Prototype graph-grounded deep regression review for high-risk OpenClaw PRs

[100yenadmin]

Plain-English Goal

ClawSweeper is trusted to help review OpenClaw pull requests. The problem is that some PRs can look fine in tests but still break important behavior, especially when they touch core runtime, auth, sessions, provider routing, release automation, or default configuration.

This issue tracks a prototype that makes ClawSweeper more cautious on those risky PRs. Instead of saying a risky PR is safe too quickly, ClawSweeper should:

• Notice when a PR touches dangerous parts of OpenClaw.
• Ask a harder set of review questions for those PRs.
• Use optional code-map context when available.
• Clearly say what evidence it used.
• Clearly say what it could not prove.
• Tell maintainers when a human architecture review is still needed.

Implementation PR: #440

Current PR head: 08d97656a837922c03ed48c848f12f2c6c400370

TLDR

This is not a promise that ClawSweeper can prevent every bad release. It is a safer review path for high-risk OpenClaw PRs.

Before this change, a review could pass while still being unclear about whether it had enough context to catch a deep regression. After this change, ClawSweeper has a specific place to say: "this PR touches risky code, here is what I checked, here is what worries me, and here is what a maintainer must decide."

The stale bot-comment cleanup was removed from this work and moved to #441. This issue is about better review judgment, not deleting old GitHub comments.

flowchart LR
  PR["OpenClaw PR"] --> Risk{"Touches risky code?"}
  Risk -->|"No"| Normal["Normal ClawSweeper review"]
  Risk -->|"Yes"| Deep["Harder regression review"]
  Map["Optional code map"] --> Deep
  Deep --> Report["Clear report: evidence, uncertainty, maintainer action"]

Loading

What The New Code Is For

The code in PR #440 adds a default-off prototype. "Default-off" means it does not turn on the optional code-map feature unless maintainers explicitly configure it.

The code mainly does five things:

1. It detects high-risk OpenClaw PRs.

   Example: a PR that touches auth/session/core runtime code should get a tougher review than a docs-only PR.

2. It adds a new review section called deepRegressionReview.

   This is where ClawSweeper must say whether the PR looks safe, needs attention, or should be blocked until a maintainer reviews the design.

3. It can include optional GitNexus context.

   GitNexus is just a code map. It helps answer questions like "what other files or symbols are related to this change?" It is only advisory. GitHub's actual diff and checked-out code remain the source of truth.

4. It records context limits.

   If ClawSweeper does not know its context-window size, it now records unknown instead of implying it had enough room to understand the whole repo.

5. It fails closed around sensitive output.

   If the optional code-map output looks like it may contain a secret, ClawSweeper omits that context instead of putting it into a prompt or report.

Why This Is Worth The Size

The PR is large because this touches the whole review pipeline, not because it adds one giant new algorithm.

The change has to update:

• how ClawSweeper classifies risky PRs,
• what it asks the model to review,
• what response shape the model must return,
• how reports are rendered,
• how review-cache keys are built,
• how automation markers decide whether a PR needs a human,
• documentation,
• and tests for the risky failure cases.

That is why the diff is big. The purpose is still simple: make ClawSweeper less likely to give overconfident approval on risky OpenClaw PRs.

Terms In Normal Language

• GitNexus: a searchable map of the codebase. It can suggest related files, but it does not replace reading the actual PR diff.
• High-risk classifier: a checklist that says "this PR touches sensitive code, review it more carefully."
• Context-window telemetry: a note saying how much code ClawSweeper was configured to fit into its model context, or unknown if it cannot tell.
• Fail closed: if something looks unsafe or unclear, ClawSweeper lowers confidence instead of pretending everything is fine.
• Deep regression: a bug that does not show up as a simple test failure but breaks important behavior because the architecture or assumptions are wrong.

Current Validation

Local validation on 08d9765:

pnpm run build: passed
node --test test/repair/update-command-status.test.ts: tests 14, pass 14, fail 0
pnpm run lint:src: passed
pnpm run format:check: passed
git diff --check: passed

Remote validation on 08d9765:

pnpm check: SUCCESS
CodeQL / Analyze (actions): SUCCESS
CodeQL / Analyze (javascript-typescript): SUCCESS
Windows Codex launcher: SUCCESS
Socket Security / Pull Request Alerts: SUCCESS
Socket Security / Project Report: SUCCESS
Merge state: CLEAN
Current-head review threads: none

Maintainer Decision

Maintainers need to decide whether this default-off review contract belongs in ClawSweeper now.

