heard-chef

Heard, Chef

Address post-submission feedback: PR review threads, PR comments, QA ticket comments, and pasted feedback. Categorize feedback, fix what belongs in scope, draft/print or post replies, run needed gates, and push.

Don't use this for: starting new ticket work, creating PRs, or reviewing someone else's PR. Use kickoff, ship-it, or pr-patrol instead.

Project config lives at <repo>/.beislid/workflow.md (typed-key fenced YAML blocks; format reference at workflow-md-format.md). Capabilities are probed lazily on first need per probe-semantics.md. Output prose follows output-templates.md and heard-chef-templates.md.

---

Read workflow.md

Read <repo>/.beislid/workflow.md. If it doesn't exist, hard-fail and stop:

🛑 No workflow.md found in .beislid/. If this is a fresh project, run /setup to create one. If you moved your config, restore it to <repo>/.beislid/workflow.md.

Validate line 1 exactly:

<!-- beislid-workflow: v1 -->

If missing or wrong, hard-fail:

⚠️ This workflow.md is <found> but heard-chef only knows v1. Update Beislið or downgrade workflow.md by hand to v1 syntax. Heard-chef will not silently mis-parse.

Compute cache identifiers:

repo_hash=$(git rev-list --max-parents=0 HEAD | sort | head -c 12)
workflow_hash=$(git hash-object .beislid/workflow.md)

Try to read ${BEISLID_STATE_DIR:-$HOME/.local/state/beislid}/probes/<repo_hash>.json:

• Missing → empty in-memory probe state; cache state cold.
• Present but workflow_hash mismatch → empty in-memory probe state; cache state stale.
• Present and hash matches → load capability entries; cache state fresh. Per-cap freshness uses probe_cache.ttl_hours from workflow.md, default 24.

Print orientation prose from heard-chef-templates.md (≤240 chars).

---

Internal: probe()

Reusable lazy-probe contract. Do not re-probe within a run; the in-memory probe state is authoritative.

1. Look up <cap> in memory.
2. If present, status: ok, and within TTL → return ok; no probe ran.
3. Else run the probe per probe-semantics.md for the cap kind. Multi-command CLI capabilities (pr_review_source, pr_review_update, ticket_update) probe as one logical cap by checking every unique binary.
4. On success → return ok.
5. On failure → use call-site-specific prose from heard-chef-templates.md.
6. On (a) → re-probe. On (b) → mark session_skip: true and use the call-site fallback for this session. On (c) → abort; do not write cache.

On clean run end, write probed/re-probed entries back to <repo_hash>.json, excluding session_skip. Preserve doctor-owned fields when present. If workflow.md changed mid-run, do not overwrite stale state; surface the write failure and continue.

---

JIT phase protocol

Heard-chef phases are authoritative only after their aux files are read. At each phase boundary, read the matching file and follow it. If a phase file cannot be read, hard-fail instead of continuing from memory.

• Phase 1: skills/heard-chef/phase-1-detect.md
• Phase 2: skills/heard-chef/phase-2-fix.md
• Phase 3: skills/heard-chef/phase-3-push.md

When BEISLID_VERBOSE=1, print one aux load stamp after successfully reading each phase file. Stamp format is ✓ heard-chef/<phase-file-stem> v1 loaded for phase-1-detect, phase-2-fix, and phase-3-push.

Checklist

1. Detect — read phase-1-detect.md. Outputs: ticket/PR identity, selected mode, normalized feedback queue with source metadata.
2. Categorize and fix — read phase-2-fix.md. Outputs: fix commits or draft-only decisions, reply drafts, change classification, approval state.
3. Push and reply — read phase-3-push.md. Outputs: gates run/skipped, branch pushed, replies posted/printed, review re-request status.

---

Phase 1: Detect

Read and follow phase-1-detect.md. Print the entry/exit one-liners from heard-chef-templates.md. Do not continue to Phase 2 until feedback is loaded or strict paste fallback has provided enough context.

Phase 2: Categorize and Fix

Read and follow phase-2-fix.md. Print the entry/exit one-liners from heard-chef-templates.md. Do not offer fast path unless every condition in the aux file is met.

Phase 3: Push, Reply, and Re-request

Read and follow phase-3-push.md. Print the entry/exit one-liners from heard-chef-templates.md. Gate, push, and update behavior must use configured workflow.md capabilities and safe temp-file payloads.

---

Common mistakes

• Using hidden GitHub writes. PR review reads/writes come from pr_review_source and pr_review_update; otherwise paste/print manually.
• Continuing without feedback context. Source failure (b) means strict paste now, not blind skip.
• Offering fast path for judgment calls. Fast path is only for obviously clear fetched feedback with non-manual update paths.
• Interpolating reply bodies into shell commands. Use JSON temp files for PR review updates and body/title temp files for ticket updates.
• Creating child tickets without approval. Always show title/body and wait for approval.
• Running from memory after aux read failure. Restore/reinstall the missing phase file instead.

Key principles

• Workflow.md is the source of truth. One per-project config for kickoff, heard-chef, and ship-it.
• Lazy probes, not preflight tables. Probe at first use and narrate failures conversationally.
• Source metadata matters. Normalize feedback together, but reply through the correct source/update channel.
• AFK only inside explicit rails. Manual/paste sources and ambiguous feedback keep the run human-in-the-loop.