fix: concise quality-gate rejection message (issue #396)

[ranxianglei]

Problem (issue #396)

When the pre-commit quality gate rejects a compression, the error returned to the model embedded the complete HOW_TO_COMPRESS_RULES (~4,936 chars ≈ ~1,234 tokens) plus a multi-line "⚠️ CRITICAL" warning paragraph — on every rejection. For a failed compression whose purpose is to save context, this re-injects ~5.4K chars of new text into the model's context. The content is also 100% redundant: the same constant is already in the ACP system prompt (lib/prompts/system.ts:58), sent with every request.

Secondary defects found while fixing:

• The gate's reason field ("Summary too short: …" / "Content coverage too low: …") was never included, even though the issue's expected behavior asks for the failure reason.
• The retry text claimed "Without acknowledgeRisk: true, the compression will be rejected again" — false under compress acknowledgeRisk failed #301/fix(compress): make preemptive acknowledgeRisk a no-op instead of an error [stable] #303 semantics: a rewritten summary that passes the gate is accepted without the flag; the flag only bypasses after a prior rejection set qualityGateRetryPending.

Fix (lib/compress/quality-gate/rejection.ts)

• Removed the HOW_TO_COMPRESS_RULES import/embedding and the CRITICAL warning paragraph.
• Added a Reason: line from result.reason.
• Rewrote retry guidance into three sentences: rewrite a more complete summary and retry (gate re-evaluates automatically); full rules already live in the system prompt; acknowledgeRisk: true bypasses this rejection once if you're confident the summary is correct despite the metrics.

Result: standard rejection payload drops from ~5.4K chars (~1.4K tokens) to ~600 chars (~150 tokens), ~85% smaller.

Not changed: the blocking gate itself, its config/metrics, the throw/bypass mechanics in lib/compress/range.ts, and the E2E fake-LLM marker strings (COMPRESSION REJECTED / QUALITY GATE FAILURE remain in the header — scripts/e2e/fake-llm-server.ts:222,503). Advisory mode / auto-bypass after N rejections stays out of scope (open design discussion in #339).

Tests (tests/quality-gate-enforcement.test.ts)

• Extracted shared fixture buildRejectionFixture().
• Old assertion "should include compress rules" → inverted to absence assertions (HOW TO COMPRESS / KEEP VERBATIM / CRITICAL must NOT appear) + reason inclusion asserted.
• New size-budget regression guard: message < 1,000 chars.
• Integration tests (real-tool rejection, acknowledgeRisk bypass, compress acknowledgeRisk failed #301 preemptive no-op, flag cleared on success) unchanged and green.

Full suite: 1,133 pass / 0 fail; typecheck + build green. New/changed lines kept prettier-conformant (repo baseline has 430 pre-existing flagged files — no unrelated reformatting mixed in).

Devlog: devlog/2026-09-15_concise-quality-rejection/ (REQ + WORKLOG).

Fixes #396

---

中文摘要：质量门拒绝压缩时返回给模型的信息不再重复注入完整的 HOW TO COMPRESS 规则（该内容已在系统提示中），改为「失败原因 + 关键指标 + 简短重试指引」，单次拒绝的注入量从约 5.4K 字符降到约 600 字符；同时修正了 acknowledgeRisk 说明与实际行为不符的问题。可以合并。

[github-actions]

📦 Built Plugin Artifact

Branch: 2026-09-15_concise-quality-rejection (6f1492e)

Option A — Install from npm PR tag (recommended)

opencode plugin opencode-acp@pr-397 --global

Each push to this PR publishes a new version under the pr-397 npm tag.

Option B — Install from GitHub

opencode plugin "github:ranxianglei/opencode-acp#2026-09-15_concise-quality-rejection" --global

Option C — Download artifact

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

tar xzf opencode-acp-pr397.tgz
cp -r package/dist ~/.cache/opencode/packages/opencode-acp@latest/node_modules/opencode-acp/dist

3. Restart opencode to pick up changes.

---

This comment is automatically updated on each push.

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 Verification complete for PR #397 (2026-09-15_concise-quality-rejection @ 3d8688f). I pulled the branch and checked every claim in the description against the code — all hold up:

Diff review — clean, exactly as described. 4 files: lib/compress/quality-gate/rejection.ts (+6/-12), tests/quality-gate-enforcement.test.ts (+29/-5), devlog REQ/WORKLOG. No unrelated changes, no version bump, diff contains nothing beyond the stated purpose.

