[JacobStephens2]

Summary

Build an isolated Go prototype of the unattended Ticket and Pickup flows, with a measured comparison against the production Rust runner. Closes #1830.

prototype/go/
├── Ticket run → Claim → Worktree + embedded Factory skills → streamed Harness
├── Delivery → Base branch reconciliation → CI watch → bounded Repair
├── Pickup → Ready selection → origin reservation → Ticket run → leased release
├── tests → CLI fixtures, state exhaustiveness, bounded reservation model
└── benchmark.py → fresh Worktrees → Go/Rust measurements → research report

The Linux prototype builds with CGO_ENABLED=0, embeds the Factory skills, validates Origin match and trusted issue context, and supervises the streaming JSON Harness with process-group signal escalation. Ticket runs retain failed work and complete Claim ending after interruption. Delivery captures one PR, checks its published head, reconciles the Base branch and performs bounded Repairs.

Pickup selects the lowest-numbered Ready issue, creates ADR-0017's parentless origin reservation with an empty lease, rechecks Claim generation before dispatch, and releases only its own token. The bounded BFS checks both dispatch invariants and retains counterexamples for the unguarded protocols. Typed state enums have a test-time exhaustiveness check. Go has its own module, CI job and affected-test routing.

The comparison report and measured artifact record cold/incremental builds, test compilation/execution, static binary size, source counts, subprocess ergonomics, exhaustiveness and model-checking trade-offs. The two measured samples are pinned to 82cc4dc2; the Rust executable and tests cover a larger product, so the report does not infer a complete rewrite's speedup. ADR-0001 remains unchanged. Specs, other run types, additional Harnesses and production continuation/re-run policies remain outside this prototype's support.

The whole-Spec review also fixed rejected acquisition incorrectly ending another actor's Claim, shared GitHub CI classification, and consolidated the two incomplete-CI readiness regressions.

Evidence

Execution evidence

• Before: cd prototype/go && PATH=/usr/local/go/bin:$PATH go test -run '^TestSpecStaleClaimMustNotReleaseSomeoneElsesClaim$' -count=1 -v exited 1: labels = map[enhancement:true ready-for-agent:true], want [enhancement ready-for-agent in-progress].
  After: the same command exits 0: --- PASS: TestSpecStaleClaimMustNotReleaseSomeoneElsesClaim (0.66s); package ok in 1.268s. A rejected acquisition now leaves the incoming Claim intact; cleanup still runs after an uncertain or interrupted PUT.
• cd prototype/go && PATH=/usr/local/go/bin:$PATH go test -run '^(TestInterruptedClaimMutationStillEndsClaim|TestPickupRefusesClaimWhenGateLabelAppearsAfterRead)$' -count=1 -v: PASS.
• python3 -m unittest discover -s scripts -p 'test_*.py': 37 tests, OK. python3 prototype/go/bundle-skills.py --check: exit 0. go vet ./... and gofmt checks: PASS.
• Final PATH=/usr/local/go/bin:$PATH scripts/affected-tests main: PASS; go test ./... passed in 160.22s and the full cargo test suite passed in 2394.68s, without retries.

Test seams: the compiled prototype CLI against temporary Git repositories and fake gh/Harness executables; the benchmark CLI against stand-in compiler tools; the state-exhaustiveness checker and bounded Pickup model.

Unaddressed findings

Standards

None. All three findings were addressed: Claim ownership, repeated CI classification, and duplicated readiness assertions.

Spec

None. The single finding was the same Claim ownership defect and was reproduced and fixed.

Security

None.

Security review outcome

No unaddressed introduced findings.

Review coverage

Both independent reviewers read all changed code, tests, documentation, workflow and benchmark artifacts, plus the review fixes. No unexplained unread changed files remain. Intentionally not read: generated prototype/go/factory-skills.json (414,908 bytes). Both reviewers accepted its exclusion after reviewing bundle-skills.py, skills.go and the CLI assertions; python3 prototype/go/bundle-skills.py --check proved byte equivalence to skills/**, and go test -run '^(TestTicketRunLinksEmbeddedSkillsAndStreamsSession|TestBinaryIsStaticLinux)$' -count=1 -v passed for embedded content, skill links, Git exclusions and static ELF behavior. Each axis's complete files-read list and exclusion evidence are recorded in the Session review reports.

Merge Danger

Door: two-way

The prototype is isolated; ADR-0001 remains unchanged.

Blast Radius: prototype

Affected-test selection and a separate Go CI workflow change how prototype checks run. The experimental executable mutates labels, origin refs and pull requests when deliberately run on a repository.

Built with codex codex-cli 0.164.0 · gpt-6.1-sol · xhigh

Tickets

• Go prototype: isolated static binary, embedded skills, and Ticket run through Harness session #1832 landed with Go prototype: embedded skills and supervised Ticket-run session #1841
• Go prototype: Ticket run Delivery, CI Repair loop, and exhaustive state modeling #1833 landed with Go prototype: deliver Ticket runs and repair failing heads #1849
• Go prototype: Pickup run (ADR-0017 reservation) and in-test BFS state-space model checker #1834 landed with Add Go Pickup and bounded reservation model checking #1850
• Go prototype: reproducible Go vs. Rust benchmark runner and comparison report #1835 landed with Benchmark the Go prototype against Rust #1854