v0.32.0
Added
- In the footer, TLH now shows how much context MCPs are consuming.
Model defaults
- The
developersubagent now defaults togpt-5.6-luna maxfor OpenAI,sonnet-4-6 mediumfor Anthropic. - The Architect, Product, Bug-hunter,
code-reviewer,oracle, andcontrariannow useopus-5.0 highfor Anthropic,gpt-5.6-sol highfor OpenAI. - All 12 bundled agents now declare thinking levels independently per provider instead of sharing a single level.
Removed
- Removed RTK from TLH (see this analysis as to why).
- Removed the pi-intercom dependency. The little we used of it is now folded into pi-subagents. This also reduces how many tokens TLH consumes.
Fixed
- Installing TLH from
mainno longer fails npm peer-dependency resolution; repository tooling now consistently uses the TypeScript version supported bytypescript-eslint. - Subagents now have two timeout caps: a soft one at ~4m30s where the architect checks on them, and a hard one with a bigger timeout that pauses them. They also support steering, and have tighter system prompts. This should make them much more responsive.
- TLH now works a bit better with Herdr.
Other minor things
/tokensnow reports median observed wall-clock latency per tool alongside the existing cost and token data. Latency is the interval between the recorded call and result events and includes any queueing or paused-run time./annotate-last-messagenow sends submitted feedback directly to the agent as a follow-up message instead of appending it to the TLH editor buffer./annotate-git-diffnow sends review feedback to the agent when you click Submit. Closing the review window with unsent comments still appends a recovery draft to the editor instead of sending, so an accidental close cannot trigger an agent turn.- Bumped the bundled Pi to
0.83.0.
Full changelog: https://github.com/diegopetrucci/the-last-harness/blob/v0.32.0/CHANGELOG.md