Verified against code:

• HOW_TO_COMPRESS_RULES import and the CRITICAL warning paragraph are gone from rejection.ts; retry guidance is now three sentences (rewrite+retry / rules already in system prompt / one-shot acknowledgeRisk bypass).
• The new Reason: line is always meaningful: context-compress-algorithms@1.3.0 populates result.reason on both failure paths (L1 "Summary too short: …", L2 "Content coverage too low: …"). The || "unknown" fallback only triggers for a hypothetical custom gate that omits the optional field (types.ts:29).
• The corrected acknowledgeRisk wording matches real behavior: lib/compress/range.ts:333 — bypass requires acknowledgeRisk && state.qualityGateRetryPending, and the flag is set pending at range.ts:350 on rejection. So "rewrite and retry without the flag → gate re-evaluates automatically" and "flag bypasses this rejection once" are both accurate; the old "will be rejected again" text was indeed wrong under compress acknowledgeRisk failed #301/fix(compress): make preemptive acknowledgeRisk a no-op instead of an error [stable] #303 semantics.
• E2E marker strings preserved: header still reads ⚠️ COMPRESSION REJECTED — QUALITY GATE FAILURE, and scripts/e2e/fake-llm-server.ts:222,503 are untouched.

Local runs on the PR head:

• npm run typecheck ✅
• Full suite: 1,133 pass / 0 fail (matches the claimed number)
• tests/quality-gate-enforcement.test.ts: 15/15 (new absence assertions + <1000 char size guard included)
• Measured an actual rejection payload end-to-end through buildQualityRejectionError: 672 chars with a real gate reason — consistent with the "~600 chars, ~85% smaller" claim.

CI status (head commit): pr-validation, build, build-artifact, test (22), test (24) all ✅; e2e still in progress (Docker suite). Given the marker strings are byte-identical where the fake LLM scans for them, e2e failure would be surprising, but I'd wait for it.

One note: the branch base (5135dfd) is 69 commits behind current master (the v1.17.x/v1.18.0 releases landed since). Neither of the two touched source/test files was modified on master in that window, so rebasing or merging will be conflict-free — but consider a quick rebase before merging for a tidy history.

Conclusion: verified, no defects found. Ready to merge once e2e goes green. Per project rules I won't merge it myself — please click merge when e2e finishes.

---

中文摘要：已拉取分支逐条核对——拒绝消息不再重注入压缩规则（实测 672 字符）、Reason 行在两条失败路径下均有内容、acknowledgeRisk 说明与 range.ts 实际语义一致、E2E 标记字符串保留、本地全量测试 1133/0 通过；CI 除 e2e（运行中）外全绿，改动文件与 master 无冲突，e2e 通过后可合并。

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 Update: the last outstanding check e2e just finished green. All 6 CI checks on head 3d8688f are now passing (pr-validation, build, build-artifact, test (22), test (24), e2e).

PR #397 is fully verified — code, tests, and CI. Ready for you to merge whenever.

中文摘要：e2e 检查已通过，全部 6 项 CI 全绿，PR 验证完毕，可以合并。

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

Review & CI status for this PR:

• Dual-agent review (AGENTS.md §5.3 code + §5.6 tests): both APPROVE. Test reviewer ran a mutation check (swapped in master's rejection.ts) — all 3 new/changed tests fail against old code as expected, pass after restore. Both NITs fixed in follow-up commits (6dd1860, a242578).
• CI: 6/6 green on 6f1492e — pr-validation, build, build-artifact, test (22), test (24), e2e. The e2e job exercises the rejection→retry path via the fake-LLM server, confirming the preserved marker strings keep detection working.
• Final payload size: 620 chars vs ~5,811 pre-fix (~89% smaller); full suite 1,133/1,133.

Ready for human merge per §5.1.1.2.

中文摘要：质量门拒绝信息去掉了与系统提示重复的全量压缩规则注入（~1.2K tokens），改为失败原因+关键指标+三句重试指引，单次拒绝注入量从 ~5.8K 降到 ~620 字符；双 agent 评审通过（含变异验证）、CI 全绿（含 e2e），可以合并。