Recipe: Lesson Cleanup

Generic, domain-invariant recipe for turning a single lesson-learned into a deterministic plan that fixes exactly what the lesson directs and nothing else. Mirrors recipe-refactor-to-profile-standards in shape: phase-3-outline loads this skill and walks its workflow once, emitting one deliverable per lesson directive. The recipe forces surgical scope so the manifest composer's surgical+tech_debt / surgical+bug_fix / surgical+enhancement cascade rules collapse Phase 5 verification and Phase 6 finalize down to the minimum safe set.

No separate analysis step — the lesson body itself is the deliverable source. Each directive in the lesson maps directly to one deliverable; the change_type is derived deterministically from the lesson kind. Q-Gate is skipped because the recipe path is its own gate.

Foundational Practices

Skill: plan-marshall:persona-plan-marshall-agent

Enforcement

Execution mode: Walk the four phase-aligned steps below in order. Each step has a single explicit job — no improvisation, no extra discovery passes.

Prohibited actions:

• Never invent extra deliverables beyond the lesson directives. The lesson body is the source of truth — if it lists three directives, the outline has three deliverables, full stop.
• Never set scope_estimate to anything other than surgical. This recipe exists specifically to drive the surgical cascade rules in the manifest composer.
• Never call Q-Gate validation from within this recipe. The recipe path bypasses Q-Gate by design (see phase-3-outline § Recipe Path).
• Never override the change_type derived from lesson kind. The mapping (bug → bug_fix, improvement → enhancement, anti-pattern → tech_debt) is fixed.

Constraints:

• Strictly comply with all rules from persona-plan-marshall-agent, especially tool usage and workflow step discipline.
• One deliverable per lesson directive — never merge or split.

Input

Parameter	Type	Required	Description
plan_id	string	Yes	Plan identifier
lesson_id	string	Yes	Lesson identifier (e.g., 2025-12-02-001)

Note: Unlike recipe-refactor-to-profile-standards, this recipe takes no recipe_domain, recipe_profile, or recipe_package_source parameters. The lesson body — already moved into the plan directory by phase-1-init Step 5b — is the sole input.

---

Step 1: Resolve Lesson Body and Kind (Phase 2 surrogate)

The recipe replaces the iterative phase-2-refine loop. Confidence is forced to 100 because the lesson is the contract.

Read the lesson document copied into the plan directory:

python3 .plan/execute-script.py plan-marshall:manage-files:manage-files read \
  --plan-id {plan_id} --file lesson-{lesson_id}.md

Parse the lesson frontmatter to extract:

• kind (one of bug, improvement, anti-pattern)
• title
• directives[] — the actionable items declared in the lesson body (one per ## Directive heading, or one per bullet under ## Actions, depending on lesson layout)

Step 1b: Premise Verification

Runs after the lesson body has been read and BEFORE Step 2 maps the lesson kind to a change_type. Without this gate, the recipe will faithfully translate stale directives into a deliverable, then phase-4-plan will faithfully translate them into tasks, and the executor will fight the live tree to apply prescriptions that no longer match it. The gate exists to catch the staleness once, at the cheapest possible point, before scope is locked.

Authoritative heuristics: This sub-step deliberately does NOT redefine the extraction or verification heuristics. They live in marketplace/bundles/plan-marshall/skills/phase-1-init/standards/lesson-source-premise-check.md, which is the canonical source for file-path / function-name / CLI-shape / anti-pattern matching, the verification-helper table, and the per-kind "stale when" rules. Read that document and apply its passes verbatim to the lesson body resolved in Step 1. Re-implementing the heuristics here would create a second source of truth that drifts out of sync with phase-1-init.

Sub-step a — Enumerate concrete factual assertions. Apply the extraction passes from the standard to the lesson body, producing a working set of (reference, kind) tuples covering every concrete factual assertion: file paths, function/method/subcommand names, CLI invocation shapes (python3 .plan/execute-script.py {notation} {subcommand}), and anti-pattern signatures. Each tuple is tied to one or more directives in the lesson — record the back-link so dropping the assertion can also drop the corresponding directive. Prose claims that do not resolve to a checkable reference are out of scope for the gate; they remain in the directive set unchanged.

Sub-step b — Quick-check each assertion against the live tree. Apply the verification-helper table from marketplace/bundles/plan-marshall/skills/phase-1-init/standards/lesson-source-premise-check.md (the canonical source) — do not restate it here.

