feat: strip reasoning from protected-exempt historical messages (gated)

[ranxianglei]

Problem

On long sessions, reasoning parts that ride on compress/skill-carrying assistant messages form a permanently-incompressible context floor. Protection is message-granular (lib/compress/protected-content.ts:202), so the whole message — reasoning included — is excluded from every compression and re-sent on every turn. Each compression round adds ~9 KB of unreclaimable reasoning; measured ~83.5% of the never-covered residual in a real long session (#368). It is a monotonic feedback loop that degrades long-session usability.

Solution

A single request-time pass, stripProtectedReasoning, that drops the reasoning parts from protected-exempt messages in closed historical turns before the request is sent. The current (possibly-open) round is always preserved. No persisted-state or internal-tag changes; no DB writes.

Five gates (request-level gates 4–5 run first; all must hold to strip)

1. Provider allowlist (added by review, 2026-09-08) — compress.stripProtectedReasoningProviders (default ["anthropic","gemini"], "*" = all, case-insensitive substring). Fail-closed: unknown or undefined provider → nothing is stripped. Rationale: GPT-family gateways may reject requests whose historical thinking blocks are incomplete.
2. Session activation (added by review) — compress.stripProtectedReasoningMinMessages (default 100, 0 = always). Below it the pass is a byte-stable no-op, so short sessions never pay prefix-cache churn.
3. Turn-closure — message index strictly < lastUserIndex (index of getLastUserMessage). The active round is never touched.
4. Selector — the message contains a tool part whose part.tool ∈ config.compress.protectedTools (compress/skill).
5. Size — total reasoning length > threshold (default 0 — review found per-message size is cache-noise: invalidation propagates from the first stripped message, so the activation gate is the real cache lever).

Action: msg.parts = parts.filter(p => p.type !== "reasoning") — drops reasoning only; the tool call + other parts are preserved (the live summary lives in the compress-call body, not the reasoning).

Placement

Inserted after hideConsumedCompressCalls and before assignMessageRefs in lib/hooks.ts — operates on the minimal surviving set (reasoning-only leftovers are already spliced there, since reasoning is "structural" per lib/compress/parts.ts:1).

Config (additive, safe defaults)

• compress.stripProtectedReasoning (bool, default true) — kill-switch.
• compress.stripProtectedReasoningThreshold (number, default 0) — size threshold in chars.
• compress.stripProtectedReasoningProviders (string[], default ["anthropic","gemini"]; "*" = all; explicit [] = strip for no provider) — added by review.
• compress.stripProtectedReasoningMinMessages (number, default 100; 0 = always) — added by review.
• All registered in VALID_CONFIG_KEYS + validateConfigTypes; excluded from CompressOverridableConfig (global-only).

Files

• lib/messages/reasoning-strip.ts — the pass, now with 5 gates + optional options param (omitted = ungated for pure-function callers).
• lib/hooks.ts — passes state.modelProviderID + config gates into the call.
• lib/config.ts — 4 config keys (interface + defaults + merge + CompressOverridableConfig exclusion).
• lib/config-validation.ts — VALID_CONFIG_KEYS + validateConfigTypes (incl. string[] element validation).
• dcp.schema.json — schema properties + defaults.
• tests/reasoning-strip.test.ts — 18 pass tests + 5 merge tests + 14 gate tests.
• tests/e2e-message-transform.test.ts — hook-level kill-switch, provider-gate (fail-closed), and activation-gate tests.
• devlog/2026-09-07_strip-protected-reasoning/ — REQ / DESIGN / WORKLOG (review updates recorded).

Testing

• 1124/1124 tests pass (was 1096 at PR open; +28 across the review rounds). Typecheck clean, build clean.
• Mutation-verified (§5.7.3): disabling the provider gate → 5 tests fail; disabling the activation gate → 2 tests fail; dropping the hook options object → 2 e2e tests fail; dropping the hook ?? fallbacks → 1 e2e test fails (each verified by temporarily mutating the code, then restoring).

Review

• Original implementation: dual-agent review, both REQUEST-CHANGES; all findings addressed.
• 2026-09-08 review session (issue Protected-message exemption inflates the incompressible context floor: per-compression reasoning is never reclaimable (~83.5% of measured residual) #368 re-review of this PR): recommended the provider allowlist + activation gate + threshold-0 changes now in this PR; owner approved direct application ("直接修改pr").
• Second dual-agent review of the additions (source + test focus): both APPROVE, all MINOR/NIT findings addressed — request-level provider resolution (requestModel?.providerID ?? state.modelProviderID, fixes undefined-provider on a fresh session's first request), allowlist entry trimming, 9 new config-validation tests (incl. fractional minMessages now rejected — Number.isInteger), hook-fallback e2e test (mutation-verified), tight boundary + presence assertions. One pre-existing follow-up noted: e2e buildConfig is structurally incomplete (tests are excluded from typecheck).

Secondary findings (filed as separate issues)

• A — display/acp_status estimator excludes reasoning (lib/messages/inject/utils.ts:586) → Display/acp_status context-usage estimator excludes reasoning tokens (undercounts real usage) #371.
• B — orphaned byMessageId entries with emptied activeBlockIds stay visible (lib/messages/prune.ts:60-66) → Orphaned byMessageId entries with emptied activeBlockIds stay visible forever #372.
• C — rewriteCompressInput full-consumption leak (lib/compress/hide-consumed.ts:42) → rewriteCompressInput full-consumption leak: compress part left intact when all blocks consumed #373.

Fixes #368

[github-actions]

📦 Built Plugin Artifact

Branch: 2026-09-07_strip-protected-reasoning (70c21ba)

Option A — Install from npm PR tag (recommended)

opencode plugin opencode-acp@pr-370 --global

Each push to this PR publishes a new version under the pr-370 npm tag.

Option B — Install from GitHub

opencode plugin "github:ranxianglei/opencode-acp#2026-09-07_strip-protected-reasoning" --global

Option C — Download artifact

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

tar xzf opencode-acp-pr370.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]

Superseded: owner redesign — the allowlist/activation-gate approach is replaced by a clean reimplementation: nested compress.reasoning { drop, threshold } config with the existing three-level provider/model cascade (#344), selector narrowed to compress tool calls only, single-thinking-size threshold. New PR incoming (refs #368).