local-run

name	local-run
description	Run the full Hydra pipeline locally (builder → reviewer → security)
metadata	{"category":"Pipeline","tags":["pipeline","test","local"]}

Run the full Hydra pipeline locally against a disposable test repo.

Input: Flags are optional. Example: /local-run --spec-dir /tmp/todo-mvp-spec --repo-url https://github.com/algorithm-conduction/todo-app

Defaults (if no flags given):

• --spec-dir: ask the user
• --repo-url: ask the user
• --github-user: derived from repo-url, or algorithm-conduction
• --repo-name: derived from repo-url, or todo-app

Steps — execute all steps without interruption. Only stop if something fails.

1. Parse flags from the user's input. If --spec-dir or --repo-url are missing, ask once.

2. Setup test repo

   • Load secrets/.env (source it, never read/display contents)
   • Rename existing repo to <name>-<timestamp> via GitHub API (museum pattern, ignore if 404)
   • Create fresh repo with auto_init: true via GitHub API
   • Upload all files from the spec dir as openspec/changes/<spec-basename>/ via GitHub Contents API

3. Build images from repo root:

   podman build -f images/builder/Dockerfile -t localhost/hydra-builder:test . -q
   podman build -f images/reviewer/Dockerfile -t localhost/hydra-reviewer:test . -q
   podman build -f images/security/Dockerfile -t localhost/hydra-security:test . -q
   

   Use ${CONTAINER_RUNTIME} if set, otherwise auto-detect podman/docker.

4. Run Builder

   bash scripts/dev-run.sh <spec-dir> builder --repo-url <repo-url>
   

   Extract result from output (turns, cost, error status).

5. Detect PR URL from GitHub API — find the open PR on the repo.

6. Run Reviewer + Security in parallel (use background Bash tasks):

   bash scripts/dev-run.sh - reviewer --repo-url <repo-url> --pr-url <pr-url>
   bash scripts/dev-run.sh - security --repo-url <repo-url> --pr-url <pr-url>
   

7. Report summary — print a table:

   • Per agent: turns, cost, verdict (pass/fail)
   • Total cost
   • PR URL
   • Log file paths

Constraints:

• Never display secrets/.env contents
• Never ask permission between steps — the user pre-approved by invoking /local-run
• If builder fails, stop and report — don't run reviewer/security
• Parse JSONL logs for result extraction (type=result → is_error, num_turns, total_cost_usd)