Record each verification as (reference, kind, status, evidence) where status is valid or stale. Do not abort on the first stale match — the gate needs the full picture to decide whether to drop selectively or abort wholesale.

Sub-step c — Drop stale directives from scope before Step 3. For every assertion marked stale, follow the recorded back-link and drop the directive(s) it underpinned from the working set. The dropped directives MUST NOT appear as deliverables in Step 3; the residual (valid) directive set is the input phase-3-outline composes from. A directive that depends on multiple assertions is dropped if ANY of its assertions is stale — the recipe never tries to "rescue" a directive by ignoring the broken half.

Sub-step d — Record each dropped scope item via the decision log. For every dropped directive, emit one decision-log entry with the prefix (plan-marshall:recipe-lesson-cleanup:premise-check) so reviewers can audit the scope reduction:

python3 .plan/execute-script.py plan-marshall:manage-logging:manage-logging \
  decision --plan-id {plan_id} --level INFO \
  --message "(plan-marshall:recipe-lesson-cleanup:premise-check) Dropped directive '{directive_title}' — stale assertion: {reference} ({evidence})"

When a directive is underpinned by multiple stale assertions, list them comma-separated in the {evidence} segment. Kept directives are NOT logged individually — only drops, since drops shrink scope and require audit visibility.

Abort condition — lesson fully obsolete. When ALL directives in the lesson have been dropped (the residual set is empty), the recipe MUST abort with an explicit error rather than silently emit an empty outline. Return:

status: error
error: lesson_fully_obsolete
plan_id: {plan_id}
lesson_id: {lesson_id}
message: "Every directive in lesson '{lesson_id}' was dropped during premise verification — the lesson no longer matches the live tree."
recovery: "Close the lesson via 'python3 .plan/execute-script.py plan-marshall:manage-lessons:manage-lessons remove --lesson-id {lesson_id} --reason lesson-obsolete --coverage-verdict obsolete', or refine it manually before re-running the recipe."

Do not call manage-solution-outline write, do not transition phase, do not emit a no-op outline. The recipe is the gate — an empty residual scope means the gate fired.

---

Step 2: Map Lesson Kind to change_type

Apply the fixed mapping:

Lesson kind	change_type
bug	bug_fix
improvement	enhancement
anti-pattern	tech_debt

If the lesson kind is anything else, abort with an explicit error — do not guess. The mapping is the contract that drives the manifest composer's cascade selection downstream.

status: error
error: unknown_lesson_kind
message: "Lesson kind '{kind}' is not mapped — supported: bug, improvement, anti-pattern"
recovery: "Tag the lesson with one of the supported kinds, then re-run."

---

Step 2b: Verify Proposed Fix Against Current Code (phase-2-refine Step 3c surrogate)

Runs after the lesson kind is mapped (Step 2) and BEFORE Step 3 composes the deterministic outline. This step closes a structural gap: lesson-sourced plans set source: lesson in request.md, which makes phase-2-refine take its Recipe Shortcut (skipping the iterative refine loop and forcing confidence to 100). That shortcut bypasses phase-2-refine Step 3c — Proposed-Fix Verification, so the lesson's ## Recommended Fix (or equivalent "Recommended Follow-Up" / "Solution / Action" / "Application" section) would otherwise flow straight into Step 3's deliverable composition without ever being challenged against the current code path.

This step is the recipe's surrogate for that bypassed verification. It runs before any request.md Clarifications block update and before any solution_outline.md content is composed, so divergences are recorded into the plan inputs rather than retro-fitted afterwards.

Authoritative patterns: This sub-step deliberately does NOT redefine the proposed-fix extraction, probe construction, or result-handling heuristics. The reusable patterns live in marketplace/bundles/plan-marshall/skills/phase-2-refine/standards/proposed-fix-verification.md, which is the canonical source for the trigger condition, extraction fields, probe budget, and Valid/Insufficient/Inconclusive result table. Read that document and apply its passes against the current lesson body.

The recipe layer adds one constraint on top of the canonical patterns: the lesson's ## Recommended Fix (or equivalent) is a starting hypothesis only — never authoritative implementation guidance. The recipe re-derives the minimal correct change from disk; it does not transcribe the recommendation.

