mattpocock-skills-opencode/docs/runtime-checks.md

Runtime checks — verified against real OpenCode v2

Environment: Windows 11, opencode v2.0.29, @opencode/plugin@2.0.29, 2026-10-14. Method: scripts/smoke.ts boots opencode serve (private server, Basic auth user opencode, password printed on stdout) in a temp fixture and asserts against the HTTP API. Result: PASS.

Verified

• Registry install by package name (1.10, 2026-10-14): @haylan/opencode-pocock@0.1.0 published to the Gitea npm registry by the dispatched release workflow (run 5684, success; tag v0.1.0 created, package linked to repo). Anonymous registry reads work (HTTP 200 without auth). A clean fixture with "plugins": ["@haylan/opencode-pocock"] + scope registry in ~/.npmrc loads the plugin: /api/plugin shows source {type:"package", target:"@haylan/opencode-pocock", version:"0.1.0"}, state.status "active"; skills inject from the bun cache path (~/.cache/opencode/npm/@haylan/opencode-pocock@latest/.../skills/...); npx -y @haylan/opencode-pocock init installs the three agents (modes decode correctly). Id collisions with a pre-existing global ~/.agents/skills Matt Pocock install resolve to the file-based copy (expected; a clean machine gets all 31 from the package).
• Plugin loads and activates from a local directory entry (plugins: ["<abs path>/dist"]); GET /api/plugin shows haylan.pocock with state.status: "active".
  • A file entry (file://.../dist/index.js) is IGNORED — local plugin entries must be directories with a root index/server file (confirms fastcrw research). npm-name entries resolve through exports (to be re-verified after first registry publish, TODO 1.10).
• Skill injection works: all 31 manifest skills registered (GET /api/skill); the 16 user-invoked ones carry autoinvoke: false. The API omits autoinvoke when true (default).
• DMI honored natively: file-loaded skills with upstream disable-model-invocation: true (e.g. a global ~/.agents/skills install) show autoinvoke: false without our plugin.
• ID collisions: when the same skill id also exists as a file-based skill (global install), the file-based entry is what /api/skill reports; plugin injection coexists without errors. (pr, absent from the tester's global install, was served from our package path.)
• Injected content must have frontmatter stripped to match file-based loading (markdown.content); the plugin does this since 2026-10-14.
• Agent markdown with V2-native frontmatter decodes correctly from .opencode/agents/: pocock (primary), pocock-worker / pocock-reviewer (subagent), descriptions intact, and the full permissions array survived decoding in order — verified on pocock-worker: skill * deny + tdd/codebase-design/diagnosing-bugs allow, subagent * deny, and all git/gh/glab shell denies present.
• Default permission floor (v2.0.29): agents start with {action:"*",resource:"*",effect:"allow"} plus specific rules, among them external_directory /tmp and /tmp/* allow — the /tmp/pocock-workers worktree root needs no extra user config on this build (README still documents the rule for other setups/subagents).
• Injected skills do NOT become slash commands (GET /api/command shows none for them; Skill.Info has no slash field). User-invoked steps are started via conversation with pocock (propose → yes → load) or by any agent explicitly loading the skill. File-based installs (global ~/.agents/skills, or a skills: [...] directory source) keep slash behavior.
• Server API: prefix /api/*, Basic auth (opencode + printed password); SPA HTML answers unknown non-/api paths with 200 — assert JSON, not status. POST /api/location/reload exists. Plugin loading is lazy (appears seconds after boot); /openapi.json requires auth.

Known environment quirks (tester machine, not product issues)

• Windows: server.kill() under shell:true only kills cmd.exe; smoke uses taskkill /PID <pid> /T /F. Temp fixture deletion can lag behind handle release (EPERM warning); harmless, retried on next run.
• The user's global ~/.agents/skills already contains most Matt Pocock skills — smoke assertions are written to tolerate id collisions (presence + flags, not source path).

2.5 orchestration run (2026-10-14, omniroute/qwen3.8-27b)

Result: PASS — full implement-spec flow, all 3 tickets, review, fixes, clean exit.

Fixture: fresh clone of the pocock-agents greeting reference (AGENTS.md ## Agent skills block → docs/agents/issue-tracker.md local tracker, GLOSSARY.md, .scratch/greeting/spec.md + 3 tickets — 01+02 independent, 03 blocked by 01+02, src/greeting.js + node:test, no remote, no deps). Agents copied to .opencode/agents/; opencode.json = {"$schema":…,"plugins":["G:/_DEV/repos/mattpocock-skills-opencode/dist"],"model":"omniroute/qwen3.8-27b"} (model is a valid top-level v2 config key — confirmed against opencode.ai/config.json).

