Treat assignment as the work signal — handle assigned PRs, not just issues

[ClaydeCode]

Summary

Make assignment the single signal for clayde to work. Today the loop only acts on assigned issues; a pull request assigned directly to clayde (with no parent issue) is never picked up, so review feedback on it is never addressed.

Reproduction

PR #80 ("feat: post-merge wrap-up") was opened by clayde on branch worktree-feat+post-merge-wrapup — outside the issue → plan → implement → PR flow. It has a CHANGES_REQUESTED review from max-tet (whitelisted) with two inline comments. Clayde never touches it: the poll loop runs fine every 5 min but only iterates the 4 assigned FreeshardBase issues. No state entry, no assigned issue → no code path to the PR.

Root cause

The loop is anchored entirely on assigned issues:

• src/clayde/github.py — get_assigned_issues() drives everything; orchestrator.main() iterates only its result.
• A PR is reachable only as a child of a tracked issue: _handle_issue() checks for review activity via issue_state["pr_url"] (src/clayde/orchestrator.py:114). State (src/clayde/state.py) is keyed by issue html_url; there is no concept of a standalone PR work item.
• Result: a PR assigned to clayde with no originating issue has no state, no parent, and is invisible to review handling.

Desired behavior

Assignment is the trigger, regardless of type:

1. The loop fetches everything assigned to the bot — issues and PRs. (Note: GitHub models PRs as issues; get_user().get_issues(filter="assigned") already returns PRs, each exposing a .pull_request attribute. Verify and branch on it rather than assuming.)
2. For an assigned PR with no linked issue: clayde addresses new review activity (reviews + inline comments) on it directly — same behavior as the existing review path, just without requiring a parent issue. State is keyed by the PR url; track last_seen_at / review ids the same way.
3. Existing assigned-issue flow (plan → approval → implement → PR → review) is unchanged.
4. Whitelist filtering still applies — only act on review content from whitelisted users, consistent with current safety.py gating.

Acceptance criteria

• A PR assigned to clayde, not linked to any tracked issue, gets its CHANGES_REQUESTED review addressed: clayde pushes commits to the PR branch and posts a summary comment.
• The assigned-issue lifecycle continues to work as before (no regression).
• PR work items persist in state.json (keyed by PR url) so review activity is detected incrementally across ticks, not re-processed each cycle.
• Reviews authored by the bot itself are ignored (as today).
• Tests cover: assigned-PR-with-no-issue review handling, and the issue-vs-PR dispatch branch. uv run pytest green.

Out of scope

Driving net-new implementation from a bare PR description (no review). This issue is about routing review feedback on assigned PRs. PR #80 is the concrete first target once this lands.

[ClaydeCode]

I've implemented this in PR #82.

What changed:

• github.is_pull_request_item(item) — detects PR items in the assigned list via html_url
• safety.filter_pr_reviews(reviews, github_username) — whitelist-gates review content, excluding the bot's own reviews
• tasks/pr_work.py + prompts/pr_work.j2 — new task that fetches PR context, filters reviews, invokes Claude, and posts a summary comment; state is keyed by PR URL
• orchestrator._handle_standalone_pr() — full lifecycle handler (skip/invoke/pure-approval/crash-recovery/usage-limit retry) mirroring _handle_issue()
• orchestrator._is_pr_tracked_as_issue() — prevents double-handling PRs already tracked under a parent issue
• orchestrator.main() — now dispatches PR items to the new handler and issue items to the existing handler

325 tests pass (uv run pytest).

[ClaydeCode]

Implemented Fix #81: extended the orchestrator to treat assigned PRs as first-class work items. Added is_pull_request_item() in github.py, filter_pr_reviews() in safety.py, a new tasks/pr_work.py task with prompts/pr_work.j2, and _handle_standalone_pr() + _is_pr_tracked_as_issue() + dispatch logic in orchestrator.py. The existing issue flow is unchanged. 325 tests pass. Opened PR #82 and posted summary comment on issue #81.