Sub-step a — Treat the lesson's recommended fix as a hypothesis. Locate the lesson's "what to do" section. Common headings include ## Recommended Fix, ## Recommended Follow-Up, ## Solution / Action, and ## Application. Capture the recommendation verbatim as hypothesis_fix for each directive that survived Step 1b. A lesson with no such section skips this sub-step entirely — there is no hypothesis to challenge, and the directives stand as written.

Sub-step b — Re-derive the minimal correct change from the live tree. For each surviving directive, read the surrounding code path on disk using the architecture inventory first — architecture files --module X to enumerate the module's components, architecture which-module --path P and architecture find --pattern P for module-spanning lookups — and fall back to Glob, Grep, and Read only for sub-module component lookup, content searches inside an already-known file, or when the architecture verb returns elision:

• Resolve every component the directive touches (skill notations, file paths, function names, scripts) in the live worktree.
• Trace the actual current behavior — what the code does today, what it calls, what its callers expect.
• Derive the minimal correct change that addresses the directive's symptom against this current state. The minimal change is the smallest edit that makes the symptom stop manifesting; it is NOT the largest edit that the recommendation suggests.

Sub-step c — Compare hypothesis against verified-correct minimal change and note divergences. For each directive with a hypothesis_fix, compare it against the minimal correct change derived in sub-step b. Classify each comparison into one of four divergence categories:

Category	Meaning
convergent	Hypothesis matches the minimal correct change — no action needed
over-broad	Hypothesis prescribes more than the minimal change requires (e.g., touches unrelated files, adds defensive code beyond the symptom)
redundant	Part or all of the hypothesis is already satisfied by code that landed elsewhere after the lesson was authored
mis-targeted	Hypothesis addresses a contract / API / pattern that no longer exists, or addresses the wrong layer

Record each non-convergent comparison as (directive, category, hypothesis_summary, verified_change_summary, evidence) for use in sub-step d.

Sub-step d — When verified-correct fix diverges, update the plan inputs. For each directive in over-broad, redundant, or mis-targeted:

1. Update the plan's request.md Clarifications block with the verified-correct fix and the rationale for divergence. The Clarifications block is the canonical place phase-3-outline reads for refined intent — by writing here, the recipe ensures Step 3's deliverable composition sees the verified version, not the hypothesis.
2. Record the divergence so it surfaces in the eventual commit message or PR body. The dispatcher's per-deliverable commit instrumentation and phase-6-finalize:create-pr consume the decision log and the request.md Clarifications block when composing their message bodies — the entry below feeds both.

Resolve the canonical request.md path, then update ## Clarifications and ## Clarified Request directly via Edit/Write — the same three-step path-allocate flow phase-2-refine uses:

python3 .plan/execute-script.py plan-marshall:manage-plan-documents:manage-plan-documents \
  request path --plan-id {plan_id}

Use Edit (or Write, when no file exists yet) against the resolved path to append a Clarifications entry per divergent directive in this shape:

- Directive: {directive_title}
  Hypothesis: {hypothesis_summary}
  Verified fix: {verified_change_summary}
  Rationale: {category} — {evidence}

When all divergences are written, mark the request clarified so phase-3-outline picks up the refined intent:

python3 .plan/execute-script.py plan-marshall:manage-plan-documents:manage-plan-documents \
  request mark-clarified --plan-id {plan_id}

Then emit a decision-log entry per divergence so reviewers can audit the scope adjustment:

python3 .plan/execute-script.py plan-marshall:manage-logging:manage-logging \
  decision --plan-id {plan_id} --level INFO \
  --message "(plan-marshall:recipe-lesson-cleanup:fix-verification) Diverged from hypothesis for '{directive_title}' — {category}: hypothesis was '{hypothesis_summary}', verified change is '{verified_change_summary}'"

convergent directives are NOT logged individually — only divergences, since divergences alter scope and require audit visibility.

Sub-step e — Log the verification summary. Emit one work.log entry summarizing the pass:

python3 .plan/execute-script.py plan-marshall:manage-logging:manage-logging \
  work --plan-id {plan_id} --level INFO \
  --message "[RECIPE:2b] (plan-marshall:recipe-lesson-cleanup) Proposed-fix verification: {N} directives probed, {C} convergent, {D} divergent ({over_broad} over-broad, {redundant} redundant, {mis_targeted} mis-targeted)"

Hand-off to Step 3. Step 3 composes deliverables from the verified directive set — i.e., the residual directives from Step 1b with each over-broad / redundant / mis-targeted directive replaced by its verified-correct minimal change. The hypothesis text is not used directly. Step 3 reads the updated Clarifications block and composes deliverables from there.

