loopdev

Prompt file imported from LARi-UQAC/ResearchTools (.github/prompts/loopdev.prompt.md). Copyright stays with the author.

Develop one feature through a design -> plan -> code -> comment -> test -> review -> score -> correct loop that stops on a composite quality gate or a hard budget cap. Thin wrapper over the loop-engineer skill - read .claude/skills/loop-engineer/SKILL.md and references/LOOP.md, then follow the pipeline and the model tiering exactly. The best cloud model (Fable 5) orchestrates and judges, cheaper tiers act, and code and comments are generated locally by the local-coder / local-writer agents so the heavy work is free.

Option contract:

/loopdev --budget <max_usd> [--score <min_score>] [--max-iters N] [--patience N] [--no-merge-pr] "<feature request>"

Procedure:

1. Parse the file(s) or topic given after the command in the chat message (if none was given, use the file currently open in the editor) for the flags and the feature request. --budget is required; refuse to start without it (no silent unbounded spend). Defaults: --score 90, --max-iters 10, --patience 2.
2. Confirm the environment: Ollama running with the local models (ollama list, and model_resolver.py --resolve --role <writer|coder> naming a qualified tag for each role - there is no fallback tag, an unqualified role is a stop), and the review plugins installed. The Agent SDK runs on your subscription auth - no gateway, no API key.
3. Run the driver:

   python ".claude/skills/loop-engineer/scripts/loop_engineer.py" --loop --budget <max_usd> --score <min_score> "<feature request>"
   

   Use --dry-run --reports r1.json r2.json to exercise the loop-control logic offline (guards, ledger) without the SDK or models.
4. The loop wraps evaluate -> correct -> rescore. It stops on the composite gate (tests green AND no CRITICAL/HIGH AND aggregate >= min_score), or on a hard stop: budget cap, max iterations, no-progress plateau, or a security regression. Security is a hard floor.
5. On a gate pass the loop stops at "ready to merge": the branch is pushed and a PR opened (token permitting, unless --no-merge-pr). The merge to a protected branch is human-gated - present the result and wait for explicit confirmation before merging.

Report at the end:

• The final aggregate score, per-axis sub-scores, and which stop rule ended the loop
• The spend vs the cap and the iteration count (from loop-budget.md / loop-run-log.md)
• The branch / PR at the merge gate, and any CRITICAL/HIGH findings that remain
• The PROCESS.md checkbox state

Never let the local model perform a merge, branch delete, or force-push; those stay with a cloud subagent or the human. Apply the pipeline directly; the only interactive checkpoint is the human merge confirmation.

the file(s) or topic given after the command in the chat message (if none was given, use the file currently open in the editor)