All stories
Thermite Ecosystem Community reports

Thermite puts every research agent on Live—and fixes stalled long prompts

The community autoresearch harness adds a cross-run activity view, individual advisor filters and fixes for large messages left waiting inside coding-agent terminals.

By AITrending2 min read
Sources (5)

Why now Version 0.1.2 was released November 3 at 21:24 UTC, making live activity easier to follow while addressing a concrete long-message failure in Claude Code and Codex terminals.
Thermite run dashboard illustration showing campaign metrics, lanes and a champion table.
Author-supplied run dashboard illustration, not the new Everything view. Displayed speed, gains and cost are illustrative screen content, not independently verified benchmark results. Source: Thermite / iammrduncan

Watching an agent fleet should show more than the workers. Thermite 0.1.2 adds an Everything view that gathers every agent’s live lines across runs and rigs, with filters to narrow the stream. Teachers, mergers, auditors and other advisors now get their own filters, colors and call-start and call-end lines.

Released November 3, the community autoresearch harness also exposes headless advisors’ tool calls as they happen. Within a run, users can switch between Terminals and Lines, keeping their chosen agent. The practical hook is visibility into which participant is working, waiting or returning an answer—not a new model or a benchmark win.

The prompt that never left the input box

The tagged changelog describes a concrete failure: a teacher’s 100 KB message entered its terminal but was never sent. The fix uses bracketed paste for long text, retries Enter while the message remains in the input and allows larger messages more time to start. It specifically discusses Claude Code and Codex behavior. Another fix retrieves a complete multiline advisor answer instead of mistaking a nested one-line object for it.

Thermite coordinates experiment campaigns and says a separate judge measures proposed changes. Its Linux/macOS release bundle starts a control service and one CPU rig from pinned Docker Compose or Podman images. The included harness images carry OpenCode and Codex CLI; they do not include Claude Code. A model is optional for the initial judge smoke check, but real agents need a supported model endpoint, key or subscription, with model costs separate.

Version 0.1.1 can update in place, according to the release; 0.1.0 requires a fresh installation. The project warns that model-written code runs on the user’s computer and can reach the local network under Compose, and that there are no default caps on turns or wall time unless configured.

Visibility does not establish that every advisor is doing the intended research. In a November 4 public issue, user shadowcodex reported seeing no web calls in teacher logs on 0.1.2 and said they would investigate configuration. That remains an unresolved user report, not a confirmed general failure. AITrending has not independently tested the release.

Related reading