---

Step 2c: Retrospective-vs-Prospective Classifier

Runs after Step 2b (Proposed-Fix Verification) and BEFORE Step 3 composes the outline. This step reuses Step 2b's per-directive divergence verdicts — it does NOT re-probe the live tree or re-run Step 2b's redundancy validation. Its job is narrow: decide, per surviving directive, whether the lesson is prospective (the corrective change still needs to be made) or retrospective (the corrective change already shipped, but the lesson carries generalizable wisdom worth porting into the component's docs).

Shared classification semantics: the retrospective-vs-prospective signals, the apply-the-wisdom / documentation-port deliverable template, and the required lesson cross-reference are defined once in `../phase-3-outline/standards/outline-workflow-detail.md` § Retrospective-vs-prospective lesson classification. This recipe and phase-3-outline are the two callers of that shared definition; do NOT restate the signals or the template here. Read the standard and apply it to each surviving directive.

Composition with Step 2b (reuse, do not re-run): Step 2b already classified each directive as convergent / over-broad / redundant / mis-targeted. The retrospective signal "fix already shipped" is exactly what Step 2b's redundant category detects — a directive whose fix landed elsewhere after the lesson was authored. This step consumes that verdict directly:

• A directive Step 2b marked redundant AND whose lesson body carries a forward-looking section (## Generalisation / ## Generalization / a "the general rule is…" passage) AND whose target component source is still editable → classify the directive retrospective. Instead of dropping it as pure redundancy, route it to the documentation-port shape: emit a documentation_only deliverable in Step 3 that lifts the forward-looking rule into the target component's canonical docs (its SKILL.md or a standards/*.md) and cross-references the originating lesson {lesson_id}.
• Surviving directives Step 2b marked convergent / over-broad / mis-targeted remain prospective → standard Step 3 composition (a code-change deliverable from the verified-correct minimal change).
• A directive Step 2b marked redundant with NO forward-looking section is obsolete — its fix already shipped and there is no generalizable wisdom to port. It is NOT composed into a deliverable (neither documentation-port nor a prospective code-change deliverable, since re-applying an already-shipped fix is a no-op). Record it as a dropped reference in the work-log so downstream phases can audit the scope reduction, and exclude it from Step 3. Do NOT mutate the original lesson body in request.md — preserve it verbatim as the source of truth.

Do NOT re-derive the "fix already shipped" state from disk — that probe already ran in Step 2b. This step only reads Step 2b's recorded verdicts and the lesson body's section structure.

Hand-off to Step 3: each retrospective directive carries a shape: documentation-port marker into Step 3; each prospective directive carries the standard shape: code-change marker; obsolete directives are dropped before Step 3 and never reach the composer. Step 3c branches on the marker (see Step 3c below). Log the classification outcome:

python3 .plan/execute-script.py plan-marshall:manage-logging:manage-logging \
  work --plan-id {plan_id} --level INFO \
  --message "[RECIPE:2c] (plan-marshall:recipe-lesson-cleanup) Retrospective classifier: {R} retrospective (documentation-port), {P} prospective (code-change), {O} obsolete (dropped)"

---

Step 3: Compose Deterministic Outline (Phase 3 surrogate)

For each lesson directive, emit one deliverable. The outline is purely structural — no LLM reasoning, no Q-Gate, no decomposition pass.

3a. Read the deliverable template:

Read: marketplace/bundles/plan-marshall/skills/manage-solution-outline/templates/deliverable-template.md

3b. Resolve the target outline path:

python3 .plan/execute-script.py plan-marshall:manage-solution-outline:manage-solution-outline \
  resolve-path --plan-id {plan_id}

3c. Compose the document. For every directive handed off by Step 2c, branch on its shape marker:

Branch shape: documentation-port (retrospective directives) — do NOT use the code-change shape below. Compose the deliverable using the apply-the-wisdom / documentation-port deliverable template defined in `../phase-3-outline/standards/outline-workflow-detail.md` § Retrospective-vs-prospective lesson classification. That template fixes the deliverable's change_type (enhancement), profile (documentation_only → implementation only), **Affected files:** (the target component's canonical docs, NOT the already-shipped code paths), and the required cross-reference to the originating lesson {lesson_id}. The directive's change_type derived in Step 2 does NOT apply to a documentation-port deliverable — the template's enhancement value supersedes it, because the code fix already shipped and only the doc-port remains.

