grep` no banned buzzword · topic maps to the 5 pillars. -> - *routing/front-door* (linkedin, create, measure): every routing target the -> simulation emits resolves to a real `commands/.md` (grep). -> - *analytics/CLI* (import, report, analyze, audit, ab-test): the documented -> graceful-degradation path is present (grep) and the directional/saves honesty -> wording is intact. -> - *guided/stateful* (onboarding, setup, calendar, firsthour, monetize, outreach, -> strategy, competitive, profile, multiplatform, newsletter, headless-review, pivot): -> the command's PRIMARY promised artifact is actually produced in the simulation -> (draft / plan / funnel / analysis / checklist / routing) and there is no dead-end -> hand-off string; for agent-invoking ones, the promised `subagent_type` targets -> resolve to real `agents/.md`. -> - (d) **agent-wiring + graceful degradation** — right `subagent_type` (literal line -> OR `--type` table); sensible fallback when an agent/tool/mcp/CLI is unavailable. -> 4. **HARDEN** — edit `commands/.md` (surgical; agent/reference only if intention -> requires). **Stopping rule (anti-gold-plating):** harden until every axis returns -> pass OR a recorded deferral; do NOT add NICE-only polish beyond axis-pass. No -> structural redesign. Record the concrete **after** output for every hardened command. -> 5. **VERIFY** — `test-runner.sh` exit 0 + `Failed: 0` + counts unchanged (29/19/25/6) -> + `node --test` green where touched + the before/after pair shows the failing axis -> now passes. Record before/after + disposition under the command's `log.md` anchor. -> -> **Independent oracle (closes plan-critic #3/#12 — self-grading):** the per-session -> `/trekreview` code-correctness reviewer reads, in `log.md`, the **before/after concrete -> output of EVERY command hardened that session** (sessions are ≤5 commands — tractable) -> and independently judges whether each claimed gap is genuinely closed. The cold -> reviewer — not the author — signs off. A session is not ALLOW until every hardened -> command's before/after has been adjudicated. -> -> **Manifest scope (closes plan-critic #10):** manifests anchor on the guaranteed -> artifact — the `log.md` entry with its unique per-command anchor (proving the command -> was *processed*). The **actual command-file hardening edits are verified by the -> per-session `/trekreview`**, which reads the real diffs. Manifest proves processing; -> review proves the edit. (A command may legitimately need zero edits if it passes all -> axes — so "command file modified" cannot be a manifest predicate.) -> -> **Prior-disposition rule (closes plan-critic #11):** where a step relies on a prior -> session's disposition (S15-B1, S17 F-VIDEO, overlap-measurement, the CWD fix), -> **spot-confirm it still holds with one grep/read** before relying on it — never "do -> not re-verify". +## Command classes & mechanical predicates (the testmethod per class) -### Step 1: Method calibration on `quick`, then harden the Start journey +- **post-emitting** — hook 110–140 present · length band present · no-body-link present · + buzzword check present · topic→5 pillars present · every `subagent_type: linkedin-studio:X` + resolves to `agents/X.md`. +- **routing** — every emitted `/linkedin:Y` resolves to `commands/Y.md`. +- **guided/stateful** — primary promised artifact actually produced · subagent targets resolve · + graceful degradation present. +- **analytics** — graceful degradation present · saves/dwell honesty intact. -- **Files:** `commands/quick.md`, `commands/onboarding.md`, `commands/first-post.md`, `commands/setup.md`, `docs/hardening/log.md` (new), `docs/hardening/field-notes.local.md` (read-if-present) -- **Changes:** First run the full 5-step method on `commands/quick.md` ALONE as a **method calibration** — produce the concrete before/after output, present the diff + the `log.md` entry shape to the operator, and lock the method before scaling. Then apply the method to the Start journey: `onboarding` (spot-confirm the S15-B1 inline-draft still delivers, then re-evaluate), `first-post`, `setup` (voice-trainer wiring + 5-pillar capture). Create `docs/hardening/log.md`; one anchored entry per command (`### /linkedin: — …`). Harden each `.md` only where an axis gaps. (`log.md` is a new file.) -- **Reuses:** the hardening method (`docs/hardening/brief.md`); `references/algorithm-signals-reference.md`; the mechanical content-quality rules (`hooks/prompts/content-quality-gate.md`); persona basis from the state file + voice-samples; the S17 record (`docs/remediation/c13-c46-triage.md`) as the log model. -- **Verify:** - - `bash scripts/test-runner.sh; echo "exit=$?"