Releases
v0.3.0
Compare
Sorry, something went wrong.
No results found
[0.3.0] - 2026-09-18
Added
Optional Pragmatic beta contracts for cited investigation and unapplied patch proposals, preserving independent workers and comparison.
Opt-in Codex session continuation, reasoning effort, native sandbox selection and native subagent controls.
Bounded waits and compact run status responses.
Changed
Show execution deadlines and permission context to workers.
Organize independent attempts and chunked work in the Pragmatic skill. Token savings and equal quality are not established.
Clarify website categories and copy a complete setup prompt from the installation control.
Fixed
Treat interrupted or truncated usage as unknown rather than a complete token total.
Preserve cancellation evidence after child output drains.
Report comparison truncation and explicit Codex quota failures.
Earlier release candidate changes
Added
Added three task recipes, two reproducible examples, and recorded development cases.
Added guides for interpreting results, recovering from failed attempts, and measuring usefulness.
Added hasOutput and finalMessageChars to task views, including compact views.
Changed
Rebuilt the homepage and README around trying alternatives from the same conversation.
Organized documentation around first use, practical tasks, evidence, and technical reference.
Made timeout advice suggest inspecting the attempt before choosing a retry strategy.
Kept failed attempts visible in published examples without claiming measured savings.
Fixed
Bounded task cancellation and cleaned up owned process groups on macOS and Linux.
Shut down active workers when the MCP input stream closes.
Canceled earlier workers when a later raw or compositional task cannot start.
Bounded version probes so one stalled CLI does not block discovery.
Protected positional prompts from being parsed as CLI options.
Retained active run evidence until comparison and receipt capture finish.
Rejected invalid Judge JSON and bounded all selected comparison candidates.
Recognized final-file answers in compact task and compositional views.
Applied output limits to final messages as well as captured streams.
Preserved new history receipts after a truncated previous line.
Rejected unintended browser origins, forged local hosts, and non-JSON chat requests in Ennodia IO.
Distinguished expired agent authentication from evidence about model quality.
Marked Antigravity's explicit no-output error as failed, even with exit code zero.
Isolated skill-discovery tests from developer home directories.
You can’t perform that action at this time.