Future AGI Files

1.58.0 (2026-11-03)

Features

• error-feed: send the agent's prompt and tool calls with each simulation call (bb3162d)
• simulate: fix my agent improvements tab, fix page and launch form [TH-8955] [agent] (f28588d)
• simulate: let a fix protect cases that must still pass (2747c4b)
• simulate: prove a fix for a failing run and open its pull request (4a4f226)
• simulate: ship a candidate the must-pass check vetoed, with a reason (fec9b17)

Bug Fixes

• agentic_eval: keep empty few-shot inputs skipped (a51310b)
• agentic_eval: keep zero and false few-shot inputs, skip empty ones by the eval rule (8102cea)
• agentic_eval: match few-shot text inputs as strings so structured inputs find feedback (47c94b3)
• agentic_eval: send JSON eval inputs to the embedder as text, not objects (214b225)
• agentic_eval: send JSON eval inputs to the embedder as text, not objects [TH-8996] [agent] (61923d5)
• error-feed: keep a tool-heavy call under the evidence page limit (25f1cb2)
• gateway: give every Gemini tool call its own id (88e8397)
• gateway: give every Gemini tool call its own id (7a2c1be)
• simulate: group accepted findings and revisit deferred batches (478e11a)
• simulate: offer only passing cases to protect [TH-8955] (0cbc866)
• simulate: protect only cases that passed and were scored (5277fd7)
• simulate: read the frozen plan for a refused candidate's held-back tier (aed84c6)
• simulate: remove hover tooltips from the environment workspace tabs (bd36504)
• simulate: remove hover tooltips from the environment workspace tabs [TH-8994] [agent] (0071539)
• simulate: ship a must-pass refusal only on a whole final check (d54614a)
• simulate: ship exactly the accepted candidate's files (773a190)
• simulate: skip the fix list when the org lacks fix my agent [TH-8955] (cb13120)