β.2: vendor + pin opencode (closes #2)

[Rchari1]

Closes #2. Stacked on #7 (β.1 workspace layout) — base will retarget to main automatically when #7 merges.

Implements spec-20260610-155955-vendor-opencode (in amico#132): the extension no longer assumes opencode on $PATH.

What this is

• opencode.lock.json — committed pin: opencode v1.17.3, per-platform SHA256 (darwin-arm64 + linux-x64). Bumping the pin = one-file PR, gated by CI.
• scripts/fetch_opencode.mjs — zero-dep download-at-build: SHA256-verified against the manifest (hard fail, no retry/override on mismatch — supply-chain signal), unpack-to-temp + atomic rename, idempotent (.sha256 stamp → offline repeat builds), --record mode for pin bumps. vendor/ is gitignored.
• src/opencode_binary.ts — resolution: config override (dev-only) → vendored path from the extension install dir → hard error with fetch/reinstall instructions. No $PATH fallback (§4 Assumption 4 stays dead). Config default "opencode" → "". On a missing binary the extension still activates (inspector/watcher/commands) with chat disabled.
• test/boot_smoke.mjs — the S35 gate: synthesized project → vendored binary serve → GET /event → 200 text/event-stream → clean SIGTERM exit. No LLM creds needed. Runs as a two-platform CI matrix (ubuntu-latest + macos-14) and as the named test:smoke script β.4's packaging step will call.
• .vscodeignore — exists so vsce does NOT fall back to .gitignore (which would silently exclude the binary from the VSIX); !vendor/** negation makes the intent explicit.
• Spike smoke tests deleted (smoke_cli, smoke_opencode_boot, opencode_boot_harness, vscode_shim); extension now has a real vitest unit suite (10 tests) in the fast CI job (pnpm -r run test).

AC → evidence

Issue AC	Evidence
VSIX ships opencode at a fixed SHA; build fails unless /event SSE 200	opencode.lock.json + fetcher mismatch tests; local smoke: GET /event → 200 (text/event-stream) … PASS on v1.17.3; CI boot-smoke matrix
ServerManager spawns the vendored binary, not $PATH	opencode_binary.ts resolution tests; extension.ts wiring; config default ""

Notes

• v1.17.3 passed the smoke on the first try — no walk-back needed despite the spike being written against 1.3.x. (β.5 owns any SSE payload drift; this PR guarantees boot + /event.)
• Spec deviations (all noted in the plan): smoke file is boot_smoke.mjs (plain node) so vitest's glob can't pull a network test into the hermetic suite; the fetcher buffers the download in memory rather than a temp file (outcome-equivalent, nothing half-written); idempotent path prints up to date: instead of exiting silently.
• For β.4: package with vsce package per platform after fetch:opencode --platform <target>; .vscodeignore already ships vendor/**.

🤖 Generated with Claude Code

[jack-champagne]

β.2 looks solid — approving in spirit. Pulled the stacked branch locally and ran it end-to-end: build/typecheck green, fast tier 44 (amico-run) + 10 (extension) passing, fetch:opencode downloaded the real v1.17.3 linux-x64 and the committed SHA verified against the live GitHub release, and test:smoke drove the real binary to GET /event → 200 text/event-stream with a clean SIGTERM exit. CI is green on all three jobs, so the darwin-arm64 pin is machine-gated too. Supply-chain handling (archive-hash-before-unpack, hard-fail with no override, atomic temp→rename install, injectable download for hermetic tests) and graceful degradation (chat disabled but the extension still activates) are exactly right.

Two inline notes — one's a question for you on the leftover AMICO_* env wiring, the other a CI-cost suggestion. Neither blocks.

[jack-champagne]

@Rchari1 — this AMICO_* env + PATH=binDir:$PATH injection looks like leftover spike wiring, and it cuts against β.1's all-argv / no-AMICO_* contract (S37 / Q90 — amico-run takes everything via argv, nothing from env). I'm inclined to kill it. Before we rip it out: is there something the opencode server still needs these for (a tool the chat slice invokes that reads them), or can β.3 pass them through amico-run flags instead? Want your read.

[jack-champagne]

Both matrix legs re-download ~157MB on every run. Worth an actions/cache step keyed on hashFiles('packages/extension/opencode.lock.json'), restoring packages/extension/vendor/ before this step — the .sha256 idempotency stamp already makes fetch:opencode a no-op on a cache hit, so unchanged pins skip the download entirely.

[Rchari1]

Thanks — on your two notes:

• AMICO_* env + PATH=binDir:$PATH (extension.ts:88): you're right it's leftover spike wiring against β.1's all-argv/no-env contract (S37/Q90). Already removed in β.3+β.5: AGENTS.md + vetted template + chat→solve→inspector slice (closes #3) #9 (β.3 Task 4): the spawn env no longer carries AMICO_JULIA_SCRIPT/PROJECT/RUNS_ROOT — amico-run is argv-only, and chat→solve drives it purely via flags. PATH now carries only the amico-run launcher dir (so opencode's bash can resolve amico-run); the Julia project reaches the agent through AGENTS.md (substituted at session-copy), passed as --project. So nothing reads those env vars anymore.
• CI cache for the 157MB opencode download: good call — I'll add an actions/cache step keyed on hashFiles('packages/extension/opencode.lock.json') restoring packages/extension/vendor/ before fetch:opencode (the .sha256 stamp already makes it a no-op on a cache hit). Tracking it as a fast-follow so it doesn't churn the stack mid-review; flag if you'd rather I fold it in now.

(Branch force-updated from the β.1 rebase — same content, just restacked on the #7 fixes.)