Retire fork compatibility shims and adopt upstream implementations #13

Tasks

Phase order matters. Each phase ends with bun run fork:verify (10/10 owned, 7/7
patched) and a build smoke test. Do not start a phase until the previous one is green.

Phase 0: Baseline (done during the 2026-08-11 sync)

• 0.1 Restore upstream catalog in package.json; keep only the 5 fork scripts
• 0.2 Restore patches/solid-js@1.9.10.patch; drop 3 orphaned downgrade-era patches
• 0.3 Revert drift in project.ts, app-runtime.ts, layer-node.ts, tool/registry.ts
• 0.4 Repoint phantom imports (max-steps.txt → core max-steps, withStatics → statics)
• 0.5 Re-apply ForkDistribution onto upstream's installation/index.ts (21 sites)
• 0.6 Fix Schema.Defect → Schema.Defect() (beta.83 made it a function)
• 0.7 Delete junk file packages/core/src/effect/dfdf
• 0.8 Confirm green: build + smoke test + fork:verify

Phase 1: Classify the divergence surface

• 1.1 Enumerate every file differing from upstream/dev under packages/*/src
  and packages/*/test, excluding generated gen/ trees
• 1.2 Classify each as feature (named fork capability) or drift (local copy,
  shim, or stale pattern). Record the verdict and a one-line reason per file
• 1.3 Cross-check every feature verdict against fork/manifest.json; anything
  claiming feature status but absent from the manifest is drift until argued otherwise
• 1.4 Publish the classification table in this change directory as inventory.md

Phase 2: Retire the legacy logger shim

• 2.1 Migrate packages/opencode/src/local/sync.ts (9 call sites) to Effect logging
• 2.2 Migrate packages/opencode/src/local/placement.ts (3 call sites)
  • pick() kept plain deliberately: it documents a synchronous slot reservation
    ("before any await"), and an Effect.gen yield boundary through that would be a
    real concurrency bug. It now returns a PickOutcome discriminated union and
    tool/task.ts logs in Effect context.
• 2.3 Migrate packages/opencode/src/beads/sync.ts (3 call sites)
• 2.4 Migrate packages/opencode/src/provider/provider.ts (2 call sites) — manifest-patched,
  keep the mergeDiscoveredModel marker intact
  • Plain promise chain, so discoverOpenAICompatibleModels now returns
    { models, warnings } and the Effect caller logs. Same shape as pick below.
• 2.5 Drop the unused import in packages/opencode/src/beads/beads.ts (0 call sites)
• 2.6 Delete packages/core/src/util/log.ts and its fork/manifest.json owned entry
• 2.7 Confirm no importer of @opencode-ai/core/util/log remains
  (note: packages/console/core/src/util/log.ts is a different, upstream file — leave it)

Phase 3: Retire the legacy defaultLayer DI pattern

• 3.1 Revert the 10 upstream-existing files that use defaultLayer to upstream,
  re-applying fork features per the Phase 1 classification
• 3.2 Migrate the 11 fork-only files to LayerNode — delete export const defaultLayer,
  keep export const node
• 3.3 Migrate remaining call sites: Effect.provide(X.defaultLayer) →
  Effect.provide(AppNodeBuilder.build(X.node))
• 3.4 Migrate the test block (~90 refs, largest single group), incl. fork-only
  test/loop/loop.test.ts and test/loop/queue-mode.test.ts
  • The merge had kept the fork's test files while taking upstream's source, so
    suites referenced defaultLayer on services that no longer export it, and
    provider.test.ts imported @/project/instance-layer, a module that does not
    exist (whole file failed to import).
  • A mechanical X.defaultLayer -> AppNodeBuilder.build(X.node) swap is WRONG:
    build() returns a closed layer, so each call constructs its own Database and
    a session written through one is invisible to the others. The working shape is
    upstream's: one module-level LayerNode.group root, compiled once with
    LayerNode.compile(root, replacements), mocks injected as node replacements.
  • prompt.test.ts restored from upstream (only 2 fork-only tests, re-added).
    Loop suites are fork-only, so they were ported to the upstream harness shape.
• 3.5 Run the suite — not just typecheck — and confirm no test silently weakened

Phase 4: Regenerate the SDK

• 4.1 Establish why sdk.gen.ts references symbols absent from types.gen.ts
  (both generated, both fork-modified — determine which is stale)
• 4.2 Regenerate via script/generate.ts; confirm the fork's llama-skein client
  additions survive regeneration
• 4.3 Confirm downstream consumers typecheck (packages/plugin, packages/tui)
• 4.4 Coordinate with the in-flight llama-skein client work before landing
  • Regenerating exposed that the fork's loop.updated and side-question.response
    events had been dropped from the event manifest by the merge, so no client could
    subscribe. Re-registered in packages/opencode's manifest.

Phase 5: Make the gate real

• 5.1 bun run typecheck passes clean at the root
• 5.2 Extend script/fork-verify.ts to fail when a file differs from upstream
  without a fork/manifest.json entry — drift becomes a build error, not a discovery
• 5.3 Document the feature-vs-drift rule in FORK_WORKFLOW.md
• 5.4 Add the classification step to the sync runbook so the next sync starts here

Phase 6: Test-suite stability (discovered during Phase 3)

• 6.1 Establish why full-suite runs die with panic(main thread): Segmentation fault — a Bun bug (panic(main thread): Segmentation fault at address 0x[...] oven-sh/bun#20643, [FEATURE]: Sliding skills context — keep skills adjacent to the active turn for consistent agent attention anomalyco/opencode#23684) triggered by bun test
  spawning child processes; 1.3.14 is latest, canary is older, so no upgrade
• 6.2 Measure the alternatives: --parallel default = one worker per core at
  1.1-2.3 GB each (~26 GB on a 15-core box, drove it into swap); --parallel=4
  = 7 GB and slower than serial; --isolate = past 6 GB and far slower
• 6.3 Adopt sharding (script/test-shards.ts, 4 fresh processes) — no crash
  in any shard, memory bounded, ~60s per shard
• 6.4 File the Bun issue with the repro (suite spawns processes via
  CrossSpawnSpawner, pty and mDNS tests, TestLLMServer)
• 6.5 Consider one shard per CI job (SHARD_INDEX/SHARD_TOTAL) so the 4-vCPU
  runners parallelise instead of running four shards in series

Unchecked items above: see Disposition in proposal.md (2026-09-18).

Plan changes

26 done