docs(testing): proactive bug-finding harness + 2026-07-16 audit (11 confirmed)#770
Merged
Merged
Conversation
…ent audit Documents the four-layer approach for finding operator-sequence bugs before operators do (the class the Canary402 field report was made of), plus two runnable artifacts: the reusable obol-audit workflow (hostile-operator lens sweep + adversarial refuters) and the run-canary.sh ephemeral-cluster skeleton. Records the 2026-07-16 first run: 11 confirmed findings. The layer-1 name-injectivity oracle (PR #767) and layer-4 agent audit independently found the same ReferenceGrant collision — two methods, one bug. Claude-Session: https://claude.ai/code/session_014YjPMViNrZ7zBVgUQzwEKk
…dings) Second layer-4 run across the 14 subsystems the first pass did not cover (wallet/key, backup, stack lifecycle, self-update, networking, buyer flow, hermes, openclaw import, model serving, storefront, secrets, infra shell-out, CRD validation). 21 confirmed (3 critical), the most severe bugs living exactly in the subsystems the core audit never reached. Claude-Session: https://claude.ai/code/session_014YjPMViNrZ7zBVgUQzwEKk
This was referenced Jul 16, 2026
Contributor
Author
|
Superseded by #778, which removes these internal tooling/docs from the production repo entirely (they don't belong in the shipped product, and the findings tables risked disclosing open issues on public |
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Proactive bug-finding harness — how we find operator-sequence bugs before operators do
Follow-up to the Canary402 field report. Every one of those 10 bugs was found by an operator doing a reasonable-but-unanticipated sequence on a live deployment — none were happy-path failures — which is why example-based unit tests missed them and a human found them first.
This PR documents the one approach that finds that class automatically (a synthetic operator driving generated sequences, checked against invariants rather than hand-written expectations) as four layers, and adds two runnable artifacts:
docs/testing/proactive-bug-finding.md— the six invariants, the four layers (invariant oracles → stateful model tests → nightly canary cluster → agent audit), how to run each, and why it beats more unit tests..claude/workflows/obol-audit.js— the reusable layer-4 workflow: N hostile-operator lenses fan out, each candidate is checked by two independent adversarial refuters, output is the ranked confirmed survivors. Re-run withWorkflow({ name: "obol-audit", args: { known: [...already-fixed...] } }).hack/canary/run-canary.sh— the layer-3 ephemeral-k3s scenario-matrix skeleton (real CLI + black-box invariant probes; fail-loud for nightly CI). Cluster bring-up and the on-chain scenarios are marked TODO for your environment.First run — 2026-07-16 (recorded in the doc)
Layer 4 was run against integration with all six field-report fixes merged, so it hunted for new bugs. 8 lenses × 2 refuters → 11 confirmed (3 rejected by a refuter, correctly). The layer-1 name-injectivity oracle independently found the same #1 grant-collision bug the agent sweep did — two methods, one bug.
Two of the eleven are already handled: the critical unvalidated-price bug (separate PR) and the ReferenceGrant dash-collision (folded into #767). The doc's triage table lists the remaining nine confirmed findings (four of them the same status-truth invariant broken in new places) as an actionable worklist.
Docs + tooling only — no production code changes.
https://claude.ai/code/session_014YjPMViNrZ7zBVgUQzwEKk