fix: warn once per distinct config-validation error set (#346)

[ranxianglei]

Closes #346

Problem

processTurn unconditionally called console.warn on every turn whenever validateConfig returned errors. With nudge.minContextLimitPct > nudge.maxContextLimitPct (a valid soft-target setup exposed by billion-context — ranxianglei/billion-context#1122), that is one log line per upstream request in proxy deployments: unbounded noise for a configuration that works fine.

Verified: the two limits gate independent code paths in decideNudge (max → pressure zone at src/compress.ts:1222; min → first-sight mass bypass + tier-count usage floor at src/compress.ts:1291,1304). Nothing depends on [min, max) being non-empty, so thresholds do fire correctly — the warning is informational only.

Fix

Deduplicate per session/state, as suggested in the issue: remember the last emitted validation-error set on CompressionState.lastConfigWarnings (new optional field) and re-warn only when the set actually changes (config edit). Warning text unchanged.

• src/compress.ts — processTurn compares the new error set against the stored one (sameWarningSet, order-insensitive) before emitting; writes the updated set back to state every turn ([] when valid).
• src/types.ts — new optional field CompressionState.lastConfigWarnings?: string[].
• src/state.ts — createInitialState() initializes it to [].
• src/sync.ts, cloneState in src/compress.ts, mergeCompressionState in src/persist/state-merge.ts — all three explicit state constructors carry the field, so dedup memory survives both the per-turn sync clone and host persist/load cycles (the exact proxy scenario from validation: minContextLimitPct > maxContextLimitPct warns once per processTurn (log spam) #346). The sync clone was dropping it initially — caught by the new round-trip test.

Behavior change (old → new)

• Before: one console.warn per processTurn call while any config-validation error exists (log spam proportional to request volume).
• After: at most one console.warn per change of the validation-error set — once at first detection, again if the config is edited into a different invalid shape, and again across an invalid→valid→invalid transition. Silent for valid configs.
• New optional state field: old persisted states load fine (field absent → first invalid turn warns once, same as before).
• Caveat: hosts that discard/rebuild state from scratch every turn will still see one warning per turn — inherent to fully stateless usage; matches the "per session/state" semantics proposed in the issue.

Verification

• 5 new tests in tests/validation.test.ts: warn-once across repeated turns (the validation: minContextLimitPct > maxContextLimitPct warns once per processTurn (log spam) #346 repro), re-warn on error-set change, re-warn after valid→invalid transition, silent + empty-set storage for valid configs, dedup memory surviving a JSON round-trip (host persist/load).
• Full suite: 771/771 pass; tsc --noEmit clean; npm run build OK.
• Note: npm run format:check fails repo-wide without this change too (no prettier config committed; ~107 files flagged incl. untouched ones like tsconfig.json). This diff follows the surrounding code style.

[github-actions]

📦 Built Package Artifact

Branch: 2026-09-22_dedup-config-warn-once (7a6f88e)

Option A — Install from npm PR tag (recommended)

In your adapter project:

npm install acp-kernel@pr-347

Each push to this PR publishes a new version under the pr-347 npm tag.

Option B — Download artifact

1. Download the artifact from the Actions run
2. Extract the tarball and install:

tar xzf acp-kernel-pr347.tgz
npm install ./package

---

This comment is automatically updated on each push.

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 Verified PR #347 (fix: warn once per distinct config-validation error set, closes #346). I did not redo the work — I pulled the branch, reviewed the code, checked CI, and fixed one blocker that was preventing merge.

Code review — passes. Implementation is correct, minimal, well-targeted:

• src/compress.ts:459 — processTurn warns only when configErrors.length > 0 && !sameWarningSet(input.state.lastConfigWarnings, configErrors), then writes the current set back every turn ([] when valid).
• sameWarningSet (src/compress.ts:1336) — order-insensitive compare, handles undefined (pre-dedup states).
• All four state constructors carry the new optional field — createInitialState (src/state.ts:19), cloneState (src/compress.ts:1665), syncBlocks (src/sync.ts:38), mergeCompressionState (src/persist/state-merge.ts:26) — so dedup memory survives both the per-turn sync clone and host persist/load (the exact validation: minContextLimitPct > maxContextLimitPct warns once per processTurn (log spam) #346 proxy scenario).
• 5 regression tests in tests/validation.test.ts: warn-once (validation: minContextLimitPct > maxContextLimitPct warns once per processTurn (log spam) #346 repro), re-warn-on-set-change, valid→invalid→valid, silent-valid, JSON round-trip. Diff is clean (+143/-1, 6 files), no unrelated churn.

Blocker found & fixed: stale-base merge conflict. The PR branched from an older master; master had since advanced through releases v0.0.79–v0.0.82 and the acp_rule persistent-rules feature, which added terminalStreak/rules/nextRuleId to the same state constructors this PR touches. mergeable_state was dirty. I rebased onto current master (b30723f) and resolved all 5 conflicts as a pure union of field additions (kept both sides — no semantic overlap).

Verification after rebase (head fe70885):

• Local: tsc --noEmit clean · full suite 874/874 pass · build OK.
• Pushed to the PR branch; GitHub CI green: test (22), test (24), pr-validation, build-artifact — all success.
• PR is now mergeable: true; mergeable_state: blocked = only the required-review/approval gate remains.

