speckit-proc
speckit-proc
Issue-to-PR pipeline. The operator triages via GitHub API, then drives a persistent worker session through specify, plan, tasks, implement, test, and deliver -- one resume per stage.
When to Use
- Implementing a GitHub issue that requires code changes
- Any dev task routed through the speckit pipeline
- MANDATORY for all code tasks dispatched through Pylot crews
Prerequisites
gh issue view $0 --repo $1 --json state --jq '.state' 2>/dev/null || echo "ERROR: cannot access issue"
Requires gh CLI with valid GH_TOKEN and worker dispatch scripts (spawn-worker.sh with --resume support, wait-for-worker.sh).
Procedure
This is a multi-stage ICM procedure. Each stage has an explicit contract (Inputs/Process/Outputs) in its CONTEXT.md. Read the contract, follow the process, write the output.
Arguments
- First positional argument: Issue number (e.g.,
42) - Second positional argument: Org/repo (e.g.,
fellowship-dev/pylot) --stage N-- resume from stage N (reads prior stage outputs)--review-- pause at checkpoint stages, exit withstatus=review
Output Location
Stage handoffs are written to:
.procedure-output/speckit-proc/
├── 01-preflight/
├── 02-specify/
├── 03-plan/
├── 04-tasks/
├── 05-implement/
├── 06-test/
└── 07-deliver/
Execution
-
Create output directory
mkdir -p ".procedure-output/speckit-proc" -
Read CONTEXT.md for the stage chain and worker session info.
-
Run stages sequentially:
For each stage:
a. Read
stages/0N-<name>/CONTEXT.mdb. Load inputs per the Inputs table (respect selective section routing) c. Follow the Process steps exactly as written d. Run Audit checks if present -- revise until all pass e. If checkpoint +--review: write output, exit withstatus=reviewf. Write handoff to.procedure-output/speckit-proc/0N-<name>/ -
Emit outcome marker:
[pylot] outcome="speckit-proc complete: <summary>" status=success
Resume
With --stage N: skip stages 1 through N-1, read their handoffs from .procedure-output/, start at stage N. If prior handoffs are missing, exit with status=failed.
Review Gates
With --review: stages with a Checkpoints section pause after completion and exit with status=review. Stages without checkpoints run straight through. Without --review, everything runs straight through.
Critical Rules
- Follow the stage contracts. The CONTEXT.md Process section is the spec.
- Load only what the Inputs table says. Extra context dilutes quality.
- Audit before output. Do not write the handoff until all checks pass.
- Docs over outputs. Reference files are authoritative, not previous handoffs.
- One stage at a time. Do not read ahead to later stages.
- Pre-flight is mandatory. No real data = garbage output. Never skip stage 01.
- One worker session. Stage 02 spawns it, stages 03-07 resume it. Do not spawn fresh workers per stage.
More from fellowship-dev/dogfooded-skills
entropy-check
Sensor — checks doc freshness and computes domain quality grades. Never fixes. Detects staleness, missing coverage, and FlowChad gaps. Updates QUALITY_SCORE.md. Skips inapplicable signals per repo.
16distill
Post-mission audit and distillation — capture mode classifies a completed mission using an 8-code failure taxonomy and writes an audit JSON; analyze mode aggregates audit JSONs into a findings report and creates GitHub issues with recommendations.
14migrate-skill
Move a skill from claude-toolkit plugin (or local .claude/skills) into the dogfooded-skills library, then import it back. Use when consolidating skills into the shared repo.
14skill-builder
Write a high-quality agent skill — covers frontmatter spec, section structure, quality criteria, and common antipatterns.
13popsicle
Agent-native onboarding doc generator — builds coverage maps, health baselines, generated docs, and agent adapters so any AI tool can autonomously navigate your repo.
8setup-harness
Scaffold the knowledge layer for a repo — ARCHITECTURE.md, QUALITY_SCORE.md, enhanced docs/code-structure.md, docs/code-guidelines.md, and FlowChad flow stubs. Gives agents a map, not a 1,000-page manual.
8