feat: share PR context in harness 2.4

[Wattenberger]

Summary

• Mint Harness 2.4 and make it the current harness version for new sessions.
• Add a v2.4 specialist bundle carrying the PR-context handoff text for coordinators, implementors, and verifiers.
• Have verifiers consume the delegated note, re-derive only missing fields, append findings, and cite base-SHA CI instead of rebuilding by default.
• Preserve all released prompt bundles byte-for-byte; add v2.4 golden coverage to pin the three overrides and unchanged inherited specialists.

Versioning rework

• Rebased onto main after feat(harness): mint v2.3 with a levers-focused Suggested Next Steps hint #1713 claimed Harness 2.3, then re-minted this feature as Harness 2.4.
• Harness 2.4 inherits Harness 2.3 agent instructions exactly and overrides only the implementor, spec-writer, and verifier specialist bodies.
• The existing Harness 2.3 module and prompt resources remain unchanged.
• The test-only v2_3_goldens current-version assertion now verifies that its registry entry remains available; the new v2.4 golden owns the current-version assertion.
• Existing sessions pinned to older harness versions retain their original immutable prompt resources.

Base-rebuild count

• Sample: 26 recent verifier sessions.
• Result: 3 of 26 used git worktree add to build a base checkout; 0 used a detached checkout.
• Conclusion: the rate does not justify the optional per-SHA record-store follow-up.

Verification

• Focused Harness 2.4 golden tests: PASS (4/4).
• Retained Harness 2.3 golden tests: PASS (4/4).
• Harness registry tests: PASS (10/10).
• make check: PASS.
• make test: PASS (8,264 passed, one leak-classified test, 4 skipped) with serial build/test execution, incremental compilation disabled, and an increased Rust stack for toolchain stability.
• Diff checks confirm released Harness 2.3 prompt resources and older bundles are unchanged.

Review fixes

• Mint a new harness version instead of mutating released prompt bundles.
• Prefer the note Base SHA; derive it from the actual PR base ref only when missing.
• Eliminate duplicate-note races by having the coordinator create exactly one note before parallel delegation.

Notes

• The user-level verifier specialist on this host is customized and not synchronized with bundled prompts, so it is intentionally unchanged.
• No wire-protocol or frontend changes.

[augmentcode]

^{Deep Code Review Agent🐛}

Review completed with 2 suggestions.

_{Reviewed commit: 056b345}

[augmentcode]

🤖 Augment PR Summary

Summary: This PR introduces a per-branch PR-context note for specialist handoffs.

• Implementors record task IDs, SHAs, gate outcomes, failures, and changed files in a tagged note.
• Coordinators are instructed to pass that note when delegating verifiers.
• Verifiers consume and extend the note, relying on base-SHA CI for pre-existing failures by default.
• Specialist prompt tests add string-level checks for the new collaboration contract.

_{🤖 Was this summary useful? React with 👍 or 👎}

[augmentcode]

Review completed. 1 suggestion posted.

Comment augment review to trigger a new review at any time.

[panghy]

Feedback from panghy: rather than editing the existing bundled prompt instructions in place, this should bump the harness/prompt-bundle version and add the new PR-context-note contract as the new version — existing instruction sets should stay as-is for sessions pinned to them. Please rework the PR to version-bump instead of mutating the current instructions.

[panghy]

Need to bump harness version if changing system prompts

[Wattenberger]

Addressed the requested prompt-versioning change in dcfc612: the new instructions now ship only in Harness 2.2, while the v1 and v1.1 resources are restored byte-for-byte to origin/main. Added Harness 2.2 registry/bundle wiring and golden tests; the full 8,129-test suite passes.

[Wattenberger]

Thanks — addressed by rebasing onto main after #1705 and re-minting this feature as Harness 2.3. Main’s Harness 2.2 instructions and specialist resources remain untouched; 2.3 inherits those instructions and overrides only implementor, spec-writer, and verifier. I also changed the v2.2 current-version golden to an availability assertion, matching the older-version pattern, while the new v2.3 golden owns the current-version check. The rebased head is 10b1f54; make check and all 8,264 tests pass. The existing re-review request remains in place.

[Wattenberger]

Rebased again after #1713 minted Harness 2.3 and re-minted this PR-context change as Harness 2.4 at 66d2e5d. Main Harness 2.3 doctrine and specialist resources remain unchanged; 2.4 inherits its instructions and changes only implementor, spec-writer, and verifier. The only edit to a 2.3 file is test scaffolding in crates/intent-services/src/v2_3_goldens.rs: its former current-version assertion is now an availability assertion, while v2_4_goldens owns the current-version check. make check, focused 2.3/2.4 and registry tests, and the full 8,264-test serial suite pass.

[Wattenberger]

@panghy re-review requested: the prompt-versioning approach you asked for is in place at 66d2e5d as Harness 2.4 (main's 2.1–2.3 untouched, byte-identical). 18/18 checks green, no open threads; the standing changes-requested review from 2026-09-03 is the only merge blocker.