fix(e2e): stop paying opencode's per-directory plugin install ~40 times a run

[Soph]

https://entire.io/gh/entireio/cli/trails/1238

What

The e2e-tests (opencode) leg went red on main from 2026-09-03 22:10 UTC and stayed red. This makes it green again: run 33867981097 — all 56 tests passed, Run E2E Tests 469s (it was 1266s and failing).

Why it was failing

Every failure was agent failed: signal: killed at the 120s prompt budget, with zero assertion failures and completely empty stdout and stderr. Nothing in the artifacts named a cause, because opencode prints nothing while it stalls.

It was not a gradual degradation. The single-process bootstrap warmup was 9–15s for all 32 preceding runs, then 40s+ on every run after. Not ours, and not opencode's code either: same opencode version (1.18.27) on both sides of the boundary, same runner image, only trail-routing commits in our repo, the claude-code leg sharing the same API key got faster, our own hook perf was better in the red runs, and opencode's install path has one commit since Aug 25 (on 08-27).

What changed is that a cost which was always pathological stopped fitting in the budgets.

The mechanism

From opencode's source at 1.18.27: ConfigPaths.directories returns the global config dir plus every existing .opencode directory walking up from cwd. config.ts:452 forks a background @opencode-ai/plugin install for each of them. plugin/index.ts:184 — if (plugins.length) yield* config.waitForDependencies() — then blocks on all of them, and waitForDependencies joins the fibers with no timeout and logs nothing.

So a repo pays a 27-package, ~61MB install once it has both a project-local .opencode directory and a plugin configured anywhere, global or project. entire enable supplies both by writing .opencode/plugins/entire.ts. Every test gets a fresh repo, so we were paying it ~40 times in parallel.

It is also much worse than an npm install sounds: npm.ts:81 runs @npmcli/arborist's reify() in-process rather than the npm CLI. Measured on one machine, identical 27-package result: arborist 5m00.9s at 1% CPU, npm install 3s.

What this does

1. Collect opencode's logs into the artifact tree. This is what made the diagnosis possible at all, and it is the piece worth keeping regardless of the rest.
2. Pre-build the dependency tree once with the npm CLI and seed it into each test repo through a new RepoSeeder hook on SetupRepo. node_modules is symlinked, not copied — 61MB across thousands of small files times ~40 live repos would cost about what it saves. The seed also writes opencode's own .gitignore, which names its own file, so nothing planted is visible to git; these repos are the subject of checkpoint assertions.
3. Raise the warmup budget and report each attempt's duration, so the one serial, uncontended measurement of opencode startup is visible in the CI log rather than only in a downloaded artifact.

Every step degrades to today's behaviour instead of failing: no npm, an unparseable version, or a platform without unprivileged symlinks all leave the repo alone and let opencode install for itself.

Verification

Don't read the pass count alone — parse the newly-collected opencode-logs/opencode.log for the gap between the last loading path= line and the next event. That is the install window.

	broken	fixed
median gap	23.5s	0.15s
p90	57.5s	0.24s
killed inside it	31 of 53	6 of 60

The 6 remaining are processes ended by test teardown, not stalls — with a 0.35s max gap there is nothing left to stall on.

Worth knowing

One intermediate revision here looked fixed and was not: Bootstrap runs in go run ./e2e/bootstrap and SetupRepo in go test ./e2e/tests, two processes, so handing the tree between them via a package variable seeded nothing. Because seeding degrades quietly, the bootstrap log reported success while 31 of 53 processes still stalled. The tree is now resolved per-process, and TestOpenCodeSeedRepoPlantsDeps runs SeedRepo from a process that never called Bootstrap — the shape of that bug.

Not in scope

The same install hits real users under the same two conditions. Moving Entire's plugin to ~/.config/opencode/plugins/ would make it a one-time global cost instead of per-repo, at the price of loading in every project including ones where Entire was never enabled — a design call, not a refactor. Upstream: anomalyco/opencode#47212.

🤖 Generated with Claude Code

---

Note

Medium Risk
Touches only E2E/CI paths but adds cross-process caching, npm installs, and symlink semantics where quiet fallback could mask regressions.

Overview
Fixes opencode E2E timeouts caused by ~40 parallel, silent @opencode-ai/plugin installs (~61MB each) after entire enable adds a project plugin.

Adds a RepoSeeder hook so agents can plant files post-entire enable; opencode SeedRepo now writes opencode.json, copies pinned package.json / lockfile into .opencode, symlinks a shared node_modules tree built once per process via npm install (version-keyed under os.TempDir()), and mirrors opencode’s .gitignore so seeded blobs don’t affect checkpoint/git assertions. Bootstrap warms up in a seeded scratch dir through RunPrompt with longer first-attempt budgets and per-attempt timing logs.

CI copies ~/.local/share/opencode/log into artifacts on the opencode matrix leg. TestOpenCodeSeedRepoPlantsDeps asserts seeding works from the test process without relying on bootstrap-only state.

OpenCode-specific setup was removed from SetupRepo in favor of the seeder interface; failures to pre-build or link deps degrade to opencode’s own install rather than failing setup.

^{Reviewed by Cursor Bugbot for commit b055dcb. Configure here.}

[Copilot]

🟡 Changes recommended

The new seeding test currently assumes symlinked node_modules unconditionally and should be excluded from Windows builds to avoid expected failures.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

Pull request overview

This PR fixes recurring e2e-tests (opencode) timeouts by avoiding Opencode’s per-repo plugin dependency install during startup, while also improving diagnosability and making warmup behavior more observable.

Changes:

• Introduces an agents.RepoSeeder hook and updates repo setup to call it after entire enable, allowing agent-specific pre-run seeding.
• Implements Opencode repo seeding + once-per-process dependency prebuild (via npm install) and updates Opencode warmup budgeting/logging.
• Collects Opencode’s own logs into CI artifacts for post-failure diagnosis.

File summaries

File	Description
e2e/testutil/repo.go	Calls an optional RepoSeeder after enabling Entire, so agents can plant required files once per fresh repo.
e2e/agents/agent.go	Adds the RepoSeeder interface and documents its best-effort contract.
e2e/agents/opencode.go	Adds Opencode dependency prebuild + repo seeding and improves warmup budgets/timing output.
e2e/agents/opencode_seed_test.go	Adds a regression test to ensure seeding works from the test process (not relying on bootstrap-only state).
.github/workflows/e2e.yml	Copies Opencode’s log directory into the artifact tree on the opencode CI leg.

Review details

• Files reviewed: 5/5 changed files
• Comments generated: 1
• Review effort level: Lite

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.