Plan Doctor Skill

Role: Post-hoc diagnostic for plan artifacts. Walks TASK-*.json files for one or all plans, scans title and description for lesson-ID-shaped tokens, verifies each token against the live manage-lessons inventory, and emits structured findings for any tokens that resolve to non-existent lessons.

In addition to the TASK-level lesson-ID sweep, scan --all also runs three plan-directory-shape diagnostics:

• orphan-plan-directory — a subdirectory under .plan/local/plans/ that lacks status.json, or has status.json but none of request.md / references.json / solution_outline.md.
• stuck-low-confidence-archive — a subdirectory under .plan/local/archived-plans/ whose status.json has metadata.confidence < 95 (or the project-configured threshold), every phase after 2-refine is pending, and metadata.archived_reason is absent.
• dangling-worktree — a subdirectory under .plan/local/worktrees/ whose corresponding .plan/local/plans/{name}/ directory does not exist.

This skill complements the at-write-time validation in manage-tasks (which prevents new bad references) by sweeping plans that may already contain stale or phantom lesson-ID references introduced before the at-write check existed (or via direct file edits that bypassed manage-tasks).

Enforcement

Execution mode: Select scan or scan-task-file based on inputs and execute immediately.

Prohibited actions:

• Do not re-implement lesson-ID detection — always import scan_lesson_id_tokens and verify_lesson_ids_exist from tools-input-validation (single source of truth — no duplication)
• Do not silently degrade to "no findings" when the live inventory is unavailable — propagate the typed LessonInventoryUnavailable / LessonRegexAnchoringError failures so live-anchor discipline is enforced
• Do not write findings outside the plan-scoped Q-Gate store — use manage-findings qgate add so they participate in the standard triage loop

Constraints:

• Strictly comply with plan-marshall:persona-plan-marshall-agent rules (one Bash command per call, no shell constructs, no improvisation)
• Findings are emitted to the Q-Gate store under phase 5-execute so they surface in the Phase 5 triage path; the script also prints a TOON summary to stdout
• Exit non-zero (1) when findings_count > 0 so CI / orchestrators can fail fast on regressions

When to Use

• After a plan has been planned (Phase 4) but before execution (Phase 5), to catch stale lesson-ID references introduced by hand-edited TASK files
• During plan retrospectives to verify all referenced lessons still exist in the inventory
• Globally (across all plans) as a periodic data-integrity sweep

Verbs

Script: plan-marshall:plan-doctor:plan_doctor

scan

Scan TASK-*.json files for one plan or for every plan under .plan/local/plans/.

# Single plan
python3 .plan/execute-script.py plan-marshall:plan-doctor:plan_doctor scan \
  --plan-id EXAMPLE-PLAN

# Every plan in the inventory
python3 .plan/execute-script.py plan-marshall:plan-doctor:plan_doctor scan --all

Parameters:

• --plan-id (mutually exclusive with --all): Scan a single plan
• --all (mutually exclusive with --plan-id): Scan every plan under .plan/local/plans/
• --no-emit (optional): Skip emitting findings to the Q-Gate store; only print the TOON summary. Useful for read-only audits.

Behavior:

• For each TASK-*.json file, scans title + description for lesson-ID-shaped tokens via scan_lesson_id_tokens
• Verifies every token against the live manage-lessons list inventory via verify_lesson_ids_exist
• Tokens that resolve to non-existent lessons become findings (rule: phantom_lesson_id)
• When --all is passed, additionally sweeps the three plan-directory-shape rules described below (orphan_plan_directory, stuck_low_confidence_archive, dangling_worktree); these rules do NOT run on a single-plan --plan-id call
• Each finding is written to the Q-Gate store (source: qgate, type: bug, severity: warning for orphan/dangling/phantom; severity: info for stuck-low-confidence; phase: 5-execute) unless --no-emit is passed
• Findings whose plan_id does not have a live .plan/local/plans/{plan_id}/ directory (dangling-worktree and stuck-low-confidence-archive cases) are emitted in the TOON payload but NOT written to a Q-Gate store — there is no destination plan-dir to receive them
• Exits 1 when any findings are produced; 0 otherwise

Plan-directory-shape rules (--all only)

orphan-plan-directory (severity: warning)

Triggers when a subdirectory of .plan/local/plans/ either:

• has no status.json, OR
• has status.json but none of request.md, references.json, solution_outline.md exist.