A good merge decision is not "this guarantees safe releases." It does not.

A good merge decision is: "this gives ClawSweeper a clearer, more honest way to handle risky OpenClaw PRs, and we accept the default-off GitNexus/code-map boundary."

Not Included

• No OpenClaw runtime changes.
• No auto-merge behavior is enabled by default.
• No claim that ClawSweeper has or needs a 1M-token context window.
• No claim that this proves release readiness.
• No stale comment deletion; that separate decision is tracked in Evaluate stale re-review command-status comment cleanup #441.

[clawsweeper]

Codex review: keeping this open for maintainer follow-up; there is still a little grit to resolve. Reviewed July 9, 2026, 11:52 AM ET / 15:52 UTC.

Summary
Keep open: current main and v0.3.0 still lack the requested deepRegressionReview/GitNexus prototype, and #440 remains the open implementation candidate.

Reproducibility: not applicable. this is a feature/prototype request, not a report of broken current behavior. Source search verifies current main lacks the requested deep-regression/GitNexus surface.

Root-cause cluster
Relationship: canonical
Canonical: #439
Summary: This issue is the canonical product/security decision thread for the deep-regression review prototype; linked PRs and issues handle implementation or adjacent GitNexus/comment-cleanup work.

Members:

• fixed_by_candidate: Add deep regression review prototype #440 - The PR uses closing syntax for this issue and implements the requested deepRegressionReview/GitNexus prototype.
• adjacent_distinct: Evaluate stale re-review command-status comment cleanup #441 - The issue body says stale re-review command-status cleanup was removed from the prototype and split into a separate policy decision.
• partial_overlap: Evaluate GitNexus-backed context for ClawSweeper cost, speed, and review accuracy #442 - That issue evaluates GitNexus-backed context for cost, speed, and accuracy, but it does not replace the deep-regression review hook requested here.
• adjacent_distinct: Add GitNexus adoption and benchmark guide #443 - That PR implements adoption and benchmark guidance for the related GitNexus evaluation issue, not this prototype hook.

Proposal only: this assessment does not dispatch repair, suppress jobs, mutate sibling items, close, or merge anything.

Ways to help us reproduce this

• Add a screenshot or short recording showing the behavior.
• Add expected vs actual behavior.
• Share version, platform, channel/provider, and relevant config details.

Next step
The open linked PR already owns the implementation; the remaining action is maintainer product/security-boundary review and branch refresh, not a new generated fix PR.

Maintainer decision needed

• Question: Should ClawSweeper accept the default-off deepRegressionReview and advisory GitNexus context contract for high-risk OpenClaw PR reviews?
• Rationale: The requested work changes review schema, prompts, cache state, report rendering, and automation markers while adding a secret-sensitive external graph-output boundary.
• Likely owner: Peter Steinberger — Peter has the strongest recent adjacent history on review/apply safety and automation gates affected by this security-sensitive review contract.
• Options:
  • Accept Default-Off Prototype (recommended): Refresh and land the linked PR if maintainers accept GitNexus as advisory, default-off, and fail-closed around secret-like output.
  • Narrow Graph Context: Keep deterministic high-risk review output but remove or further gate GitNexus packet insertion before landing.
  • Pause Prototype: Leave this issue and the linked PR open until maintainers define the permanent review and security contract separately.

Security
Needs attention: Security-sensitive: optional GitNexus context must not leak secret-like output into prompts, reports, markers, or cache state.

Review details

Best possible solution:

Use #440 as the implementation vehicle after it is refreshed; keep GitNexus adoption benchmarking in #442 and stale comment cleanup in #441 separate.

Do we have a high-confidence way to reproduce the issue?

Not applicable: this is a feature/prototype request, not a report of broken current behavior. Source search verifies current main lacks the requested deep-regression/GitNexus surface.

Is this the best way to solve the issue?

Unclear until maintainer review: #440 is the intended implementation path, but the GitNexus/schema/security boundary is a product and security decision.

AGENTS.md: found and applied where relevant.

Remaining risk / open question:

• The linked implementation changes durable review schema, prompt expectations, cache state, report rendering, and automation markers, so maintainers should accept that rollout before this issue closes.
• The optional GitNexus path may process secret-like advisory output, so the redaction, omission, hashing, and fail-closed behavior need explicit security review.
• Add deep regression review prototype #440 currently reports CONFLICTING against main, so the implementation is not landable until refreshed.

