adamast-failure-modes

AdaMAST for Codex

Use AdaMAST as a lightweight runtime discipline while doing the user's task.

Runtime behavior

• Keep the active taxonomy out of startup context unless the user or an AdaMAST command explicitly supplies it.
• At meaningful boundaries, inspect the recent trajectory before continuing:
  • finishing a sub-task;
  • recovering from a failed tool command;
  • switching strategy;
  • preparing to submit a final answer.
• Use the reflection order from AdaMAST prompts:
  1. Observe concrete events or missing expected steps.
  2. Correlate only evidence-supported causes.
  3. Map to taxonomy codes only when evidence supports the match.
  4. Decide whether to make one focused repair or continue.
• Treat none apply as valid. Do not invent a failure mode or force an edit.
• When the SessionStart context says taxonomy selection is pending, show the supplied selector verbatim and do no task work until the user chooses. After selection, resume any held task without asking the user to repeat it.
• When the selector reports No taxonomy, do not emit AdaMAST checkpoints or describe AdaMAST as active for that conversation.
• Before final submission, complete a final AdaMAST gate and only report ready when no unresolved taxonomy-relevant issue remains.
• When runtime context announces an AdaMAST native taxonomy learning job, immediately launch exactly one native Codex subagent with the supplied task prompt. Continue the user's main work while it runs. Do not perform the taxonomy job in the main agent, invoke codex exec, request an API key, or recursively launch another taxonomy agent. The subagent must return the supplied receipt envelope exactly; AdaMAST validates and activates it.
• After any failed tool call, privately map the evidence to the active taxonomy and, before the next tool call, emit one concise user-facing progress checkpoint. Use none apply when the failure was intentional or harmless; do not invent a failure code.
• Keep routine hook polls silent. Show each learning lifecycle notice once when runtime context requests it, and keep the longer reflection private.
• End each substantive Codex final answer with the compact checkpoint required by the active runtime context: Checkpoint, Relevant codes, Evidence, and Next action. Keep the longer Observe/Correlate/Map/Decide reflection internal unless a hook explicitly requests it.