Chimera Agent 0.65.0: Pairing, Deadlines, and Sandbox Guarantees

First boot pairs the bot, not the room

Before this release, the desktop app asked "is this a new install?" on the Messaging card when the data folder was already populated. That meant every bot the app started ran open—answering anyone who reached it, using the owner's tools and spend. A pairing code, if one was generated, went to a console nobody reads.

The app now records which bots were already configured at startup. Those bots keep answering as before. A bot configured after that point starts by pairing, and its code appears on the Messaging card where someone actually looks. The API reflects this: GET /api/messaging reports pairing_code, pairing, and open states. The desktop bridge sees pairing but never the code itself, so automation cannot bypass the human step.

A sandbox that cannot run refuses to run

Setting CHIMERA_SANDBOX=docker without a Docker daemon used to fall back to the host with a single log line. A command that should have been isolated ran directly on the machine. That fallback is gone. Now run_shell, execute_code, and workflow steps refuse before they even ask the host-exec question. The refusal message states what happened, that nothing ran, and the two ways out: start Docker, or set CHIMERA_SANDBOX=local. The verifier abstains rather than reverting work because a daemon was down.

A caller that passes fallback= explicitly keeps the old behaviour. Everyone else gets a hard stop. The same boundary applies to inbound media: Telegram and WhatsApp downloads now stop at the 20 MB limit before the first byte when the size is declared, instead of reading the whole file into memory and measuring afterward.

Deadlines reach the governance band

A decision can now carry a deadline, and missing it pushes the REVIEW band toward scrutiny. DecisionSpec.deadline_s declares one per decision; Decider.decide(..., deadline_s=) overrides it per call. An answer not back in time halts with deadline_missed on the receipt and in the decision log, so misses become countable.

With CHIMERA_GOVERNANCE_BAND_DEADLINE_S set, a miss produces a REVIEW card with band: deadline—never ALLOW. A late answer is never cached or applied after the fact. This is off by default, and the new Settings › General › Governance and audit panel exposes it alongside five other opt-in controls that were previously reachable only through .env or could be saved but not read back. GET /api/config reports them in a governance_audit block. Clearing the deadline no longer leaves an app that cannot start: an empty value now reads as no deadline.

Ollama models read the whole prompt

Calls to ollama_chat/ or ollama/ models never named a context window, so Ollama served its machine default—4,096 tokens in the case that surfaced this. Every longer prompt was cut to about half that size, silently. The first request of a chimera solve run, roughly 20,500 characters, was read as 2,050 tokens, and the model never saw the task.

Every gateway path now sends num_ctx to Ollama routes only, drawn from CHIMERA_OLLAMA_NUM_CTX (default 32,768; set to 0 to keep the server's default). The compaction budget of an Ollama model is sized from that window instead of the 128,000-token fallback. A response whose prompt count shows a cut logs a PROMPT TRUNCATED warning instead of passing in silence. Hosted providers are unchanged.

What to do now

Upgrade with pip install chimera-agent==0.65.0. If you run the desktop app, restart it and check the Messaging card for any bot that needs pairing. If you use CHIMERA_SANDBOX=docker, verify the daemon is running or switch to local. If you call Ollama models, set CHIMERA_OLLAMA_NUM_CTX to match your model's window. The full release notes are at Chimera Agent v0.65.0.