feat: issue-level implement_issue reasoner — use SWE-AF as a sub-harness

[AbirAbbas]

Summary

Adds an issue-level entry point, implement_issue, so a main coding harness (Claude Code / Codex / OpenCode) can delegate one fully-scoped issue to SWE-AF as a cheap sub-harness. The feature-level build pipeline is unchanged — this is the fast path for the second prompt shape ("change this code in this file, like this"), where the caller has already done the planning.

No planning agents run. Deterministic (non-LLM) git setup creates an isolated worktree + issue/<build_id>-<slug> branch off base_branch, the existing coding loop implements the issue (reduced budget: 3 iterations, reviewer-gated; QA + synthesizer when needs_deeper_qa), an optional single verifier pass checks the acceptance criteria, and the branch is returned as the deliverable. Nothing is pushed and no PR is opened by default — the caller owns merge and CI. Typical cost is 4–8 LLM calls instead of hundreds, targeting ~10–30 min on cheap models instead of hours.

Registered on both nodes (swe-planner.implement_issue, swe-fast.implement_issue) and mirrored 1:1 in the Go port (swe-planner-go / swe-fast-go).

Changes

• swe_af/issue/ — new package: IssueSpec/IssueBuildConfig/IssueBuildResult schemas, deterministic git ops (worktree/branch/checkpoint-commit, repo-local .git/info/exclude masking so the caller's git status stays clean), and the implement_issue orchestrator driving run_coding_loop
• swe_af/app.py, swe_af/fast/app.py — include the shared issue_router on both nodes
• go/internal/issue/, go/internal/config/issueconfig.go, go/internal/node/register.go — Go port with identical reasoner name, input schema, and behavior; node surface tests updated to 31/30
• README.md — "Use SWE-AF as a Sub-Harness" section: when to use implement_issue vs build, async curl + webhook/poll recipes, result shape, config table, fan-out guidance
• .claude/skills/delegate-issue/ — ready-made Claude Code skill for the delegation flow (.gitignore narrowed from .claude/ to .claude/* + !.claude/skills/ so versioned skills ship while local state stays ignored)

Validation contract (tests derive from these)

• The result branch contains commits implementing the issue; the caller's working tree, current branch, and git status are untouched
• N concurrent calls on the same repo_path yield N independent branches (worktree-per-call, retry on repo-lock contention)
• No planning roles are ever invoked; total agent calls stay within the configured budget
• Default config pushes nothing to any remote and opens no PR
• Verifier failure or iteration exhaustion surfaces unmet criteria/debt (success: false keeps a salvageable branch; zero-commit builds delete their branch)
• Same reasoner name and input schema on all four nodes (Python + Go, planner + fast)

Test plan

• Python: 47 new contract tests in tests/issue/ (real temp repos + real coding loop; only the agent dispatch is scripted)
• Full suite at CI parity (Python 3.12, no .env): 1084 passed, 1 skipped + compileall clean
• Go: new tests in internal/issue + internal/config; full CI gate green — gofmt clean, go build, go vet, go test -race -count=1 ./... with GOWORK=off

Note for reviewers: origin/feat/pipeline-optimization contains related skip-worktree/direct-git commits; this PR deliberately implements its own minimal git ops in swe_af/issue/git_ops.py rather than cherry-picking, since the issue-level path only ever owns a single worktree/branch.

---

🤖 Generated with Claude Code