0.50.0 — the agent tells you what it is doing, and can reach you

Five things in this release were happening without you being able to see them, reach them, or turn them on.

The agent tells you what it is doing

It keeps a task list — one item at a time marked as in progress, finished items struck through — and it is on the screen while the run goes. RunState.tasks had existed for a while as a field that nothing ever wrote to; the list now survives a context compaction, so a long run does not forget its own plan.

It can reach you when nobody is at a console

Set CHIMERA_APPROVAL_WEBHOOK to a channel webhook and an approval question follows you there. Three unattended surfaces — cron jobs among them — were written to be able to ask a person, and the parameter that carried the question was dropped one layer down. They now say unreachable when there is a home directory but no way to deliver, instead of silently deciding for themselves.

You can turn the trust kernel on

CHIMERA_GOVERNANCE ships off on every surface, and the Security screen reported an audit log with nothing anywhere to switch on the thing that writes it. There is now a control, and the screen says which state it is in.

chimera find has a semantic half

Hybrid retrieval — keyword and vector together — adopted by a rule fixed before the run and not after it. Measured on the project's own corpus, hybrid beats keyword by 6.25 points (p = 1.7e-04). Vector search alone loses to keyword, which is why the fusion and not the embedding is what earns its place. The pass announces its bill before it starts.

Runs stop dying on models nobody catalogued

A model outside the nineteen hand-checked entries was assumed to have a 128,000-token window. Of the 431 models the index publishes, 31 serve 64,000 or less — for those, the compaction trigger sat past the wall it exists to avoid, and a context overflow ends a run with no recovery. The window now comes from the live index when the catalogue does not know the model. Five catalogue entries also promised a window their provider does not serve, and one price was wrong by 2.2x while its own note claimed it had been verified.

Also

Traces now record which backend served each step, not just which model answered. A model slug on OpenRouter is a pool — one of them has thirty endpoints whose context windows span 5x and whose prices span 8.8x — and until now nothing in this project could tell two of them apart. Several of its own published measurements turned out to be about pools rather than models, and the affected claim in bench/parallel_tools is retracted in the changelog rather than quietly corrected.

Honest caveats

• The installers are unsigned. First run shows a SmartScreen warning on Windows and a Gatekeeper warning on macOS. That is expected; the updater is signed, which is the part that matters for what lands on your machine after install.
• Governance ships off. The control exists so you can turn it on, not because it is on.
• The compaction summariser ships off, behind AgentConfig.summarise_compaction. Compaction itself has never fired in ordinary use — measured at 0 times across 137 runs — so the summariser is built and unproven rather than built and needed.
• Cancellation is cooperative. Stopping a run stops it before its next model call; calls already in flight finish and are billed.

Full detail, including the measurements that said do not build this, is in the changelog.