fix: allow upstream attribution in translated readmes

[markijbema]

What

Make the kilocode global-config change-detection baseline (globalStamp) single-owned and seeded eagerly at service construction, in packages/opencode/src/config/config.ts.

Why

globalStamp records the on-disk content of the global config so edits made by other Kilo processes can be detected and hot-reloaded. It had two problems:

1. Two writers. Both loadGlobal and refreshGlobal wrote globalStamp. A cache reload (loadGlobal) could advance the baseline to whatever was on disk at reload time, silently absorbing an external edit that landed between invalidation and the reload — so the next refreshGlobal would compare against the already-advanced baseline and miss the change.
2. Empty-string swallow. The baseline started as "" and the guard if (!globalStamp || …) made the very first refreshGlobal a no-op that didn't record anything, leaving detection dependent on whichever writer ran first.

Combined with Server.Default() being memoized once per process (so the Config service closure — globalStamp + cachedGlobal — is shared), this is a latent flake surface: test/kilocode/global-config-refresh.test.ts ("detects external global config edits") intermittently fails on CI with "ask" instead of "allow", even though it passes in isolation. The test and the affected code are identical on main; this is not specific to any feature branch.

Change

• Seed globalStamp once when the Config service is built.
• Remove both loadGlobal writes of globalStamp.
• refreshGlobal is now the sole owner: it advances the baseline only when on-disk content actually differs, then invalidates the cache so the next read reloads.

A wrong initial seed (e.g. service built before the config dir is finalized) is self-correcting: the first refreshGlobal simply triggers one idempotent invalidate+reload — never stale data.

Honesty note

This is a correctness/hardening change for multi-process config-edit detection and it removes the shared-state vector behind the CI flake. I could not write a regression test that fails on the old code and passes on the new one in a single process — the old single-process logic is self-correcting on every reload, and the flake only manifests via shared-closure leakage across the three tests in the file under the full parallel suite. Rather than add a test that passes on both versions (test theater), I left the existing suite as the guard. All 115 config-related tests pass; global-config-refresh.test.ts was run 8× in a loop with no failures.

Testing

• bun test ./test/kilocode/global-config-refresh.test.ts (8 loops, green)
• bun test across config/permission/global suites: 115 pass
• bun run typecheck: clean
• opencode annotation guard: clean

[kilo-code-bot]

Code Review Summary

Status: No Issues Found | Recommendation: Merge

Files Reviewed (1 files)

• script/check-forbidden-strings.ts

---

_{Reviewed by gpt-5.4-20260305 · Input: 31.5K · Output: 3.5K · Cached: 179.7K}

_{Review guidance: REVIEW.md from base branch main}