Codex review notes: model internal, reasoning high; reviewed against 411e97156ab4.

Label changes

Label changes:

• remove clawsweeper:fix-shape-clear: Current issue advisory state no longer selects this label.

Label justifications:

• P2: This is a normal-priority review automation feature with bounded but security-sensitive maintainer acceptance work.
• impact:security: The issue is explicitly about high-risk review and optional graph context that must avoid leaking secret-like tool output.

Evidence reviewed

Security concerns:

• [medium] Approve the GitNexus output boundary
  The linked implementation proposes advisory graph context with redaction, omission, hashing, and fail-closed behavior; maintainers should approve that boundary before accepting it as review infrastructure.
  Confidence: 0.87

What I checked:

• Repository policy read: AGENTS.md was read fully; its proposal-only review lane, automation-safety, protected handling, and full-URL reference guidance affected this review. (AGENTS.md:27, 411e97156ab4)
• Current main decision contract: The current Decision interface lists existing review fields and has no deepRegressionReview field on main. (src/clawsweeper.ts:530, 411e97156ab4)
• Current main schema keys: The current DECISION_SCHEMA_KEYS list has no deepRegressionReview entry, so the requested response contract is not implemented on main. (src/clawsweeper.ts:1777, 411e97156ab4)
• Current main search: A targeted source/docs/test/workflow search found no deepRegressionReview, GitNexus, gitnexus, CLAWSWEEPER_GITNEXUS_CONTEXT, or context-window telemetry surface on current main. (411e97156ab4)
• Latest release search: The latest release tag v0.3.0 points at dc82491 and also has no requested deep-regression or GitNexus terms in src/clawsweeper.ts. (src/clawsweeper.ts, dc824915bb6c)
• Open linked implementation PR: GitHub reports Add deep regression review prototype #440 open with closing reference to this issue, head 08d9765, 15 changed files, and current mergeStateStatus CONFLICTING. (08d97656a837)

Likely related people:

• 100yenadmin: Git blame ties much of the current Decision interface and schema-key baseline to Eva's review-policy work, and the linked implementation PR is authored from the same account. (role: feature-history contributor; confidence: high; commits: d20cf36edb6b, 08d97656a837; files: src/clawsweeper.ts, schema/clawsweeper-decision.schema.json, prompts/review-item.md)
• brokemac79: Recent history added maintainer decision packets and review-output policy that this issue extends. (role: recent area contributor; confidence: high; commits: 232898fb6edc, 7af92c124673, 7e27e5dc8055; files: src/clawsweeper.ts, schema/clawsweeper-decision.schema.json, prompts/review-item.md)
• Peter Steinberger: Recent main-history commits touched issue-review retrying, apply throughput, review marker safety, runtime bounds, and local review behavior near the automation gates this issue would affect. (role: recent adjacent automation owner; confidence: high; commits: 411e97156ab4, 76360b98742b, 25d7e74e6dc5; files: src/clawsweeper.ts, .github/workflows/sweep.yml)
• Yuval Dinodia: Review-content cache and re-review history are adjacent surfaces that the linked implementation says it changes. (role: adjacent review-cache contributor; confidence: medium; commits: eca6fbfa0e6d, 4db0631ce933; files: src/clawsweeper.ts, test/review-content-cache.test.ts)

How this review workflow works

• ClawSweeper keeps one durable marker-backed review comment per issue or PR.
• Re-runs edit this comment so the latest verdict, findings, and automation markers stay together instead of adding duplicate bot comments.
• A fresh review can be triggered by eligible @clawsweeper re-review comments, exact-item GitHub events, scheduled/background review runs, or manual workflow dispatch.
• PR/issue authors and users with repository write access can comment @clawsweeper re-review or @clawsweeper re-run on an open PR or issue to request a fresh review only.
• Maintainers can also comment @clawsweeper review to request a fresh review only.
• Fresh-review commands do not start repair, autofix, rebase, CI repair, or automerge.
• Maintainer-only repair and merge flows require explicit commands such as @clawsweeper autofix, @clawsweeper automerge, @clawsweeper fix ci, or @clawsweeper address review.
• Maintainers can comment @clawsweeper explain to ask for more context, or @clawsweeper stop to stop active automation.