Not merging — merge is human-only per repo rules. Please approve + merge: #347

Notes:

• Behavior change (already disclosed in PR body): one console.warn per change of the validation-error set instead of per turn; silent for valid configs; old persisted states load fine (field absent → first invalid turn warns once). Fully-stateless hosts still see one warn/turn (inherent).
• npm run format:check failing repo-wide is pre-existing (no prettier config committed); my conflict resolution matched surrounding style, adds no new formatting delta.

中文摘要:修复了 PR #347 因落后于 master(期间合入 v0.0.79–v0.0.82 与 acp_rule 功能)导致的合并冲突——已 rebase 到最新 master 并按字段并集解决 5 处冲突;本地 typecheck + 874 项测试 + 构建全绿,GitHub CI(test 22/24、pr-validation、build-artifact)也全绿。代码实现正确、测试充分,diff 干净,现只差人工审批,可以合并。

[ranxianglei]

@ework-daemon 与 master 冲突了,请 rebase 到当前 master 后重推,CI 绿后待合。

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 已完成 rebase + 重推,CI 全绿,待合。

做了什么:

1. master 又前进了(我上次 rebase 的基点 v0.0.82 → 现在已到 v0.0.89,期间合入 compress 端点落在被 orphan-hiding 隐藏的失败 compress 调用上时整条区间被拒为 cannot be anchored（应吸附到相邻可见消息） #396/fix: snap compress range endpoints onto hidden orphan calls (#396) #397/fix: preserve non-text blocks in Anthropic tool_result via sidecar #367/fix(persist): fence async commits against late-landing stale writes (#373) #409/fix: drop top-level anyOf from COMPRESS_PARAMETERS (Anthropic 400, #404) #405 等,新增了 imageFullRestored/imageShrinks/hiddenOrphanRefs 状态字段和 processTurn 里的 contentStore)→ 确实产生了新冲突。
2. 把 PR 分支 rebase 到当前 master(615347a),4 处冲突全部按字段并集解决(两边各自新增的字段都保留,无语义重叠):src/state.ts、src/persist/state-merge.ts、src/types.ts、src/compress.ts(processTurn 里同时保留 lastConfigWarnings 写入与 contentStore;cloneState 自动合并且仍带该字段)。
3. 重推:fe70885 → d00e96b(force-with-lease)。

验证:

• 本地 pre-flight(rebased head d00e96b):tsc --noEmit 干净 · 全量测试 1086/1086 通过 · build OK。diff 相对 master 仍是干净的 +143/-1(6 文件),未混入 master 新代码。
• GitHub CI 全绿:test (22) ✅ · test (24) ✅ · build-artifact ✅ · pr-validation ✅。
• 分支已与当前 master 对齐(head 的父提交即 master 最新 tip 615347a,mergeable: true)。GitHub 页面若短暂显示 "behind" 是 force-push 后的异步计算延迟,git 侧已确认最新。

待合: 只差人工审批/合并(合并为人工操作,我不代合)。链接: #347

中文摘要:已将 PR #347 rebase 到当前 master(v0.0.89)并按字段并集解决 4 处冲突后重推(d00e96b),本地 typecheck+1086 测试+构建全绿,GitHub CI(test 22/24、build-artifact、pr-validation)也全绿,分支已对齐最新 master,可以合并。

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 说明:本条 rebase 请求发出约 40 分钟后,本 PR 的专属会话仍无任何进展(远端分支未动、无新提交),由 #412 审计会话代为执行 rebase。若本 PR 会话随后恢复运行,看到已推的 rebase 结果即可停止,无需重复。rebase 完成后我会在此回报(含门禁结果)。

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 Rebase 完成,分支已更新,可合。

• 新 head:7a6f88e(原 d00e96b → rebase 到 master b86c064,含 fix: unify tool-message accounting across status report surfaces (issue #390) #177/test(persist): retarget missed fsp.rename mock from #409 (#410) #411 合并)。
• 冲突 4 处,全部为同模式并集解(master 侧新增的 state 字段 vs 本 PR 新增字段,互不覆盖):
  • src/compress.ts processTurn():lastConfigWarnings = configErrors.length > 0 ? [...configErrors] : [](PR)+ const contentStore = input.contentStore ?? createContentStore();(master)
  • src/persist/state-merge.ts:master 三个 image 字段 + PR lastConfigWarnings: parsed.lastConfigWarnings ?? fresh.lastConfigWarnings
  • src/state.ts fresh-state:imageFullRestored/imageShrinks/lastConfigWarnings 均初始化
  • src/types.ts CompressionState:两侧可选字段全保留(含各自 doc comment)
  • src/sync.ts、tests/validation.test.ts 自动合并干净(master 未动过该测试)
• diffstat 与原 PR 完全一致(+143/−1),无内容丢失。
• 门禁(本地真实 Node):typecheck 干净 ✅ / 1093/1093 tests ✅ / build 成功 ✅。
• 行为语义与 PR 描述一致:warn-once 去重按「错误集变化」触发,不触压缩决策/wire/state schema 契约(state 仅加一个 optional 字段,旧持久化状态缺省安全)。

CI 将在 force-push 后重跑;合并请人工执行(PR merges are human-only)。

一句话总结:#347 已 rebase 到当前 master 并四处并集解冲突,本地三门禁全绿(1093/1093),可以合并。