feat: design provider-agnostic issue-to-PR automation pipeline

[ClaudiaFang]

Goal

Build a bot that automates the issue-to-PR lifecycle on a GitHub Project (v2) board, and later other trackers, driving multiple possible AI coding agents — not locked to a single vendor.

Pipeline

1. Issue status → Ready: bot analyzes the requirement and writes a design/plan back to the issue as a comment.
2. Human reviews and approves the plan.
3. An AI coding agent clones the repo, creates a git worktree/branch, implements the change, and runs the target repo's tests locally.
4. Bot opens a PR linked to the issue (Closes #N), moves the issue to Review.
5. If the human requests changes on the PR, the bot re-invokes the AI agent against the same branch to address feedback.
6. Human merges the PR → issue auto-closes (native GitHub behavior) → bot reconciles the project board status to Done.

Hard architectural requirement: provider-agnostic

Two integration points must be abstracted behind interfaces from day one, not hard-coded:

• Tracker provider — GitHub Issues + Projects v2 today, GitLab Issues/Boards as a future provider. Interface covers: watch/poll status transitions, read issue body/comments, post a comment, open a PR/MR linked to an issue, read PR/MR review state, update the project board status field.
• AI agent provider — Claude Code today, Codex CLI (and others) as future providers. Interface covers: given a repo path + task description + worktree, run the agent and return success/failure + a summary. The "write plan" step and the "implement" step both go through this interface and may use different concrete providers.

No orchestration code may call gh/Octokit or a specific agent CLI directly — only through the provider interfaces (src/providers/tracker/, src/providers/agent/).

Notes

• Repo is public; secrets (tracker tokens, agent API keys) must never be committed — env vars only, secret-scanning skill guidance applies.
• Local harness (feature_list.json) already breaks this into 12 trackable sub-features covering both provider interfaces, GitHub + Claude Code concrete implementations, and the orchestration state machine.

[ClaudiaFang]

Backlog priority review — 2026-08-04

Priority: P2

Reason: This automation directly supports the desired high-agent-throughput workflow, but it is a large platform project with tracker, agent, orchestration, credentials, isolation, review, and recovery boundaries. The current open PR only establishes the harness.

Scheduling note: Continue only as small vertical slices with explicit human approval gates. Finish and reconcile PR #2 first; prove one GitHub + one agent happy path before expanding providers or autonomous behavior.