v5.8 integration verification
Status: unit and live matrix passed on Linux/tmux; macOS/Herdr remains manual Date: 2026-11-05 Ticket: #104 Parent specification: #99 Branch: feat/v58-104 Base: integration/ensemble-v5.8 at 972d02e
This record covers System Agent init and deliberate skill invocation. It does not close the parent specification or the integration ticket.
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 pins @opencode/plugin 2.2.1 and OpenTUI 0.7.0 for type checking; the live checks exercise the installed OpenCode 2.2.5 host. The 2.2.1 compatibility report remains the compatibility baseline, and the v5.8 behavior described there is the same behavior exercised here.
Commands
| Command | Result |
|---|---|
npm run verify | Pass — strict type check, then 632 tests passed, 0 failed, 0 skipped. |
npm run test:live:fake | Pass (final run) — two earlier executions failed for intermediate reasons described below; the final run passed every v5.8 assertion. |
npm run test:live:fake:empty | Pass — empty model catalog, explained empty state, wizard cancellation, no memory writes. |
npm run test:live | Pass — packaged plugin against the real host catalog; the model-dependent branch ran (no “no models available” skip line). |
Both live scenarios used the harness's isolated project, HOME, XDG directories, and loopback fake provider where applicable. No real workspace data was read or modified.
Live coverage
The fake scenario is the deterministic regression net for the v5.8 interactions:
| Acceptance area | Observed live evidence |
|---|---|
| System Agent init progress and report | I, :init, and /init each produced a visible /init audit line and a finalized Ensemble init complete System Agent report; .ensemble/init.json recorded version 2. |
/init audit lines | Snapshot asserted exactly three user messages with text /init, one per trigger. |
| ` | ` palette flags |
| Assigned-skill invocation | Selecting a leader-owned skill inserted /ensemble-live-skill, created one owner-assigned turn, and preserved the exact instruction. |
| Multi-skill grouping | $FakeAgent /ensemble-live-skill /ensemble-live-skill-two multi-skill-check produced one replied turn, one assignment row, the grouped skill title, one attributed reply, and the exact text. |
| Multi-agent routing | $FakeAgent ... !Helper ... produced one assigned row per anchor, each with its own resolved skill and reply. |
| Mixed anchors | A skilled anchor plus a normal anchor produced the skill turn and the normal turn with the same exact request text. |
| Leader triage | A General skill routed to the leader with its skillGrants visible (the single triage source of truth); the native session showed the allow rule while running and the Skillbox baseline after the turn. |
| Delegatee boundary | The leader's delegate received the self-authored guidance with no skillGrants, not a host permission. |
| Owner picker | A message-start skill owned outside the leader opened the owner picker; the chosen owner received the message with its text unchanged. |
| Validation errors | Wrong-owner, non-leader General, Direct-conversation General, and over-cap requests were rejected atomically; the wrong-owner and non-leader cases kept the draft and named the exact skill/owner/leader; no work rows were created. |
| Solo init | The built-in /init General row executed the System Agent job, and /init extra was rejected as solo-only. |
The real scenario covers the packaged plugin, Ensemble page, leader help, agent/workspace wizards, composer agent/file completions, group send, command palette, native-key handoff, the v5.8 palette flags, prose /init as ordinary text, and a visible System Agent init run to a terminal report on OpenCode 2.2.5. It is not a duplicate of every fake interaction assertion.
Documentation review
docs/implementation.mdnow records exact-text routing, one prompt per unique agent per message, grouped skills, separate-message sequencing, no per-skill bodies/tasks, the System Agent identity, General-skill leader triage, one-run grants and restoration, no attachment bypass, and the deliberate internal native init call.GLOSSARY.mddefines Skill, General skill, Skill invocation, System Agent, and Init job with the v5.8 behavior.src/v2/tui/help.tsteaches/and$Agent /palette scoping,Skill | Generalversus owner flags, General leader triage,/initsolo-only behavior, exact-text grouping and ordering, and validation recovery. The help catalog completeness test remains green.docs/compatibility/opencode-2.2.1.mdrecords the no-attachment boundary, one prompt per agent, one-run leader grant, and the deliberate nativeinitcommand inside the temporary session.README.md,scripts/README.md, anddocs/agents/opencode-tui-runtime.mdpoint at this record and describe the v5.8 live coverage and current host version.
Not run / remaining manual check
- macOS/Herdr: not run. This environment is Linux/tmux. Follow the repeatable macOS/Herdr checklist on that host; it is manual and unexecuted until a run is recorded there. The collision-free shortcut intent, inline leader label,
Esc/Ctrl+Gunder a live native run, System Agent streaming,/and$Agent /palettes, and/initunder Herdr remain a manual check. Do not treat this record as macOS evidence. - Pixel-level TUI proof: the harness asserts content and behavior, not exact ANSI frames or colors.
- Cross-process service lifecycle: the live checks use isolated harness services, not the shared hosted service.
Intermediate runs
The final passing runs are the ones recorded above. Two intermediate fake executions failed while this ticket was being prepared:
- The first
npm run test:live:fakefailed at the shared help filter (filtered help names the first match): the newcomposer.skill-validationhelp prose also matched the existingdelete workspacefuzzy query, changing the expectedmatch 1/1to two results. The help wording was tightened without weakening the test or changing the live harness, and the unit help suite stayed green. - The next fake execution failed at the pre-existing T4 work-list fixture with
T4 work fixture did not reply: the delegated row was visible aspending, the workspace reported0 working, and it did not reachrepliedwithin the fixture's 60-second poll. No v5.8 assertion had failed, and the next rerun passed every v5.8 check.
Both failures are recorded rather than hidden. The T4 scheduling timeout should be watched if it recurs; it is not claimed as a v5.8 regression from this run.
Commands not run
- No test was skipped. The only unperformed verification is the manual macOS/Herdr check described above.
Code review fixes
After the five v5.8 tickets were merged, the final two-axis code review found spec and standards issues. They were fixed in 91d455b and merged at 6265df1.
Spec fixes:
- Native init now receives Ensemble's full initialization instructions, not the short steering note.
- Real and empty-catalog live checks gained v5.8 assertions.
task.delegaterefuses a triage parent; reassignment was already guarded.- Prose
/initis ordinary text; only message-start or anchor-attached init tokens start the System Agent job. - Restart reconciliation settles only running init jobs.
message.sendcannot bypass planner ownership with explicit recipient IDs.- Skill messages retain
Referenced files:blocks. :initwith extra text or mentions is rejected as solo-only.
Standards fixes:
- Removed source-text and constant-only tests where behavior assertions were practical.
- Trimmed glossary entries to domain definitions.
- Consolidated triage metadata to one source of truth.
- Extracted the duplicated planner skill-error cascade.
- Consolidated init aliases and the init progress constant.
- Removed the unused init cancellation-state parameter.
- Moved the composer command catalog beside composer parsing.
- Refactored transcript attribution to one rendered author source.
Final verification after the review fixes: npm run verify passed with 637 tests; test:live:fake, test:live:fake:empty, and test:live all passed; git diff --check clean.