` → expected: output contains `Failed: 0` and `exit=0` (gate on `Failed: 0` + exit 0 + unchanged `EXPECT_*` counts, NOT the literal `Passed: 74` — it is incidental and `set -e`-fragile). - - `ls commands/*.md | wc -l` → `29`. - - `for c in quick onboarding first-post setup; do grep -qE "^### /linkedin:$c" docs/hardening/log.md || echo "MISSING $c"; done` → no output. -- **On failure:** **HALT** — S1 is the method-calibration gate that gates S2–S8 (hard dependency, not advisory). If calibration reveals the method itself is unsound, do NOT proceed to S2; revert command edits (`git checkout -- commands/quick.md commands/onboarding.md commands/first-post.md commands/setup.md`, keep `log.md`), and re-lock the method WITH THE OPERATOR before any later session runs. -- **Checkpoint:** `git commit -m "fix(linkedin-studio): S1 hardening — method calibration (quick) + Start journey"` -- **Manifest:** - ```yaml - manifest: - expected_paths: - - docs/hardening/log.md - min_file_count: 1 - commit_message_pattern: "^fix\\(linkedin-studio\\): S1 hardening" - bash_syntax_check: [] - forbidden_paths: - - docs/linkedin-studio-persona-brief.md - - docs/linkedin-studio-ui-brief.md - - docs/voyage-build/progress.json - must_contain: - - path: docs/hardening/log.md - pattern: "^### /linkedin:quick" - ``` +## Synthetic test data (fixtures) — for data-dependent commands -### Step 2: Harden Create I (atomic short-form) +Some commands cannot be meaningfully tested without input data — analytics (`import`, `report`, +`analyze`, `audit`, `ab-test`) need a CSV/JSON; stateful commands (`calendar`, `firsthour`) need a +queue/state. For these, **step 2 (agree test-method) includes agreeing a small fictitious dataset** +that exercises the command's real path: -> *quick is COMPLETE after S1 and is NOT re-hardened here (closes plan-critic #16).* -> *Session-grouping refinement (closes plan-critic #6): the brief's session table was -> explicitly "to be confirmed/refined by /trekplan". multiplatform moves to S2 (atomic -> cross-format short-form), batch to S3 (with the visual cluster), headless-review+pivot -> to S4 (the long-form cluster with newsletter). Coverage preserved — 29, none dropped/doubled.* +- **Isolated + throwaway** — built in a temp/scratch location (e.g. `/tmp/…` or a gitignored path), + **NEVER** the real analytics/state dirs, **NEVER** committed. +- **Minimal but realistic** — just enough rows/fields to trigger the behavior under test (e.g. a + 3–5-row CSV with the real column names; a 2-experiment ab-test file). +- **Clearly fictitious** — obviously-fake values, impossible to mistake for real data. +- **Discarded after the test** — the fixture proves the command works; it is not a deliverable. -- **Files:** `commands/post.md`, `commands/react.md`, `commands/multiplatform.md`, `docs/hardening/log.md` -- **Changes:** Apply the 5-step method to `post` (content-optimizer + differentiation-checker; hook/length mechanical pass), `react` (URL→post; de-AI gate), `multiplatform` (cross-format adaptation; confirm long-form correctly routes to `/linkedin:newsletter`). Append anchored entries; harden where an axis gaps. -- **Reuses:** hardening method; `algorithm-signals-reference.md`; content-quality gate; the `differentiation-checker` wiring pattern from S1. -- **Verify:** - - `bash scripts/test-runner.sh; echo "exit=$?"` → `Failed: 0` + `exit=0` + counts unchanged. - - `for c in post react multiplatform; do grep -qE "^### /linkedin:$c" docs/hardening/log.md || echo "MISSING $c"; done` → no output. -- **On failure:** revert — `git checkout -- commands/post.md commands/react.md commands/multiplatform.md` (keep `log.md`); if `/trekreview` ≠ ALLOW, HALT until resolved. -- **Checkpoint:** `git commit -m "fix(linkedin-studio): S2 hardening — Create I (post, react, multiplatform)"` -- **Manifest:** - ```yaml - manifest: - expected_paths: - - docs/hardening/log.md - min_file_count: 1 - commit_message_pattern: "^fix\\(linkedin-studio\\): S2 hardening" - bash_syntax_check: [] - forbidden_paths: - - docs/linkedin-studio-persona-brief.md - - docs/linkedin-studio-ui-brief.md - - docs/voyage-build/progress.json - must_contain: - - path: docs/hardening/log.md - pattern: "^### /linkedin:post" - ``` +## Session queue (ONE command each) -### Step 3: Harden Create II (visual + batch) +`S1 ✅` quick · onboarding · first-post · setup — **DONE under v1; do NOT re-harden.** -- **Files:** `commands/carousel.md`, `commands/video.md`, `commands/batch.md`, `docs/hardening/log.md` -- **Changes:** Apply the method to `carousel` (spot-confirm the S15-B3 full-deck clipboard still holds; differentiation-checker; mcp-image graceful degradation to text-only), `video` (video-scripter + differentiation-checker; spot-confirm the S17 F-VIDEO disposition (deliberate aspect-ratio decision) before relying on it; caption guidance; graceful degradation), `batch` (content-planner + trend-spotter; pillar-rotation across the week). Focus axis (d) graceful degradation when mcp-image/tools are unavailable. -- **Reuses:** hardening method; `references/linkedin-formats.md`, `references/video-strategy-guide.md`; the S17 F-VIDEO disposition (spot-confirmed, not assumed). -- **Verify:** - - `bash scripts/test-runner.sh; echo "exit=$?"