perf: align nudge/system prompts with pi-side direct-action discipline

[ranxianglei]

Problem (#436)

With compress.candidates enabled, the standing prompt and nudge guidance steered the model into a speculative acp_status pre-check turn before compressing listed candidates ("use acp_status for a fresh view", "call acp_status first"). On local models each extra round trip costs 15–23s, and the extra small turn amplifies prompt-prefix invalidation after compression.

Fix (prompt-only, aligned with billion-context-pi / acp-kernel discipline)

The candidate list is already authoritative: planCompressionCandidates() validates every MICRO/EPISODE entry against the real range executor at nudge construction (lib/messages/inject/candidates.ts) — stronger than the kernel's static suggestions. So only the prompt discipline needed porting; no arming/state/tool changes were required.

• lib/prompts/context-limit-nudge.ts — CANDIDATE_GUIDANCE now states candidates were validated when the nudge was built; compress directly; on a stale-ID failure run acp_status once and re-issue using only reported refs.
• lib/prompts/system.ts — candidate bullet de-hedged ("do not call acp_status first"); new failure-recovery bullet (no arithmetic range adjustment; re-issue same turn, batched); removed the MULTI-TIER "call acp_status first" paragraph.
• lib/prompts/compress-range.ts — CANDIDATE GUIDANCE tail: direct-action + post-failure recovery instead of "use acp_status when stale".
• acp_status remains fully functional as an optional refresh/diagnostic path; stale-plan and structure-version validation untouched.

Not ported verbatim from the kernel: "every successful compress renumbers the remaining refs" — opencode-acp refs are sticky per raw message ID (lib/message-ids.ts), so the recovery line is phrased generically.

Tests

• tests/prompts.test.ts, tests/nudge-text.test.ts: constant-level regression tests pinning the new wording and asserting removal of the old hedging phrases.
• tests/compression-candidates.test.ts: multi-turn injectCompressNudges regression test per AGENTS.md §5.7.1 — two consecutive calls sharing one SessionState, asserts shouldInjectThisTurn + lastPerMessageNudgeTokens/lastNudgeShownTokens side effects after each turn, rendered nudge carries the new guidance without old hedging, production-style preserveRecentMessages: 20. Verified to exercise the max-limit anchor path (guidance renders only once usage crosses maxContextLimit).

Verification: typecheck ✓, prettier (changed files) ✓, build ✓, full suite 1270/1272 — both failures are environmental in the review sandbox (read-only /tmp; tests/soft-block.test.ts:14 mkdir and tests/inactive-block-decompress.test.ts:206 hardcoded /tmp toFile path) and reproduce independently of this diff.

No version bump (release branches only).

Closes #436

[github-actions]

📦 Built Plugin Artifact

Branch: 2026-09-18_nudge-prompt-align-bcp (1501b2c)

Option A — Install from npm PR tag (recommended)

opencode plugin opencode-acp@pr-439 --global

Each push to this PR publishes a new version under the pr-439 npm tag.

Option B — Install from GitHub

opencode plugin "github:ranxianglei/opencode-acp#2026-09-18_nudge-prompt-align-bcp" --global

Option C — Download artifact

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

tar xzf opencode-acp-pr439.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

PR #439 verification — code confirmed, CI green, ready for human merge.

Treated this as already-completed work (per the ework-agent-pr marker): pulled the branch, verified the diff + CI rather than redoing it. No changes made — CI is green and nothing needed fixing.

What the change does (prompt-only; no logic/state/tool changes):

• Removes the standing instruction to run acp_status before compressing listed candidates. The premise holds up: planCompressionCandidates() (lib/messages/inject/candidates.ts:466) validates every surviving MICRO/EPISODE entry against the real executor (prepareExecutableRangePlans) at nudge-build time, gated by a selectionMatchesSource() drift check (candidates.ts:429) → omitted as executor-selection-drift/executor-rejected otherwise. So the list is authoritative when built, and the pre-check turn was pure overhead (15–23s/round-trip on local models + extra prefix invalidation).
• Adds same-turn failure-recovery wording: stale ref → run acp_status once, re-issue with only reported refs, batched; no arithmetic range math.
• acp_status stays fully functional as an optional diagnostic; stale-plan / structure-version validation untouched.

Files: lib/prompts/context-limit-nudge.ts, lib/prompts/system.ts, lib/prompts/compress-range.ts (+ devlog REQ/WORKLOG, 3 test files). I swept every remaining acp_status mention in lib/prompts/ — none still tell the model to pre-check before compressing; the rest are tool-definition lines, a historical-ref safety note (system.ts:63, about past compress calls), or optional "for details" pointers.

Independent verification I ran:

• typecheck ✓
• Full suite 1270/1272 — the 2 failures are environmental and reproduce without this diff: tests/soft-block.test.ts (EACCES … mkdir '/tmp/…') and tests/inactive-block-decompress.test.ts (hardcoded /tmp toFile path rejected by the allowed-paths guard). Neither file is touched by this PR.
• Targeted tests/prompts.test.ts + tests/nudge-text.test.ts + tests/compression-candidates.test.ts: 35/35 pass, including the 4 new wording-pinning regression tests (direct-action bullet, same-turn failure recovery, old-hedging absence).
• CI on head 1501b2c0: e2e ✓ build ✓ build-artifact ✓ test(22) ✓ test(24) ✓ pr-validation ✓ — all green.
• No version bump — correct per §5.1.1.1 (release branches only); pr-validation confirms branch name + devlog presence.

Two things to note before merge:

1. Branch is 9 commits behind master (06efd39c → 026dbe73). GitHub reports mergeable: true (no conflicts) and I confirmed zero file overlap between this PR and those 9 commits (they touch index.ts, lib/hooks.ts, lib/messages/*, and unrelated tests). A quick rebase onto current master gives clean history + fresh CI on the latest base; merging as-is is also safe.
2. No formal reviews recorded yet. §5.3 requires ≥2 independent agent reviews for lib/ changes. This pass counts as one; a second independent review is still outstanding before a human merges.

Bottom line: code is correct and tightly scoped, premise verified, CI fully green. Ready for a human to merge (rebase-then-merge recommended); the merge itself stays human-only per policy.

---

中文摘要：核对已完成的 PR #439(纯提示词改动——让模型直接压缩"构建时已用真实执行器校验过"的候选、不再先调 acp_status,失败才在同轮兜底重试)。逐条验证了 diff 与描述一致、核心前提成立(candidates.ts 确实在 nudge 构建时对每个候选跑真实执行器并做 drift 校验)、typecheck 通过、全量测试 1270/1272(仅 2 个与本 diff 无关的只读 /tmp 环境失败)、CI 六项全绿。分支落后 master 9 个提交但与本 PR 零文件重叠、可无冲突合并(建议 rebase 后合并);仅剩 §5.3 要求的第二次独立 agent review 尚未记录,合并操作按规范由人工执行。