v5.9 verification — agent guardrails and feedback loop
Status: verified locally on Linux/tmux; hosted CI not run; macOS/Herdr checklist manual and unexecuted Date: 2026-11-06 Specification: #105 Tickets: #106 through #112 Branch: fix/v59-review Base: integration/ensemble-v5.9 at 7c79e48
This is the specification-level verification record. It covers all seven v5.9 tickets and the post-merge two-axis code review fixes. The earlier 2026-11-05-v59-t2.md and 2026-11-05-v59-t7.md records remain ticket-scoped; this -v59.md record is the evidence the tracker audit looks for. The record does not close #105 or any ticket. At the time of writing, GitHub reports #105 with 7/7 children closed, so it is ready for a maintainer to close after this record lands.
v5.9 is environment and tooling work. It does not change composer behavior, init, skill routing, permissions, or any other product runtime surface.
Environment
| Item | Value |
|---|---|
| OS | Linux |
| Terminal host | tmux 3.5a |
| Node | 26.12.1 |
| npm | 12.2.1 |
| OpenCode | 2.2.5 |
fd | fdfind 10.2.0 |
| Git | 2.47.3 |
The repository type-checks against @opencode/plugin 2.2.1. The live checks exercise the installed OpenCode 2.2.5 host. These are different version surfaces, as recorded in the runtime guide.
Ticket coverage
| Ticket | Delivered |
|---|---|
| T1 #106 | test:live:fake:all, test:live:all, the fail-fast PASS/FAIL/SKIP matrix, and test:live:latency. The known flaky T4 fixture now uses a bounded 180s wait with a subtask diagnostic and no retry loop. |
| T2 #107 | CI live job for the fake/empty suite with pinned Ubuntu, Node, OpenCode, tmux, and fd; require-mode preflight; local skip-and-exit-zero preserved. |
| T3 #108 | Reviewer standards for judgement calls and the two-axis review workflow, linked from the contribution guide. |
| T4 #109 | Table-driven service routing regression matrix covering skill text against explicit recipients, file references, DM/group, multi-agent, and mixed anchors. |
| T5 #110 | ADR 0001: Native-skill boundary, linked from implementation and agent instructions. |
| T6 #111 | One-time unused-symbol cleanup, then noUnusedLocals and noUnusedParameters in tsconfig.json; npm run verify green. |
| T7 #112 | Version-spec closure rule, read-only npm run issue:hygiene, handoff reference split, and the repeatable macOS/Herdr checklist. |
Review fixes
Spec findings:
| Finding | Resolution |
|---|---|
Matrix assumed /tmp/opencode existed | scripts/live-matrix.sh creates the temp parent before mktemp. src/v2/live-matrix.test.ts runs the matrix with a clean TMPDIR and no pre-existing opencode directory. |
| Skip detection scraped log text | The harness writes ENSEMBLE_LIVE_SKIP_MARKER on the local skip path; the matrix reads that file. Local skip still exits zero and require mode still fails. Covered by the matrix tests. |
| Hygiene matched the T2 ticket record as spec evidence | verificationRecord now requires the spec's own -v59.md record, not -v59-t2.md. src/v2/issue-hygiene.test.ts covers ticket-only versus final-record directories. |
Preflight required fd for the fake suite | scripts/live-preflight.sh defaults to fake scope and requires/prints fd only for all. CI keeps its fd install. |
| Historical v5.6/v5.7 records were edited | Reverted those two files to their delivered content. Current v5.8 and v5.9 records keep the checklist link. |
| Only ticket records existed for v5.9 | This final record. |
| Documentation index omitted v5.9 and v5.6 | docs/README.md lists v5.9, v5.8, v5.7, and v5.6. |
| Hosted CI was not run | Recorded under Not run below; no hosted CI pass is claimed. |
Standards findings:
| Finding | Resolution |
|---|---|
Latency probe duplicated the › cursor literal | The probe reads settingsCursorMark(true) from src/v2/tui/settings.ts instead of restating the marker. |
Dead railAgentLabel helper and test | Removed both; railGroupLine remains the live helper. |
| Duplicated subtask node reader | live_subtask_read is the single reader for field polling and timeout diagnostics. |
issue-hygiene.mjs exported internal helpers | Exports removed; tests exercise the CLI at its public seam. |
| Duplicated skip-versus-fail prose | Scripts README is the authoritative contract; contribution, runtime, and review docs point to it. |
| Preflight tool/version checks were inconsistent | One version normalizer compares Node and OpenCode exactly; the fd gate uses the same skip/require path as other tools. Covered by src/v2/live-preflight.test.ts. |
Commands
All commands ran sequentially on the branch before handoff. No hosted CI job was triggered.
| Command | Result |
|---|---|
npm run verify | Pass — strict type check, then 673 tests passed, 0 failed, 0 skipped. |
npm run test:live:fake:all | Pass — fake 167s, fake-empty 8s; 2 passed, 0 failed, 0 skipped, 0 not run in 175s. The suite also passed an earlier run (174s) before the final harness-reader refactor. |
npm run test:live | Pass — real host catalog; packaged plugin, palette, prose /init, and visible System Agent init verified. No skip line. |
npm run test:live:preflight | Pass — fake scope: node v26.12.1, opencode v2.2.5, tmux 3.5a; fd not required. |
npm run test:live:preflight -- all | Pass — also reports fd 10.2.0. |
npm run issue:hygiene | Pass — reports #105 open with 7/7 children closed and docs/verification/2026-11-06-v59.md. |
npm run issue:hygiene -- --all | Pass — #105 resolves to this record; #99, #90, and #78 resolve to their final records; no ticket record is used as spec evidence. |
node --test --experimental-strip-types src/v2/live-matrix.test.ts | Pass — 3 tests: clean-host temp root, structural skip marker, required-skip failure. |
bash -n scripts/live-matrix.sh scripts/live-preflight.sh scripts/live-latency-probe.sh scripts/lib/live-harness.sh scripts/live-check-fake.sh scripts/live-check-fake-empty.sh scripts/live-check.sh | Pass — no shell syntax errors. |
git diff --check | Pass — no whitespace errors. |
The aggregate matrix ran the scenarios in order and printed one summary. The fresh-host matrix test uses a temporary TMPDIR with no pre-existing opencode directory and a HOME without the nvm OpenCode install, so the scenarios take the local skip path without starting a server or model.
Not run
- Hosted GitHub Actions: not triggered or executed from this environment. The workflow file and equivalent local commands were checked, and the deterministic suite passed locally, but this record does not claim a passing hosted CI run. The
livejob still requires its first hosted execution. npm run test:live:allas one command: not invoked. Its fake, empty, and real scenarios were run separately (test:live:fake:allthentest:live), so the full matrix content is covered without a second three-minute fake run.- macOS/Herdr: not run. This environment is Linux/tmux. The repeatable checklist remains manual and unexecuted until a run is recorded there.
- Real-model checks in CI: intentionally not part of the CI job; they stay local and are not required.
Tracker state
- #105 remains open and was not closed by this branch.
- #106–#112 are not closed by this record.
- #99, #90, and #78 were not reopened or rewritten.
- The v5.6 and v5.7 verification records were restored to their delivered content; only the current v5.8 record gained a checklist pointer.