` → `Failed: 0` + `exit=0` + counts unchanged. - - `for c in carousel video batch; do grep -qE "^### /linkedin:$c" docs/hardening/log.md || echo "MISSING $c"; done` → no output. -- **On failure:** revert — `git checkout -- commands/carousel.md commands/video.md commands/batch.md` (keep `log.md`); if `/trekreview` ≠ ALLOW, HALT. -- **Checkpoint:** `git commit -m "fix(linkedin-studio): S3 hardening — Create II (carousel, video, batch)"` -- **Manifest:** - ```yaml - manifest: - expected_paths: - - docs/hardening/log.md - min_file_count: 1 - commit_message_pattern: "^fix\\(linkedin-studio\\): S3 hardening" - bash_syntax_check: [] - forbidden_paths: - - docs/linkedin-studio-persona-brief.md - - docs/linkedin-studio-ui-brief.md - - docs/voyage-build/progress.json - must_contain: - - path: docs/hardening/log.md - pattern: "^### /linkedin:carousel" - ``` +| | | | | +|---|---|---|---| +| S2 post | S9 newsletter* | S16 analyze | S23 profile | +| S3 react | S10 headless-review | S17 audit | S24 create | +| S4 multiplatform | S11 pivot | S18 ab-test | S25 measure | +| S5 carousel | S12 firsthour | S19 strategy | S26 linkedin | +| S6 video | S13 calendar | S20 competitive | | +| S7 batch | S14 import | S21 monetize | | +| S8 pipeline | S15 report | S22 outreach | | -### Step 4: Harden Create III (long-form cluster) +*S9 newsletter (16-phase) may split into S9a/S9b. Otherwise one command = one session. -- **Files:** `commands/newsletter.md`, `commands/headless-review.md`, `commands/pivot.md`, `docs/hardening/log.md` -- **Changes:** Apply the method to the long-form cluster. `newsletter` = **orchestration only** (the 16-phase gate sequence + per-gate agents were heavily reviewed in remediation; simulate the operator's path THROUGH the phases — banner, phase transitions, lock/visual-assets/headless gates — NOT a re-review of each gate agent; spot-confirm `overlap-measurement.md` exists before citing "do not re-measure"). `headless-review` — wiring is **CONFIRMED present** (`headless-review.md:77-79,141-142`, `--type → subagent_type` table); the step VERIFIES that table resolves to real agents (mechanical axis-c) and simulates the consolidated-report output — no wiring fix expected. `pivot` (re-open-gates heuristic). **May split into its own session** if newsletter's orchestration simulation is large — decide at session start, log the split explicitly (not a silent cap). -- **Reuses:** hardening method; `references/newsletter-strategy-guide.md`; `references/longform-quality-rules.md`; `docs/remediation/overlap-measurement.md` (spot-confirmed). -- **Verify:** - - `bash scripts/test-runner.sh; echo "exit=$?"` → `Failed: 0` + `exit=0` + counts unchanged. - - `for c in newsletter headless-review pivot; do grep -qE "^### /linkedin:$c" docs/hardening/log.md || echo "MISSING $c"; done` → no output. - - `grep -qE "linkedin-studio:(content|language|fact)-reviewer" commands/headless-review.md` → present (wiring confirmed). -- **On failure:** revert — `git checkout -- commands/newsletter.md commands/headless-review.md commands/pivot.md` (keep `log.md`); if `/trekreview` ≠ ALLOW, HALT. -- **Checkpoint:** `git commit -m "fix(linkedin-studio): S4 hardening — Create III long-form (newsletter, headless-review, pivot)"` -- **Manifest:** - ```yaml - manifest: - expected_paths: - - docs/hardening/log.md - min_file_count: 1 - commit_message_pattern: "^fix\\(linkedin-studio\\): S4 hardening" - bash_syntax_check: [] - forbidden_paths: - - docs/linkedin-studio-persona-brief.md - - docs/linkedin-studio-ui-brief.md - - docs/voyage-build/progress.json - must_contain: - - path: docs/hardening/log.md - pattern: "^### /linkedin:newsletter" - ``` +## End-of-session ritual (every session — STATE.md handoff baked in) -### Step 5: Harden the Engage journey +1. `bash scripts/test-runner.sh` → `Failed: 0` + counts 29/19/25/6. +2. Terse, tool-grounded `log.md` line for the command (ONLY after verify): `PASS` / edit + `file:line`. +3. Commit own file(s) only, **LOCALLY** (`fix:` for command edits; explicit `git add `, never `-u`/`-A`). **Do NOT push hardening work** — see Push policy. +4. **Overwrite STATE.md:** mark this command ✅; point at the NEXT single command + its class; + carry the method + discipline rules → ready for the next "Les STATE.md og følg instruksjonene". +5. (Optional) update Voyage `.session-state.local.json` next-label for `/trekcontinue` coherence. -- **Files:** `commands/firsthour.md`, `commands/calendar.md`, `commands/pipeline.md`, `docs/hardening/log.md` -- **Changes:** Apply the method to `firsthour` (engagement-coach + post-feedback-monitor; `recordFirstHourPlan` state path), `calendar` (queue view + publish action; spot-confirm the auto-publish honesty boundary holds; state mutation), `pipeline` (content-planner + trend-spotter). Focus axis (c) guided/stateful predicate + axis (d) state-mutation + honesty boundaries. -- **Reuses:** hardening method; `hooks/scripts/state-updater.mjs` (`recordFirstHourPlan`, queue); the remediation scheduling-boundary wording (spot-confirmed). -- **Verify:** - - `bash scripts/test-runner.sh; echo "exit=$?"