| name | adamast-failure-modes |
| description | Use when working on an agent task where AdaMAST failure-mode checkpoints, final submission gates, trace capture, taxonomy generation/refinement, or AdaMAST CLI setup should guide Codex. This skill helps Codex apply AdaMAST during software or research tasks, diagnose its own trajectory against the active taxonomy, and avoid claiming completion before the AdaMAST final gate is satisfied. |
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:
- Observe concrete events or missing expected steps.
- Correlate only evidence-supported causes.
- Map to taxonomy codes only when evidence supports the match.
- 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.
If the AdaMAST package is available
Prefer the package CLIs over hand-rolled state:
- Use
adamast-doctor --config adamast.json to check setup.
- Use
adamast-find --list or adamast-find --inherit <taxonomy_id> to resolve stored taxonomies.
- Use
adamast-single-run for no-harness single-model tasks.
- Use
adamast-dashboard to inspect recorded evidence and fired codes.
- Use
adamast-traces status to inspect trace growth.
If a command asks for --trace-output, use the project-specific trace folder supplied by the user or ./adamast-program for local experiments.
If no AdaMAST command is available
Still follow the AdaMAST final-gate shape in the final reasoning pass:
Final AdaMAST status: READY_TO_SUBMIT or REPAIR_REQUIRED
Codes checked: relevant taxonomy ids, or none
Evidence: concrete task or verification evidence
Repair attempts used: integer count
Final decision: submit, repair, or report unresolved
Do not expose private chain-of-thought. Keep the final user-facing answer concise; mention only the actionable result and verification status.