feat(compaction): optional compaction.model override + re-enable senpi compaction on SDK-native claude-sdk-oauth lanes

[HeiTuz]

Problem

The claude-sdk-oauth lane keeps a resident Claude Agent SDK session, and the compaction lane policy (lane-policy.ts) stands down senpi compaction entirely on the assumption that the SDK runs its own native compaction. In practice that native compaction never fires for these sessions — senpi's session.log shows threshold/pre_prompt compactions repeatedly rejected with the Claude Agent SDK owns compaction for this session while context grows unbounded (observed past 335k tokens) until overflow, with no recovery path.

Change

Add an optional compaction.model setting ("provider/model", e.g. deepseek/deepseek-chat) that redirects only the compaction summarization call to a different model. When set:

1. Lane policy (lane-policy.ts): the SDK-native stand-down lifts, so senpi auto/threshold compaction runs on the lane again. This is the escape hatch for lanes whose SDK never compacts.
2. Execution (agent-session.ts → _executeCompaction): a new _resolveCompactionModel() resolves the override through the model runtime and uses it for the summarization auth (_getCompactionRequestAuth) + compact() call. Session bookkeeping (lifecycle record, preparation, branch) still tracks the session model, so history/diagnostics are unaffected.
3. Safe fallback: any resolution failure (unset / malformed provider/model / unknown model) falls back to the session model, so compaction can never silently break.

Default behavior is unchanged for every provider — with no override, nothing changes.

Surface

• CompactionSettings.model?: string (settings-manager.ts, compaction/compaction.ts), exposed through getCompactionSettings().
• LaneContext gains an optional getCompactionSettings getter so the lane policy can read the override.

Validation

• New tests: lane-policy stand-down lift with override + continued stand-down when the override resolves to no model; settings model passthrough.
• test/compaction/ suite: 47 files / 339 tests pass.
• Root npm run check (biome + pinned-deps + ts-imports + shrinkwrap + tsc --noEmit): clean.

---

Summary by cubic

Adds an optional compaction.model override that redirects only the compaction summarization call to a specified model and re-enables senpi compaction on claude-sdk-oauth lanes when set. Prevents unbounded context growth when SDK-native compaction never fires, with no change to default behavior when unset.

• New Features
  • compaction.model accepts "provider/model" (e.g., deepseek/deepseek-chat); used only for summarization and falls back to the session model if invalid or unresolved.
  • Lane policy: when an override is set on claude-sdk-oauth lanes, senpi auto/threshold compaction runs again instead of standing down.
  • Setting is exposed via getCompactionSettings() and resolved at execution time without affecting session bookkeeping or diagnostics.

^{Written for commit 95f7dfb. Summary will update on new commits.}

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, add credits to your account and enable them for code reviews in your settings.

[1vivy]

Another production datapoint for the "native compaction never fires" claim, from a lane that ran to the window edge rather than to 335k.

Session: claude-sdk-oauth / claude-opus-5 (contextWindow: 1000000 per pi-ai anthropic.json), resumeMode: "auto", senpi 2026.8.9-2, compaction.enabled: true.

Prompt tokens (input + cacheRead + cacheWrite) per assistant turn, from the session JSONL:

03:37   81,308      06:04  659,972
04:07  212,414      07:13  807,897
04:37  371,067      19:12  938,654
05:19  530,435      19:50  983,261  <- max
                    20:20   75,098  <- after model_change off the lane

983,261 / 1,000,000 = 98.3% of the context window. senpi's adaptive threshold for a 1M window is 0.65 (policy.js: computeAdaptiveThresholdRatio), so compaction was due at ~650k, around 06:00 — it ran another ~14h past that, 290 billed assistant turns, $229.48, stopping ~17k tokens short of hard overflow. Context only dropped when the model was switched off the lane.

Two details that may be worth folding into the fix:

1. Nothing is logged. logs/compaction.log has zero lines for this session id and zero lane-rejection lines. The message_end monitor hook is itself gated (compaction/index.js:650, isMonitorableMessageEvent(event) && !lanePolicy.disablesSenpiCompaction(ctx)), so usage is never tracked on the lane and there is no rejection to log. The failure is silent, not just unrecovered — which makes it hard for users to notice before the bill.

2. Zero boundaries were ever mirrored. No claude-sdk-oauth-compact entries in the session, so the SDK emitted no compact_boundary at any point up to 98.3%. That is the production observation feat(claude-sdk-oauth): compaction alignment, restart continuity, and the permanent gate #637's gate matrix explicitly could not produce (phase (f) was skipped because "the loopback can't make the real CLI emit a compact_boundary").

One question on the shape of this PR: since the stand-down only lifts when compaction.model is set, a default claude-sdk-oauth + resumeMode: auto session still never compacts after this merges. Is a default-side backstop in scope here (e.g. re-arming the usage monitor on the lane, or lifting the stand-down once usage crosses the threshold with no boundary observed), or is that intended as a separate change?