` → `Failed: 0` + `exit=0` + counts unchanged. - - `node --test hooks/scripts/__tests__/*.test.mjs` → `Failed: 0` (pass 98) if any hook touched. - - `for c in firsthour calendar pipeline; do grep -qE "^### /linkedin:$c" docs/hardening/log.md || echo "MISSING $c"; done` → no output. -- **On failure:** revert — `git checkout -- commands/firsthour.md commands/calendar.md commands/pipeline.md` (keep `log.md`); if `/trekreview` ≠ ALLOW, HALT. -- **Checkpoint:** `git commit -m "fix(linkedin-studio): S5 hardening — Engage journey (firsthour, calendar, pipeline)"` -- **Manifest:** - ```yaml - manifest: - expected_paths: - - docs/hardening/log.md - min_file_count: 1 - commit_message_pattern: "^fix\\(linkedin-studio\\): S5 hardening" - bash_syntax_check: [] - forbidden_paths: - - docs/linkedin-studio-persona-brief.md - - docs/linkedin-studio-ui-brief.md - - docs/voyage-build/progress.json - must_contain: - - path: docs/hardening/log.md - pattern: "^### /linkedin:firsthour" - ``` +## Defaults (adjustable between sessions) -### Step 6: Harden the Measure journey +1. **Log:** keep, terse, tool-grounded, after-verify only. +2. **Persona:** ICP author default; fresh-adopter for onboarding/first-post/setup; deviate explicitly elsewhere. +3. **Gate/commit:** once per session = per command. -- **Files:** `commands/import.md`, `commands/report.md`, `commands/analyze.md`, `commands/audit.md`, `commands/ab-test.md`, `docs/hardening/log.md` -- **Changes:** Apply the method to `import` (analytics CLI graceful degradation — spot-confirm the remediation CWD/`npm install` fix holds; the S16 saves manual-entry path), `report`/`analyze` (analytics-interpreter; saves + directional framing), `audit` (quarterly strategy audit; type-c = guided/stateful predicate), `ab-test` (directional-not-significant framing; content-optimizer). Focus axis (c) analytics graceful-degradation + (a) directional-A/B + saves honesty. -- **Reuses:** hardening method; `scripts/analytics/`; the remediation's saves/dwell honesty wording; the S16 `parseOptionalCount()` saves path (spot-confirmed). -- **Verify:** - - `bash scripts/test-runner.sh; echo "exit=$?"` → `Failed: 0` + `exit=0` + counts unchanged. - - if analytics touched: `cd scripts/analytics && npm test` → `Failed: 0` (pass 116). - - `for c in import report analyze audit ab-test; do grep -qE "^### /linkedin:$c" docs/hardening/log.md || echo "MISSING $c"; done` → no output. -- **On failure:** revert — `git checkout -- commands/import.md commands/report.md commands/analyze.md commands/audit.md commands/ab-test.md` (keep `log.md`); if `/trekreview` ≠ ALLOW, HALT. -- **Checkpoint:** `git commit -m "fix(linkedin-studio): S6 hardening — Measure journey (import, report, analyze, audit, ab-test)"` -- **Manifest:** - ```yaml - manifest: - expected_paths: - - docs/hardening/log.md - min_file_count: 1 - commit_message_pattern: "^fix\\(linkedin-studio\\): S6 hardening" - bash_syntax_check: [] - forbidden_paths: - - docs/linkedin-studio-persona-brief.md - - docs/linkedin-studio-ui-brief.md - - docs/voyage-build/progress.json - must_contain: - - path: docs/hardening/log.md - pattern: "^### /linkedin:ab-test" - ``` +## Push policy (operator rule, 2026-05-31) -### Step 7: Harden the Grow journey +Hardening work — *"det vi GJØR"*: command edits + `log.md` entries — is committed **locally +only, never pushed**. Only **method / process / tooling improvements** (this plan, a future +checker, STATE conventions) may be pushed, and only on **explicit operator OK** +(*"eventuelle forbedringer"*). -- **Files:** `commands/strategy.md`, `commands/competitive.md`, `commands/monetize.md`, `commands/outreach.md`, `commands/profile.md`, `docs/hardening/log.md` -- **Changes:** Apply the method to `strategy` (phase guidance/trajectory; strategy-advisor), `competitive` (niche analysis; type-c guided predicate; remediation un-gated it), `monetize` (~1K soft-gating; readiness scoring), `outreach` (collab + speaking; network-builder; tracked-pipeline via `recordOutreachContact`), `profile` (profile-SEO surface from remediation). Focus (a) intention for the ~1K-gated commands (soft-gate guidance, not dead-ends) + (c)/(d) tracked-pipeline state. -- **Reuses:** hardening method; the remediation's profile-SEO additions, outreach pipeline tracker (`state-updater.mjs`), ~1K soft-gating rule. -- **Verify:** - - `bash scripts/test-runner.sh; echo "exit=$?"