AI-assisted — drafted and iterated with Claude Code (claude-sonnet-4-6). Lightly tested: unit tests pass, pnpm tsgo clean, pnpm check passes; timeout-retry path validated by code review; not exercised against a live timeout event. I have personally reviewed all code changes prior to this commit and take responsibility for their correctness.

Summary

• Problem: compaction runs on the session's active thinking level, so sessions using slow/high-think models (e.g. Opus + thinking:high) can take 4+ minutes to compact — exceeding channel timeout windows (Discord 30s, Telegram 240s) and causing a stale-response retry loop (Compaction timeout races against channel timeout, causing stale-response loop #25272).
• Why it matters: affects any user running a thinking-enabled model on a chat surface with a strict reply timeout; the loop is deterministic and requires manual intervention to break.
• What changed: compaction now runs with thinking=off by default regardless of session model. New agents.defaults.compaction.thinking key — "off" (default) or "on" (inherit the session model's current thinking level). When a compaction run times out with thinking enabled, the runner retries once without thinking automatically. The before_compaction hook is guarded to fire exactly once per compaction event.
• What did NOT change: primary session model and thinking level are unchanged; compaction prompt content, threshold logic, and retry limits in run.ts are unchanged; no new network calls or auth paths.

Schema

agents.defaults.compaction.thinking: "off" (default) | "on"

• "off": compaction always runs without extended thinking regardless of the session model — preventing timeout races on channels with strict reply windows.
• "on": compaction inherits the session model's current thinking level. On timeout, automatically retries once without thinking.

Renders as a two-button segmented control in the config UI.

UI — Before / After

Before (upstream main — no Compaction Thinking field):

After — new Compaction Thinking field with off / on segmented control:

Change Type

• Feature
• Docs

Scope

• Gateway / orchestration
• API / contracts
• UI / DX

Linked Issue/PR

• Closes Compaction timeout races against channel timeout, causing stale-response loop #25272
• Related Feature: Allow compaction model override (use cheaper model for summarization) #7926 [Feature]: Compaction: configurable model override + flush-then-reset mode #30452
• Discussion: #33430
• Companion: feat(compaction): add fallbackModel to retry compaction on quota/rate-limit errors #33396 (superseded by new companion PR for compaction.fallbackModel)

User-visible / Behavior Changes

• Compaction now always runs with thinking: off by default. Existing sessions unaffected — this only changes the thinking level used during the compaction summarization turn.
• New config key agents.defaults.compaction.thinking (opt-in; defaults to "off").

Security Impact

• New permissions/capabilities? No
• Secrets/tokens handling changed? No
• New/changed network calls? No
• Command/tool execution surface changed? No
• Data access scope changed? No

Repro + Verification

1. Set a thinking-enabled model (e.g. Opus + thinking: high)
2. Trigger compaction via context threshold
3. Confirm compaction completes without thinking ([compaction-diag] log shows thinking=off)
4. Set agents.defaults.compaction.thinking: "on", repeat — confirm thinking is used and timeout-retry path activates if needed

Evidence

• 5 unit tests covering all resolveCompactionThinkLevel branches (absent, "off", "on" with various session levels)
• Diagnostic log fields: thinking=${effectiveThinkLevel} thinkAttempt=${thinkAttempt} on every [compaction-diag] start/end line

Human Verification

• Verified: unit tests pass, pnpm tsgo clean, pnpm check passes
• Edge cases: absent config, "off" with thinking session, "on" with non-thinking session, hookFired guard
• Not verified: live timeout race against a real channel deadline

Compatibility / Migration

• Backward compatible: thinking is optional; default is "off" (no change for existing sessions without explicit opt-in)
• To opt in: agents.defaults.compaction.thinking: "on"

Failure Recovery

• To revert: remove agents.defaults.compaction.thinking from config (or set to "off")
• Bad symptom: compaction always uses thinking regardless of config — check resolveCompactionThinkLevel return value in logs

Risks and Mitigations

• Risk: users with thinking: "on" who expect fast compaction may see slower compaction on first attempt before timeout-retry kicks in.
  • Mitigation: "on" is opt-in; default is "off" which is strictly faster.