Skip to content

opencode-prune.md

Latest commit

 

History

History
100 lines (76 loc) · 5.7 KB

opencode-prune.md

File metadata and controls

100 lines (76 loc) · 5.7 KB
Raw

Pruning the opencode experiment copy

opencode's experiment copy is the whole monorepo (34 packages), but a headless opencode run only touches ~a dozen of them. The rest — web/desktop/marketing apps, demo videos, test fixtures — is dead weight. Two sizes matter (don't conflate them):

.git objects (git-fetch download) working tree (checkout) on disk
full (current) 79 MB 184 MB 262 MB
after D ~12 MB ~73 MB ~86 MB

(Measured on a real fresh clone of the pruned seed: .git 13 MB + tree 73 MB.)

du on the copy shows the 262 MB on-disk total (objects + checkout). The trial container runs git fetch --depth 1, which transfers only the .git packfile (~79 MB) — dominated by non-compressible binary assets — and checks out the working tree locally. That download per container is the install slowdown.

Patch-back safety (the hard constraint). Pruning only ever deletes whole files/dirs; it never edits a kept file. So every kept file stays byte-identical to upstream, an evolution diff (git diff base..HEAD) touches only kept files and applies cleanly to upstream, and the deletions never show up in that diff. (Removed files also aren't reachable from the evolver's edits, so it can't depend on them.)

D — the conservative prune (dependency-safe; done first)

Removes only what is outside the headless-run dependency closure:

removed size what it is
packages/console 55 MB web dashboard (~40 MB of marketing .mp4)
packages/web 17 MB marketing website
packages/app (except vendor/) ~15 MB web app — but keep app/vendor/opencode-ai-client-1.17.13-v2.tgz (67 KB)
packages/desktop 13 MB Electron desktop app (1 MB .icns icons)
artifacts/ 7.4 MB demo videos
screenshot-uk.png + README.<lang>.md (~20) ~0.2 MB marketing image + translated READMEs

Gotcha (found empirically, not by static analysis): packages/session-ui — which D keeps — depends on @opencode-ai/client via file:../app/vendor/opencode-ai-client-1.17.13-v2.tgz, a 67 KB vendored tarball inside packages/app. Delete all of app and bun install fails (ENOENT extracting tarball from @opencode-ai/client). So keep that one tarball; drop the rest of app. (This is a different @opencode-ai/client than the runtime one: sdk-next uses workspace:* → packages/client, unaffected.)

Result: per-container download (git packfile) 79 MB → ~12 MB (~6.5×); on-disk 262 MB → ~86 MB (the kept 67 KB tarball is noise).

Verified (bun 1.3.14, the container's pinned version)

On the pruned tree (packages/app reduced to just vendor/):

  • bun install → 1539 packages installed, exit 0 (plain bun install, no --frozen-lockfile, so the seed's bun.lock referencing the deleted packages is fine — bun re-resolves in-container; the shipped bun.lock/package.json stay byte-identical to upstream ⇒ patch-safe).
  • opencode … run --help → exit 0, full run-command module graph loads, no import errors.
  • opencode … run --print-logs → agent fully boots: loads tools + session, selects the ai-sdk LLM runtime, and makes the real model streaming call (only fails on a deliberately-dead endpoint). No removed package is needed at runtime.

The earlier segfault was purely a bun 1.2.14 vs required 1.3.14 mismatch — not pruning.

Kept — the runtime closure

opencode + its workspace:* deps core, server, tui, sdk, sdk-next, llm, schema, codemode, plugin, http-recorder, protocol, script; the transitive pull-ins ui (via tui), client (via sdk-next), effect-drizzle-sqlite / effect-sqlite-node (via core); and root config (package.json, bun.lock, tsconfig*, bunfig.toml) + patches/ (bun patchedDependencies).

A — deeper prune candidates (not yet applied; need a bun install + run check)

Bigger but riskier — verify bun install + bun run packages/opencode/src/index.ts run --format json still work on the pruned tree first:

  • non-dep packages stats, session-ui, storybook, docs, enterprise, cli, containers, function, httpapi-codegen, identity, slack (~8 MB; some are explicit workspace entries, so bun.lock may need regenerating — a baseline-only change, still patch-safe). Note: dropping session-ui also removes the only reason to keep app/vendor/…tgz, so the whole of packages/app could then go.
  • packages/opencode/test/ fixtures (~10 MB — a 5 MB base64 PNG, a 2.6 MB image, a 4.7 MB JSON).
  • packages/ui assets (fonts/sprites, ~20 MB) — droppable only if headless run --format json never loads the TUI.

How it's applied

Wired at seed time in beagle/tools/onboard.py. The set lives in PRUNE_PROFILES["opencode"] (a list of git pathspecs); seed_from_upstream(..., prune=…) builds the reduced tree in the index (read-tree → git rm -r --cached --ignore-unmatch <pathspecs> → write-tree) and commits that as the orphan baseline — no working-tree checkout. Pathspecs that match nothing are ignored, so the profile degrades gracefully if a future opencode version restructures.

Onboard with --prune opencode:

python -m beagle.tools.onboard --upstream https://github.com/anomalyco/opencode \
    --ref <sha> --repo <you>/opencode_v1.18.16 --private --version 1.18.16 \
    --prune opencode --profile-name opencode_v1.18.16

Patch-safety holds because pruning only removes whole paths: kept blobs are byte-identical to upstream (the integration test asserts the kept packages/core/… blob hash equals upstream's), so an evolution git diff base..HEAD touches only kept files and applies cleanly to upstream.