` → `Failed: 0` + `exit=0` + counts unchanged. - - `for c in strategy competitive monetize outreach profile; do grep -qE "^### /linkedin:$c" docs/hardening/log.md || echo "MISSING $c"; done` → no output. -- **On failure:** revert — `git checkout -- commands/strategy.md commands/competitive.md commands/monetize.md commands/outreach.md commands/profile.md` (keep `log.md`); if `/trekreview` ≠ ALLOW, HALT. -- **Checkpoint:** `git commit -m "fix(linkedin-studio): S7 hardening — Grow journey (strategy, competitive, monetize, outreach, profile)"` -- **Manifest:** - ```yaml - manifest: - expected_paths: - - docs/hardening/log.md - min_file_count: 1 - commit_message_pattern: "^fix\\(linkedin-studio\\): S7 hardening" - bash_syntax_check: [] - forbidden_paths: - - docs/linkedin-studio-persona-brief.md - - docs/linkedin-studio-ui-brief.md - - docs/voyage-build/progress.json - must_contain: - - path: docs/hardening/log.md - pattern: "^### /linkedin:outreach" - ``` +## Amendment policy -### Step 8: Harden the front-doors + router, and close the phase - -- **Files:** `commands/create.md`, `commands/measure.md`, `commands/linkedin.md`, `docs/hardening/log.md` -- **Changes:** Apply the method to the routing surfaces: `create` (Create front-door — simulate "what should I make?" and confirm axis-c routing predicate: every target it emits resolves to a real, now-hardened creation command), `measure` (Measure front-door — routes to the correct analytics command), `linkedin` (the five-journey router — every journey + command listed and reachable, no stale entries). **Brand-consistency judgment (from plan-critic #5):** evaluate `commands/linkedin.md:4,16` ("LinkedIn thought leadership assistant/commands") — these are domain descriptors (S17 ruled them not-stale plugin-name misnomers), but the router is the brand front-door, so harden to "LinkedIn Studio" phrasing IF it improves intention clarity (a hardening judgment, not forced). **Phase close:** confirm `log.md` has all 29 anchored entries (SC-F); decide the optional v4.2.0 bump (if taken, it is the ONE feat-gated, three-doc, version-synced commit of the phase — run the full version grep — NOT a `fix:`). -- **Reuses:** hardening method; `docs/remediation/journey-layer-design.md` (the five-journey design); the hardened commands S1–S7 as routing targets. -- **Verify:** - - `bash scripts/test-runner.sh; echo "exit=$?"` → `Failed: 0` + `exit=0` + counts unchanged. - - **SC-F coverage (anchored, closes blockers #1/#2):** `for c in $(ls commands/*.md | xargs -n1 basename | sed 's/.md//'); do grep -qE "^### /linkedin:$c( |—|$)" docs/hardening/log.md || echo "MISSING: $c"; done` → no output (all 29 anchored entries present). - - **SC-G stale-brand (broadened, closes #5):** `grep -rn "thought leadership plugin" commands/ skills/ references/ README.md` → 0 (the plugin-NAME misnomer; domain phrases are allowed). -- **On failure:** revert — `git checkout -- commands/create.md commands/measure.md commands/linkedin.md` (keep `log.md`); escalate if SC-F coverage is incomplete. -- **Checkpoint:** `git commit -m "fix(linkedin-studio): S8 hardening — front-doors + router; phase complete"` -- **Manifest:** - ```yaml - manifest: - expected_paths: - - docs/hardening/log.md - min_file_count: 1 - commit_message_pattern: "^fix\\(linkedin-studio\\): S8 hardening" - bash_syntax_check: [] - forbidden_paths: - - docs/linkedin-studio-persona-brief.md - - docs/linkedin-studio-ui-brief.md - - docs/voyage-build/progress.json - must_contain: - - path: docs/hardening/log.md - pattern: "^### /linkedin:linkedin" - ``` - -## Alternatives Considered - -| Approach | Pros | Cons | Why rejected | -|----------|------|------|--------------| -| One Voyage session per command (29 sessions) | Deepest per command; granular checkpoints | ~29 sessions; high startup overhead; uneven | Operator locked "per journey, large groups split" — 8 sessions | -| Spec-audit only (no simulation) | Faster, cheaper | Misses output-quality gaps — the phase's whole reason | Operator locked "hybrid: run + spec-audit" | -| Allow structural redesign (merge/split commands) | Could fix root-cause structure | Churn right after remediation; rocks the count/journey invariants; re-opens 14a | Operator locked "intention-fidelity + prompt-quality, no structural" | -| Full exploration swarm (6–8 agents) in planning | Canonical /trekplan rigor | Re-maps an already-fully-mapped, frozen codebase — wasted spend | Scaled down per the "adaptive / don't over-spawn" rule; kept the high-value brief-review + plan-critic + scope-guardian gates | -| Oracle = "≥1 command/session spot-check" (original draft) | Cheaper review | Leaves 80%+ of a session self-graded — does not break the loop | Replaced (plan-critic #3) by "reviewer adjudicates EVERY hardened command's before/after" | - -## Test Strategy - -- **Framework:** no unit tests are written *by* this phase (it edits Markdown command - prompts). The "test" per command is the **simulation + per-type mechanical predicate + - the concrete before/after output pair**, recorded in `docs/hardening/log.md` and - **adjudicated by the cold `/trekreview` reviewer** (the