speckit-harness-status

name	speckit-harness-status
description	Render a compact, budget-aware slice of the harness state with a recommended next action
compatibility	Requires spec-kit project structure with .specify/ directory
metadata	{"author":"github-spec-kit","source":"harness:commands/speckit.harness.status.md"}

Harness Status — Budget-Aware Context Rendering

Render the externalized harness state as a compact working slice. This is the Harness-1 (arXiv:2606.02373) "budget-aware context rendering" step exposed as a command: instead of re-reading transcripts or whole files, the agent (and the user) get a bounded snapshot that is sufficient to decide the next action.

Run it at any time — especially at the start of a fresh session to resume research with zero context carried over, or before deciding whether more exploration is worth the budget.

This command is read-only: it writes nothing and consumes no budget.

User Input

$ARGUMENTS

Optional. full renders larger slices (3× the configured sizes); a topic string filters curated/candidate rows to those matching it.

Steps

1. Locate state

Resolve HARNESS_DIR as in /speckit.harness.init. If no harness exists, say so and point to /speckit.harness.init — do not create anything.

2. Load slices only

Read configuration for slice sizes (rendering.*), then load:

• budget.md — mission(s), budget table, stop conditions, last 5 action-log rows.
• curated.md — top curated_slice entries by importance (critical first); count the rest.
• candidates.md — up to candidates_slice rows with status new/inspected; count the rest.
• verification.md — all refuted rows, plus counts by verdict; list critical curated entries lacking a verified record.
• observations.md — last observations_slice entries.

Do not read evidence.md bodies, spec/plan files, or anything beyond the slices. Keep the rendered output within the configured context_tokens cap; truncate lowest-importance material first and say what was truncated.

3. Render the snapshot

# Harness Status — <feature> (<date>)

**Mission**: <mission lines>

**Budget**: searches <r>/<b> · inspections <r>/<b> · verifications <r>/<b> remaining

**Curated set** (<shown>/<total>, cap <cap>):
| ID | Importance | Finding | Verified? |
...

**Open frontier** (<shown>/<total> candidates): ...

**Verification**: <n> verified · <n> refuted · <n> unverifiable
⚠ Unverified critical claims: <list or "none">
⚠ Refuted (do not rely on): <list or "none">

**Recent activity**: <observation slice, one line each>

4. Recommend exactly one next action

Close with a single recommendation, derived from the snapshot:

• Unverified critical claims → /speckit.harness.verify.
• Mission gaps with budget remaining → /speckit.harness.explore <question>.
• Marginal gain exhausted or budget empty → /speckit.harness.report.
• Coverage complete and verified → /speckit.harness.report, then proceed to /speckit.plan or /speckit.tasks.
• Empty harness (no curated entries, no actions logged) → /speckit.harness.explore <mission>.

State the reason in one sentence (e.g. "2 critical claims unverified and 12 verification budget remaining → verify before planning").

Guardrails

• Read-only. No file writes, no budget changes, not even timestamp updates.
• Slices, never full files; counts stand in for what is not shown.
• The recommendation must follow from the rendered state alone — if the state files are missing or malformed, report that instead of guessing.