feat: add coverage mode to task-manager-agent and goal-manager-agent

[bborbe]

Adds the reverse map both manager agents were missing.

task-manager-agent and goal-manager-agent each checked necessity in the forward direction only — a task → the goal box it serves, and a goal's tasks → the box each serves. Nothing answered the inverse: for each open box, what drives it, and what evidence exists for it.

What ships

A new coverage action on both agents, emitting one row per closure box:

<box verbatim> | <driving item or "none"> | <evidence: measured(<source>) | asserted | none> | <proposal: create | re-scope | split | diagnosis-task | —>

• goal subject — boxes: # Success Criteria + # Definition of Done; drivers: the tasks linked in # Tasks
• topic subject — boxes: # Success Criteria + # Completion Gate (a topic carries no # Definition of Done); drivers: the member goals under # Goals

The subject kind is resolved goal-first and reported, because the two kinds differ in both halves of the map.

/vault-cli:task-status --coverage and /vault-cli:goal-status --coverage render the table below the anchor pair.

Why measured must name a source

A rating that scans a box's own text for a quoted result reads nothing, and is indistinguishable from a real reading unless the row names where the result came from. So measured(<source>) names the Progress entry or artifact — measured(progress 2026-10-10), measured(PR #240), measured(tag v0.126.0) — and a measured row carrying no source is a defect rather than a pass.

Likewise a none driving cell must carry a proposal naming the missing prerequisite; — against none is an omission, not a pass. Each run ends with a reconciliation line whose parts must sum to the box count.

Both actions are read-only — they never modify, create, re-link or tick anything, on the subject or on any driver.

Verification

• make precommit — passes (8 Go packages, 10 rule suites, golangci-lint, go vet, osv-scanner, trivy fs, CHANGELOG structure, addlicense).
• make check-versions — all four version strings equal at 0.168.7; the new ## Unreleased heading does not disturb the alignment check.
• Markdown-only change: agents/ and commands/ are executable surface, not docs.

[ben-s-pull-request-reviewer]

Now I have all the information needed for classification. Let me do the in-session classify and adjudicate:

Classification (Step 4c-sel):

• agent-cmd/agent-frontmatter → applicable (agents/*.md changed)
• agent-cmd/command-frontmatter → applicable (commands/*.md changed)
• agent-cmd/command-thin → applicable (agents/.md + commands/.md changed)
• agent-cmd/gap-driven-feedback → applicable (agents/.md + commands/.md changed)
• agent-cmd/no-user-prompts → applicable (agents/.md + commands/.md changed)
• agent-cmd/scripts-in-claude-dir → applicable (agents/.md + commands/.md changed)
• agent-cmd/single-source-of-truth → applicable (agents/.md + commands/.md changed)
• changelog/conventional-prefix-required → applicable (CHANGELOG.md changed)
• git-commit/feature-branch-no-tag → skipped — @commits trigger, not a diff rule
• git-commit/imperative-mood → skipped — @commits trigger, not a diff rule
• git-workflow/never-direct-commit-to-master → skipped — @commits trigger, not a diff rule
• go-security/resource-ownership → skipped — @commits trigger, not a diff rule
• go-security/tenant-isolation → skipped — @commits trigger, not a diff rule
• markdown-todo/lowercase-x-for-complete → applicable (**/*.md changed)

Adjudication (Step 4d-sel):

All 8 applicable rules were reviewed against the diff. All agents have valid frontmatter (name, description, tools, model), all commands have valid frontmatter (description, argument-hint, allowed-tools), CHANGELOG uses feat: prefix. No violations found. No mechanical findings from Step 4a.

Traceability: 15 candidates, 8 applicable, 6 skipped (all @commits-triggered), 1 skipped (not a code change)

---

PR Review: feat: add coverage mode to task-manager-agent and goal-manager-agent

Reviewer: selector mode, no sub-agents spawned
Files changed: 5 (CHANGELOG.md, agents/goal-manager-agent.md, agents/task-manager-agent.md, commands/goal-status.md, commands/task-status.md)
+134 lines, −6 lines

Selector Mode Traceability

• Candidates: 15 rules matched by Step 4b-i glob filter
• Applicable: 8 rules selected by Step 4c-sel (8 ≤ 15)
• Skipped:
  • git-commit/feature-branch-no-tag → @commits trigger not applicable to diff review
  • git-commit/imperative-mood → @commits trigger not applicable to diff review
  • git-workflow/never-direct-commit-to-master → @commits trigger not applicable to diff review
  • go-security/resource-ownership → @commits trigger not applicable to diff review
  • go-security/tenant-isolation → @commits trigger not applicable to diff review
• Mechanical funnel: 0 findings from 79 YAMLs run

Must Fix (Critical)

None.

Should Fix (Important)

None.

Nice to Have (Optional)

None.

Notes

• The feature is well-scoped: a new coverage action added to both agents, gated behind a --coverage flag on the two slash commands. The specification correctly identifies the inverse-map gap (forward necessity checking already existed; reverse coverage hadn't been answered).
• The goal-manager-agent coverage action correctly handles the goal/topic distinction for subject resolution and for closure-box identification (topic has no # Definition of Done — uses # Completion Gate instead).
• The task-manager-agent coverage action correctly excludes parse_checkboxes for box collection (it greps the whole file including # Tasks which are drivers, not boxes), requiring section-bounded parsing instead.
• Both actions correctly enforce: measured requires an external named source; — proposal against none driving cell is an omission; reconciliation line has two independent partitions; evidence shape is exactly one row per box.
• The commands correctly delegate to the agents without re-deriving the rating, per spec.

selector clean — no adjudication needed

---

{
  "verdict": "approve",
  "summary": "Feature adds a well-specified coverage action to both agents and commands. The reverse-map specification correctly handles the goal/topic subject distinction, box vs driver separation, evidence rating rules, reconciliation line partitions, and read-only constraint. No violations found in applicable rule set (agent/command frontmatter, changelog prefix, thin-command, gap-driven-feedback, no-user-prompts, scripts-in-claude-dir, single-source-of-truth, markdown-todo).",
  "comments": [],
  "concerns_addressed": [
    {
      "concern": "correctness: coverage action specification must accurately describe forward+reverse mapping logic",
      "disposition": "addressed",
      "detail": "Both agents' coverage actions correctly implement the reverse map (box→driver), distinct from the existing forward necessity check (driver→box). goal-manager-agent properly distinguishes goal vs topic subjects and their different closure boxes; task-manager-agent correctly uses section-bounded parsing to exclude # Tasks from box collection."
    },
    {
      "concern": "correctness: --coverage flag behavior must correctly render the coverage table per specification",
      "disposition": "addressed",
      "detail": "Both commands (goal-status.md, task-status.md) delegate to their respective agents with the coverage action, render the table verbatim below the anchor pair, suppress the status body, and explicitly do not re-derive the rating — matching the spec's requirements."
    }
  ]
}