independent oracle). The - existing suites (`test-runner.sh`, hooks `node --test`, analytics `npm test`) are - **regression guards**, gated on `Failed: 0` (not a literal pass-count). -- **Existing patterns:** evidence-per-row disposition records (S17 `c13-c46-triage.md`). -- **New tests in this plan:** 0 new automated tests; the concrete before/after pair + - per-type mechanical predicate is the falsifiable evidence, adjudicated independently. - -### Tests to write - -| Type | File | Verifies | Model | -|------|------|----------|-------| -| Evidence record | `docs/hardening/log.md` | per-command INTENT→VERIFY + concrete before/after + mechanical predicate; unique anchor per entry | `docs/remediation/c13-c46-triage.md` | -| Regression | (existing) `scripts/test-runner.sh` | counts 29/19/25/6 + structural guards; gate on `Failed: 0` + exit 0 | — | -| Regression | (existing) hooks/analytics `node --test` | `Failed: 0` (98 / 116) when those surfaces touched | — | - -## Risks and Mitigations - -| Priority | Risk | Location | Impact | Mitigation | -|----------|------|----------|--------|------------| -| High | **Self-graded quality axis** (brief-review #1, plan-critic #3/#12) | method 3+5 | "hardened" claimed without provable improvement | Concrete (reproducible) before/after output + per-type mechanical predicate + `/trekreview` reviewer adjudicates **every** hardened command's before/after (not a sample) | -| High | **Coverage gate false-positive** (plan-critic #1/#2) | SC-F + per-step greps | a dropped command reads as covered | Unique anchored log header `^### /linkedin:`; all greps anchored, never bare-word | -| Medium | **Gold-plating / unbounded depth** (brief-review #3) | method 4 | token blowup; uneven coverage | Stopping rule: axes pass-or-defer; no NICE polish; per-session split logged if scope balloons | -| Medium | **SC-H vs deferral contradiction** (brief-review #2) | brief SC-H | gate unsatisfiable, or "open finding" redefined | SC-H = "0 **un-triaged** findings; each fixed or recorded-deferral; `/trekreview` ALLOW with deferrals noted" | -| Medium | **`Passed: 74` brittle under `set -e`** (plan-critic #7, scope-guardian caveat) | every Verify | spurious gate failure on any added check | Gate on `Failed: 0` + exit 0 + unchanged `EXPECT_*` counts; `74` is incidental | -| Medium | **No circuit-breaker on the S1 method gate** (plan-critic #9) | S1 On-failure | headless run proceeds with an unsound method | S1 On-failure = HALT; S1→S2..S8 is a hard dependency; any non-ALLOW `/trekreview` halts the phase | -| Medium | **Mechanical axis absent on ~13 no-output commands** (plan-critic #4) | method 3(c) | the non-circular axis evaporates → pure judgment | Per-type mechanical predicate defined for every command class (routing/analytics/guided), never "N/A → judgment" | -| Low | Simulation ≠ live run | method 2 | a gap only real testing surfaces is missed | Field-notes inbox: operator's live findings outrank simulated; checked before each journey; SC-I | -| Low | Prior-disposition premise misremembered (plan-critic #11) | S1/S3/S4/S6 | an axis skipped on a false premise | Spot-confirm each cited disposition with one grep/read before relying | -| Low | Stale-brand grep mis-scoped (plan-critic #5) | SC-G | a plugin-name misnomer in commands/ slips | SC-G greps commands/+skills/+references/+README; S8 also judges the router's domain phrasing | -| Low | Touching a not-mine untracked file | repo `docs/` | accidental commit of operator/UI briefs | `forbidden_paths` in every manifest + explicit staging (own files only) | -| Low | Optional v4.2.0 bump fumbled as `fix:` (brief-review #4) | S8 phase-close | version-skew | If taken, the bump is the ONE feat-gated three-doc version-synced commit; run the full version grep | - -## Assumptions - -| # | Assumption | Why unverifiable | Impact if wrong | -|---|-----------|-----------------|-----------------| -| 1 | `quick` is the right S1 calibration command | operator preference, not yet confirmed | trivial — swap at S1 start | -| 2 | SC-H means "0 un-triaged findings", deferrals allowed | reconciling brief SC-H with SC-D | a legitimately-deferred finding would otherwise fail the gate | -| 3 | `newsletter` orchestration fits S4 with headless-review + pivot | depends on simulation size | S4 splits into its own session (logged) — permitted | -| 4 | ~~headless-review wiring may be broken~~ **RESOLVED** — wiring confirmed present (`headless-review.md:77-79,141-142`) | n/a | n/a — S4 verifies the table resolves, no fix expected | -| 5 | Version stays v4.1.0 through the phase; optional v4.2.0 at phase end | phase-end decision | none — pre-committed as optional | - -## Verification - -*Per-step manifest verification runs automatically during execution. End-to-end / -phase-level checks:* - -- [ ] **SC-F (coverage, anchored):** `for c in $(ls commands/*.md | xargs -n1 basename | sed 's/.md//'); do