The finding carries a remediation field with one of three values:

• rm_rf — no logs/ content; the partial init produced nothing worth keeping.
• archive_with_reason — logs/ has content; archive with manage-status archive --reason orphan-init-incomplete (see D2 — manage-status archive --reason).
• operator_review — status.json claims current_phase: 6-finalize but artifacts are absent; needs operator decision (likely "stuck finalize on a shell" from a parallel session).

stuck-low-confidence-archive (severity: info)

Triggers when an archived plan (subdirectory of .plan/local/archived-plans/) meets ALL of:

• status.json has every phase after 2-refine in status: pending.
• status.metadata.confidence is a number strictly less than 95 (the default threshold).
• status.metadata.archived_reason is missing or empty.

This is advisory only — the archive is a record of an operator decision. The finding carries confidence and threshold fields so the operator can decide whether to (a) restore + raise the threshold or (b) annotate the archive with archived_reason. Do NOT auto-remediate.

dangling-worktree (severity: warning)

Triggers when a subdirectory of .plan/local/worktrees/ does not have a corresponding .plan/local/plans/{name}/ directory. The likely cause is a cleanup race or a failed git worktree remove on a prior finalize. Operator should inspect the worktree for uncommitted work, then remove it.

scan-task-file

Scan a single explicit TASK-*.json file. Useful for ad-hoc validation when a plan-id is not convenient (e.g., scratch files in .plan/temp/).

python3 .plan/execute-script.py plan-marshall:plan-doctor:plan_doctor scan-task-file \
  --plan-id EXAMPLE-PLAN \
  --task-file /absolute/path/to/TASK-001.json

Parameters:

• --plan-id (required): Plan identifier the task file belongs to (used as the plan-id in findings; not used to resolve the path)
• --task-file (required): Absolute or repo-relative path to a single TASK-*.json file
• --no-emit (optional): Skip emitting findings; only print the TOON summary

Behavior:

• Identical scanning rules to scan, but operates on a single file
• Findings still go to the Q-Gate store under --plan-id so the standard triage loop can pick them up

Output Contract (TOON)

Both verbs print:

status: success
checked_files: 12
findings_count: 4
findings[4]{plan_id,task_file,line,token,rule,reason,remediation,confidence,threshold}:
  EXAMPLE-PLAN,TASK-003.json,1,2099-01-01-00-001,phantom_lesson_id,phantom_lesson_id,,,
  EXAMPLE-PLAN,TASK-007.json,1,2099-12-31-23-999,phantom_lesson_id,phantom_lesson_id,,,
  doc-revamp,,,,orphan_plan_directory,orphan_plan_directory,operator_review,,
  2026-05-18-lesson,,,,stuck_low_confidence_archive,stuck_low_confidence_archive,,47.0,95.0
summary:
  plans_scanned: 1
  emit_to_qgate: true

The TOON table widens to the union of fields across all rule types — TASK-level rows leave the rule-specific cells empty, and rule-level rows leave the TASK-specific cells empty. The TOON parser surfaces each row as a dict keyed by the column names; missing cells parse as empty strings.

When the Q-Gate primitive REJECTS one or more of the emitted findings, two further top-level fields follow summary:

qgate_persist_failed: true
qgate_persist_failures[1]{title,detail,message}:
  "plan-doctor finding (phantom_lesson_id): EXAMPLE-PLAN","TASK-003.json cites 2099-01-01-00-001","plan_not_found"

Both fields are absent when every finding reached the store. A rejected finding is counted in findings_count but is NOT in the Q-Gate store, so a caller that reads findings_count alone would report findings the run failed to file — branch on qgate_persist_failed before treating findings[] as retrievable via manage-findings qgate list.

