Skip to main content

asterweave:failure-analyst

Role​

Diagnoses repeated build, test, verification, CI, or review failures using competing hypotheses, state/diff comparison, and stable failure signatures — without making changes.

Purpose​

Reproduces a repeated failure with the smallest safe command, normalizes a stable signature, and compares the current repository/diff/environment with prior attempts. Generates competing hypotheses, stating supporting and contradicting evidence and the cheapest discriminating check for each. Detects no-progress cycles — an identical signature plus a materially unchanged diff/config/environment — and refuses to rerun the same command twice without a new hypothesis or state change.

When invoked​

Escalated to when a stable failure signature recurs across implement/test/verify/review/monitor-pipeline back-edges without resolution — see Pipeline failures.

Inputs​

The recurring failure signature, the current and prior diffs/environment, and relevant CI/test/review output.

Context received​

A bounded diagnostic assignment with the failure history, not the full conversation.

Tools / permissions​

Read, Grep, Glob, Bash, PowerShell. Explicitly disallowed: Write, Edit, Agent.

Writes allowed?​

No.

Output format​

Root-cause confidence, evidence, ruled-out hypotheses, a recommended retry, replan, needs-human, or security-escalation edge, and the stable failure signature. Separates product defects, test defects, environment failures, flaky behavior, dependency/toolchain issues, permission/policy denials, and stale evidence.

Typical workflow​

A node fails repeatedly with the same signature → failure-analyst is invoked instead of retrying blindly → its recommended edge routes the graph forward (retry, replan, or escalate to a human).

Invoked from within /asterweave:deliver's repair loop; also useful for /asterweave:retro.