grep -qE "^### /linkedin:$c( |—|$)" docs/hardening/log.md || echo "MISSING $c"; done` → no output (all 29 anchored entries; no substring false-positives). -- [ ] **SC-A…SC-E (per command):** each anchored `log.md` entry has intention + simulation (persona + CONCRETE before-output) + 4-axis evaluation + (if hardened) concrete after-output + verify line. -- [ ] **SC-C (mechanical, hard, per type):** every command records a pass/fail on its type's mechanical predicate (post-emitting → 4 content rules; routing → targets resolve; analytics → graceful-degradation present; guided → primary artifact produced) — never "N/A → judgment". -- [ ] **SC-G (no structural drift):** `ls commands/*.md | wc -l` → `29` every session; `bash scripts/test-runner.sh` → `Failed: 0` + exit 0 + `EXPECT_*` counts unchanged; `grep -rn "thought leadership plugin" commands/ skills/ references/ README.md` → 0. -- [ ] **SC-H (clean gate, reconciled):** each session's `review.md` shows **0 un-triaged findings** (each fixed in-session or carrying a recorded deferral rationale) and `/trekreview` returns **ALLOW** (deferrals noted, not WARN-override). -- [ ] **Oracle (independent):** each session's `/trekreview` adjudicated the before/after of **every** command hardened that session (recorded in `review.md`) — not a sample. -- [ ] **SC-I (field-notes):** where `docs/hardening/field-notes.local.md` has a finding for a session's commands, the log addresses or triages it; absent file → graceful no-op (does not block; the inbox is operator-owned). -- [ ] **Regression:** `node --test hooks/scripts/__tests__/*.test.mjs` → `Failed: 0` and `cd scripts/analytics && npm test` → `Failed: 0` whenever those surfaces are touched. - -## Estimated Scope - -- **Files to modify:** up to 29 command `.md` files (only where an axis gaps — some pass - clean with no edit) + occasional agent/reference edits where intention requires. -- **Files to create:** 1 (`docs/hardening/log.md`); `field-notes.local.md` is - operator-created (gitignored). -- **Complexity:** medium — low risk per edit (surgical prompt changes, no code/structure), - broad (29 surfaces) and judgment-heavy, de-risked by the per-type mechanical predicate + - the full-session cold-reviewer oracle. - -## Execution Strategy - -*8 sessions, sequential (each depends on the method locked in S1 and shares the single -`log.md` append target → not parallelizable without log-merge conflicts). Run one per -Voyage session via `/trekcontinue`. **Session-grouping note (plan-critic #6):** the -brief's session table was explicitly "to be confirmed/refined by /trekplan"; this plan -refines it — multiplatform→S2, batch→S3, headless-review+pivot→S4 — with full coverage -preserved (29, none dropped/doubled). `quick` is complete after S1.* - -### Session 1: Method calibration + Start -- **Steps:** 1 · **Wave:** 1 · **Depends on:** none -- **Scope fence:** Touch: `commands/{quick,onboarding,first-post,setup}.md`, `docs/hardening/log.md`. Never touch: other commands, the 3 not-mine untracked files. -- **Circuit-breaker:** if S1 calibration fails, HALT — no later session runs until the method is re-locked with the operator. - -### Session 2: Create I (atomic short-form) -- **Steps:** 2 · **Wave:** 2 · **Depends on:** Session 1 (hard gate — method locked) -- **Scope fence:** Touch: `commands/{post,react,multiplatform}.md`, `log.md`. Never touch: other commands. - -### Session 3: Create II (visual + batch) -- **Steps:** 3 · **Wave:** 3 · **Depends on:** Session 1 -- **Scope fence:** Touch: `commands/{carousel,video,batch}.md`, `log.md`. Never touch: other commands. - -### Session 4: Create III (long-form cluster) -- **Steps:** 4 · **Wave:** 4 · **Depends on:** Session 1 -- **Scope fence:** Touch: `commands/{newsletter,headless-review,pivot}.md`, `log.md`. Never touch: other commands. (May split 4a/4b if newsletter orchestration is large — logged.) - -### Session 5: Engage -- **Steps:** 5 · **Wave:** 5 · **Depends on:** Session 1 -- **Scope fence:** Touch: `commands/{firsthour,calendar,pipeline}.md`, `log.md`. Never touch: other commands. - -### Session 6: Measure -- **Steps:** 6 · **Wave:** 6 · **Depends on:** Session 1 -- **Scope fence:** Touch: `commands/{import,report,analyze,audit,ab-test}.md`, `log.md`. Never touch: other commands. - -### Session 7: Grow -- **Steps:** 7 · **Wave:** 7 · **Depends on:** Session 1 -- **Scope fence:** Touch: `commands/{strategy,competitive,monetize,outreach,profile}.md`, `log.md`. Never touch: other commands. - -### Session 8: Front-doors + router (phase close) -- **Steps:** 8 · **Wave:** 8 · **Depends on:** Sessions 1–7 (routing targets must be hardened first) -- **Scope fence:** Touch: `commands/{create,measure,linkedin}.md`, `log.md`. Never touch: other commands. - -### Execution Order - -- **Wave 1:** Session 1 (calibration — hard gate; HALT-on-failure) -- **Waves 2–7:** Sessions 2–7 (each depends only on S1; sequential because they share `log.md`) -- **Wave 8:** Session 8 (after all targets hardened) - -### Grouping rules applied - -- Unit = one command; grouped by journey per the locked cadence. -- Create (8 cmd) split across S2–S4; small journeys = one session each. -- Sessions share the single `log.md` append target → sequential, not parallel. -- Session 8 (routers) last — depends on its targets being hardened. - -## Plan Quality Score - -*Recomputed post-revision (the pre-review self-score was struck per plan-critic #17).