Branch shape: code-change (prospective directives) — compose the standard code-change deliverable:

• Title: Lesson cleanup: {directive_title}
• Description: The directive body verbatim from the lesson
• Metadata:
  • change_type: derived in Step 2
  • execution_mode: automated
  • domain: from the lesson's component field (fall back to plan-marshall-plugin-dev for plugin-bundle lessons)
  • scope_estimate: surgical — REQUIRED. This drives the surgical+{change_type} cascade rule in the manifest composer.
• Skills: derived from the lesson's related field (skill notations) — passed verbatim to the composed deliverable.
• Affected files: from the lesson's affected_files field if present; otherwise empty (the executor task will discover them from the directive body).

Obsolete directives (Step 2c dropped them as redundant-with-no-forward-looking) never reach Step 3c — they are not composed.

The top-level outline metadata MUST also include scope_estimate: surgical so phase-3-outline writes the value into status metadata for the manifest composer to read.

3d. Write the outline with the Write tool to {resolved_path}. Sections in order:

• # Solution: Lesson Cleanup — {lesson_title} header with plan_id, lesson_id, kind, change_type, scope_estimate: surgical metadata
• ## Summary — one sentence echoing the lesson title and directive count (e.g., "Apply the 3 directives from lesson {lesson_id} as surgical changes.")
• ## Overview — list of derived skills and the lesson kind → change_type mapping
• ## Deliverables — one deliverable per directive, using the template structure from 3a

3e. Validate the written outline:

python3 .plan/execute-script.py plan-marshall:manage-solution-outline:manage-solution-outline \
  write --plan-id {plan_id}

3f. Skip Q-Gate — the recipe path is its own gate. Log the bypass:

python3 .plan/execute-script.py plan-marshall:manage-logging:manage-logging \
  decision --plan-id {plan_id} --level INFO \
  --message "(plan-marshall:recipe-lesson-cleanup) Skipped Q-Gate validation (recipe path)"

---

Step 4: Hand Off to Phase 4 (Manifest Composition)

Phase-4-plan reads the outline, sees scope_estimate: surgical, and applies the surgical cascade rules per change_type:

• surgical + bug_fix → Phase 5 keeps quality-gate only; Phase 6 keeps push, create-pr, lessons-capture only.
• surgical + enhancement → Phase 5 keeps quality-gate only; Phase 6 keeps push, create-pr, lessons-capture only.
• surgical + tech_debt → Phase 5 keeps quality-gate only; Phase 6 keeps push, create-pr, lessons-capture only.

In all three surgical cases, the composer drops automated-review and sonar-roundtrip from Phase 6, matching the success criteria for this recipe.

This recipe does not invoke phase-4-plan directly — phase-3-outline returns control to the orchestrator, which advances to phase-4-plan as usual. The manifest composer reads scope_estimate from the outline metadata.

---

Output

status: success
plan_id: {plan_id}
lesson_id: {lesson_id}
lesson_kind: {kind}
change_type: {derived}
scope_estimate: surgical
deliverables_count: {N}
outline_path: {resolved_path}
next_phase: 4-plan

---

Related

• plan-marshall:phase-1-init Step 5b — Moves the lesson into the plan directory before this recipe runs.
• plan-marshall:phase-2-refine Step 3c — Canonical proposed-fix verification, defined in marketplace/bundles/plan-marshall/skills/phase-2-refine/standards/proposed-fix-verification.md. Lesson-sourced plans (source: lesson in request.md) take phase-2-refine's Recipe Shortcut and bypass Step 3c, so this recipe's Step 2b is the surrogate that closes the gap.
• plan-marshall:phase-3-outline § Recipe Path — Loads this skill with the input parameters and skips Q-Gate accordingly.
• plan-marshall:phase-4-plan § Manifest Composition — Reads scope_estimate=surgical and applies cascade rules.
• plan-marshall:recipe-refactor-to-profile-standards — Sister recipe; same shape, different scope (codebase-wide tech-debt sweep vs. single-lesson surgical cleanup).