Errors (TOON, exit 1):

status: error
error: lesson_inventory_unavailable
message: "manage-lessons list exited 2: 'boom'"

Field	Description
status	success (always for non-fatal results) or error (fatal — see error table)
checked_files	Number of TASK-*.json files actually parsed
findings_count	Number of phantom lesson-ID references discovered
findings[]{plan_id, task_file, line, token, rule, reason, remediation, confidence, threshold}	Per-finding rows. line is 1 for phantom rows (the file is JSON; per-line attribution is best-effort). Rule-level rows leave TASK-specific cells empty.
rule	Identifies the diagnostic: phantom_lesson_id, orphan_plan_directory, stuck_low_confidence_archive, or dangling_worktree. Mirrors reason for backward compatibility.
reason	Same value as rule (retained for backward compatibility with the original single-rule contract).
remediation	Set only on orphan_plan_directory rows: rm_rf, archive_with_reason, or operator_review.
confidence / threshold	Set only on stuck_low_confidence_archive rows.
summary.plans_scanned	Distinct plans inspected
summary.emit_to_qgate	true when findings were emitted, false for --no-emit
qgate_persist_failed	Present and true only when the Q-Gate primitive rejected at least one finding; absent when every finding reached the store
qgate_persist_failures[]{title, detail, message}	Present alongside qgate_persist_failed. One row per rejected finding, carrying the finding content that never landed plus the primitive's rejection message

Error codes

Error	Meaning
invalid_plan_id	--plan-id failed canonical validation (kebab-case)
plan_not_found	The plan directory does not exist for --plan-id
task_file_not_found	--task-file does not exist on disk
task_file_unreadable	The file exists but failed to parse as JSON
mutually_exclusive	--plan-id and --all were both supplied
lesson_inventory_unavailable	The manage-lessons list subprocess failed — surfaces as a fatal error to enforce live-anchor discipline
lesson_regex_anchored_to_drifted_inventory	The canonical LESSON_ID_RE matches none of the live IDs — the regex shape has drifted from reality

Canonical invocations

The canonical argparse surface for plan_doctor.py. The plugin-doctor analyzer (_analyze_manage_invocation.py) reads this section as source-of-truth for the manage-invocation-invalid and missing-canonical-block rules. Consuming docs xref this section by name instead of restating the command inline. See `pm-plugin-development:plugin-script-architecture` cross-skill-integration.md § "Script invocation in documentation".

scan

python3 .plan/execute-script.py plan-marshall:plan-doctor:plan_doctor scan \
  [--plan-id PLAN_ID] [--all] [--no-emit]

scan-task-file

python3 .plan/execute-script.py plan-marshall:plan-doctor:plan_doctor scan-task-file \
  --plan-id PLAN_ID --task-file TASK_FILE [--no-emit]

Cross-References

• plan-marshall:tools-input-validation — Owns scan_lesson_id_tokens and verify_lesson_ids_exist (the single source of truth for lesson-ID detection; this skill never re-implements either function)
• plan-marshall:manage-findings — Receives the Q-Gate findings (qgate add) so the standard Phase 5 triage loop can resolve them (FIX / SUPPRESS / ACCEPT)
• plan-marshall:manage-tasks — At-write-time validator (companion to this post-hoc sweep); together they cover both the prevention and detection paths for lesson-ID drift
• Lessons — The live-anchor discipline enforced here originates in lessons 2026-04-29-10-001 (regex/data drift surfacing as silent green) and 2026-05-03-21-002 (live-anchored test fixtures). See standards/check-lesson-id-references.md for the rationale and how to interpret/resolve findings.

Standards (Load On-Demand)

Read standards/check-lesson-id-references.md