Command (cwd = fixture, background, > run.log 2>&1):

opencode run --agent pocock --auto "Implement the spec in .scratch/greeting/spec.md using
implement-spec: build all three tickets with parallel workers, review, and finish. The tracker is
local markdown (see docs/agents/issue-tracker.md). Do not ask me anything else; proceed through the
whole flow autonomously."

Duration: ~63 min (11:00:26Z → 12:03:47Z) on the local 27b model. EXIT_CODE=0.

#	Check	Result	Evidence
1	Integration + ticket branches exist and are merged	PASS	spec/greeting is an ancestor of main; git merge-base --is-ancestor confirms 01, 02, 03 all in spec/greeting. 01 + 03 ff-merged; 02 diverged after 01 landed → resolved via merge commit 0308c30 "Merge branch 'spec/greeting' into ticket/greeting/02", exactly the agent's documented resume-worker path.
2	Worktree dir used + cleaned	PASS	All worktrees under /tmp/pocock-workers/pocock-25-fixture/greeting/ ($WT default). git worktree list shows only the main checkout after the run; orchestrator logged git worktree remove --force … tickets/{01,02,03,review-fixes} → CLEANED. Leftover empty notes/ + tickets/ dirs (no worktree admin entries) — harmless.
3	Tests pass on integration branch	PASS	Fresh worktree of spec/greeting: npm test = 16/16 pass, 0 fail (also 16/16 at every in-run merge point).
4	Flow evidence in run.log	PASS	→ Skill "ask-matt", → Skill "implement-spec", → Skill "code-review"; 4 pocock-worker dispatches (ticket 01/02 parallel, "Worker 02: merge integration tip" resume, ticket 03) + 4 pocock-reviewer dispatches (Spec axis, Standards axis, + compact re-runs); TDD red→green commit pairs per ticket (test(…): red… → feat(…)); no ✗ permission denials — zero worker escape attempts (no push/stash/out-of-allow-list skill). Only failure in the whole log: one ✗ Glob hit by the user's global codebase-memory "graph-first" rule (environment quirk, below).
5	Model actually served	PASS	Every turn banner: > pocock · qwen3.8-27b — top-level "model" in opencode.json took effect for orchestrator and all nested subagents.
6	Ticket statuses updated	PASS	All three issues/NN-*.md now read Status: done with a ## Comments note naming the branch — per the local-tracker convention in docs/agents/issue-tracker.md.

Flow detail worth keeping: 02's divergence was handled by resuming worker 02 to merge the new integration tip (agent text: "It knows its own change best, so it's the merger") — the non-ff path worked, not just the happy ff path. Review produced 2 actionable Standards findings + 1 real Spec gap (farewell error-message criterion untested); orchestrator dispatched a review-fixes worker for them and merged back to main. spec/greeting then ff-merged to main (no remote → delivered on default branch, per agent text).

Product bugs: none. Everything above is agent/plugin behaving per its documented design.

Findings / model-weakness notes (not product bugs):

• Skill-id collision resolved to the user's global copy. When the orchestrator loaded code-review, the content came from ~/.agents/skills/code-review/SKILL.md (pre-existing global Matt Pocock install), not the plugin-injected copy — same collision documented in the 1.10 note. The model then said the skill text was "truncated" and hunted for the SKILL.md on disk (found the global one). Our plugin ships the full body (code-review body = 6177 chars, no slicing in src/index.ts), so this is the collision + local-model behaviour, not truncation by the plugin.
• graph-first rule collision. The user's global codebase-memory-mcp injects a rule that forces a graph query before any glob. One orchestrator glob was rejected with that error; the model recovered by using find. Not our product; note for anyone running this with codebase-memory installed.
• Local 27b is slower than capable. The whole run took ~63 min; reviewer subagents were CPU-busy for ~14 min with no log growth (subagent output streams on completion, not incrementally — expected, not a stall). No stalls requiring intervention; no permission denials; the only judgment shortcuts were self-declared in the final summary (skipped a second post-fix review pass; left spec-prescribed module duplication alone).

Suggested agent improvements (NOT applied — for the owner):

• pocock.md: consider noting that when a skill's content looks truncated, the model may re-read the SKILL.md from disk; if the user also has a global Matt Pocock install, the id resolves to the file copy — worth a one-line "skills may resolve to a global copy; content is authoritative either way".
• pocock-worker.md/pocock.md: the non-ff resume path works but is model-dependent; a short reminder that "a ticket that can't ff-merge is normal when a parallel ticket merged first — resume its worker" could make the 27b class of model take the documented path more reliably.