* - -| Dimension | Weight | Score | Notes | -|-----------|--------|-------|-------| -| Structural integrity | 0.15 | 90 | 8 steps dependency-ordered; S1 now a HALT-enforced gate; brief-vs-plan reassignment justified | -| Step quality | 0.20 | 85 | per-command 5-step method; VERIFY now backed by an independent oracle, not self-re-sim; headless-review resolved | -| Coverage completeness | 0.20 | 90 | anchored coverage predicate (no substring false-positives); all 29 mapped, none dropped/doubled | -| Specification quality | 0.15 | 86 | concrete files/verify; gate on `Failed: 0` not literal 74; stale-brand grep broadened; prior dispositions spot-confirmed | -| Risk & pre-mortem | 0.15 | 88 | brief-review + plan-critic blockers/majors folded with mitigations | -| Headless readiness | 0.10 | 85 | On-failure + Checkpoint per step; S1 HALT circuit-breaker; non-ALLOW halts the phase | -| Manifest quality | 0.05 | 78 | manifests anchor on the unique-anchored log entry; edit-correctness explicitly delegated to `/trekreview` (honest separation) | -| **Weighted total** | **1.00** | **87** | **Grade: B+** | - -**Adversarial review:** -- **Plan critic:** REVISE — 2 blockers + 10 major + 5 minor (Grade C, 66). All blockers + majors addressed in this revision; see Revisions. (1 claim, #8 headless-review, found RESOLVED on verification; #5 found overstated, scope broadened anyway.) -- **Scope guardian:** ALIGNED — 0 creep, 0 material gaps; the `Failed: 0`-not-`74` caveat is adopted. - -## Revisions - -*Added by adversarial review (Phase 9) — plan-critic REVISE + scope-guardian caveat.* - -| # | Finding | Severity | Resolution | -|---|---------|----------|------------| -| 1 | SC-F coverage grep matches command names as substrings (dropped command reads as covered) | blocker | Unique anchored log header `### /linkedin: —`; all coverage/per-step greps anchored `^### /linkedin:$c` | -| 2 | `measure`/`create` collide with the Measure/Create journey labels | blocker | Same anchored-header fix; `:measure` ≠ the word "Measure" | -| 3 | Oracle samples ≥1/session of author-written sketches (80%+ self-graded) | major | Oracle now adjudicates **every** hardened command's before/after per session; simulation captures CONCRETE reproducible output, not a sketch | -| 4 | Mechanical axis inapplicable to ~13 no-output commands → pure judgment | major | Per-type mechanical predicate defined for every command class (routing/analytics/guided); never "N/A → judgment" | -| 5 | Stale-brand grep scoped to skills/ only | major→adjusted | Verified: `"thought leadership plugin"` is absent everywhere; grep broadened to commands/+skills/+references/+README; S8 also judges the router's domain phrasing | -| 6 | Brief-vs-plan session reassignment unjustified | major | Explicit refinement note (brief table was "to be confirmed by /trekplan"); coverage preserved | -| 7 | `Passed: 74` brittle under `set -e`; 74 hardcoded | major | All Verify/gate predicates → `Failed: 0` + exit 0 + unchanged `EXPECT_*` counts | -| 8 | headless-review wiring assumed-broken-then-deferred | major→resolved | Verified wiring is PRESENT (`--type → subagent_type` table, lines 77-79/141-142); S4 verifies, no fix expected; Assumption #4 struck | -| 9 | No circuit-breaker; S1 failure doesn't halt S2–S8 | major | S1 On-failure = HALT; S1→S2..S8 hard dependency; any non-ALLOW `/trekreview` halts the phase | -| 10 | Manifests anchor on log.md; hardened command file invisible | major | Documented honest separation: manifest proves processing (unique anchor); `/trekreview` proves the edit (reads real diffs); a clean-pass command may have 0 edits | -| 11 | Prior dispositions cited but unverified ("do not re-verify") | major | Spot-confirm each cited disposition with one grep/read before relying | -| 12 | SC-E re-simulation is circular | major | Same as #3 — independent oracle + concrete captured output | -| 13 | node --test counts hardcoded/conditional | minor | Framed as `Failed: 0` (counts incidental) | -| 16 | `quick` double-count (S1 + brief's S2 spillover) | minor | Explicit "quick complete after S1; not re-hardened in S2" note | -| 17 | Plan pre-assigned Grade A before review | minor | Score recomputed post-revision (B+ 87); header note added | +Improve the method/plan **between sessions** if a session surfaces a gap (operator-approved). +This plan is a living contract, not frozen.