Failure-process hardening: annotate silent results, detect same-failure repairs, thread requirements into checkers#88
Open
davidrhodus wants to merge 5 commits into
Open
Conversation
Empty and stderr-only results are where models form false premises
("did it work?" / "stderr means it failed") — the largest failure
trigger in Zhao et al.'s CLI-agent trajectory study (arXiv:2607.09510,
false premises = 30.7% of decisive errors). Make the verdict explicit:
- empty success: [no output — command succeeded (exit 0)]
- stderr-only success: [command succeeded (exit 0) — output above is stderr]
- empty failure: [exit code N — no output on stdout or stderr]
- successful commands containing a `cd` segment get a note that the
working directory resets before the next command (just-in-time
version of the system-prompt warning)
Success annotations never contain "[exit code " — steering classifies
that marker as failure. Also fix the steering timed-out matcher, which
looked for "[timed out after " while the emitted text is
"[timed out — process killed]".
Co-Authored-By: Claude Fable 5 <[email protected]>
"Repairing the wrong problem" is the costliest coding-agent failure behavior (39% of wasted execution in arXiv:2607.09510). The verifier now hashes each failure (stage, exit code, digit/hex-masked output) and compares rounds: when a repair didn't change the failure signature, the verify nudge stops saying "reconsider" and instead demands a validated diagnosis via a discriminating probe before any further edit. Line/column drift is masked so an edit that only shifts code is still recognized as the same failure. Unstable-stage failures keep their own escalation (stage_mutation_counts) and are excluded. Nudge-only in v1 — no control-flow change; `repeated_verify_failures` telemetry (report key added) will show whether burning rounds is worth it. Co-Authored-By: Claude Fable 5 <[email protected]>
… review Supplying the task's requirements to a checker is the single biggest recall lever for specification-relative failures (arXiv:2607.09510: ignored-requirement recall 3%->22%, false premises 15%->32% once the monitor sees the spec). The independent review previously saw only the derived TaskContract + diff; the verify-failure nudge saw no requirement text at all. TaskContract::requirements_digest builds a bounded (~1.5 KB) digest — derived acceptance sentences plus a verbatim prompt excerpt — computed once per turn and injected into (1) the verify-failure nudge and (2) the independent review context. Deterministic only: no side-call, no config, no latency. Co-Authored-By: Claude Fable 5 <[email protected]>
Fabricated success clusters exactly where verification is absent (26% of failed trajectories in arXiv:2607.09510, beginning at the point of no return). VerificationStatus::Unverified already gates status and exit codes end-to-end; this closes the residual gaps: - trio review stays fail-open, but an Unavailable verdict now books skeptic_unavailable_count / skeptic_last_status telemetry and the TUI prints an explicit "approved WITHOUT review" warning instead of a dim aside - one-shot runs print a deterministic stderr trailer when changes were made but no verification stage ran — emitted regardless of --allow-unverified (the flag changes the exit code, not the truth) - regression tests lock the honest outcomes: round-cap exhaustion reports Failed+Incomplete, mutation without stages reports Unverified, and --allow-unverified completes the turn without laundering the verification status Co-Authored-By: Claude Fable 5 <[email protected]>
cargo fmt on the core crates plus the two clippy -D warnings failures the trio commit introduced (doc list continuation in trio.rs, nested if in parse_trio_args). hi-mlx formatting is left untouched — it doesn't build everywhere and is outside the CI crate list, though note that `cargo fmt --all --check` still fails on it. Co-Authored-By: Claude Fable 5 <[email protected]>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Four changes motivated by "Failure as a Process: An Anatomy of CLI Coding Agent Trajectories" (arXiv:2607.09510) — 1,794 trajectories showing coding-agent failures dominated by epistemic errors (57.9%; false premises 30.7%), with wrong-problem repair consuming 39% of wasted execution and 26% of failed runs fabricating success.
[no output — command succeeded (exit 0)]; stderr-only success and empty failure annotated; successfulcdcommands get a just-in-time cwd-reset note. Fixes the steering timed-out matcher (matched a never-emitted string).repeated_verify_failurestelemetry + report key. Nudge-only in v1.Evidence
repeated_verify_failuresfired on real Terminal-Bench tasks in the new Harbor integration.🤖 Generated with Claude Code