A new conversation compaction system that automatically summarizes older messages when the prompt exceeds context limits, using binary search to find the optimal compaction boundary.
Feature-flagged behind tengu_cobalt_raccoon (default: false). Disabled in non-interactive mode.
- Replaces the removed
tengu_sm_compactsystem with a more sophisticated approach - Uses binary search to determine how many message groups to preserve vs. summarize
- Sends older messages to a model fork for summarization
- Includes retry logic (strips media on size errors), empty-response detection, and PreCompact hook support
- Extensive telemetry: trigger, success, duration, pre/post token counts, group counts
- When enabled, triggers automatically during conversation turns when token limits are approached
Feature gate (search for "tengu_cobalt_raccoon") returns !1 (false) as default
Strings lifted out of the shipped bundle, so the claim above can be checked against them.
tengu_cobalt_raccoon On for this account, and not off by defaultThe flag server returned on for the one account this site reads, and nothing in this release compiles it off by default. The compiled default is shown below, and says which it is when we cannot read one: a fifth of gates compile in a string or a number rather than on or off, and most published releases have no gate table behind them at all. No client can see what the server returns for your account.
This account: on · anonymous baseline: on · compiled default in v2.1.105: no gate table built for this version
These values were read against a different version of Claude Code, so treat them as the nearest reading available instead of one taken on this release.
Read once, for one account on one subscription tier, against v2.1.105. It isn't a statement about your account. What a flag value here can and cannot tell you
A reading is one sample. Claude Code evaluates its flags remotely, so no client sees the targeting rule behind a value and this says nothing about your account.
A reading is one sample. Claude Code evaluates its flags remotely, so no client sees the targeting rule behind a value and this says nothing about your account.