Conversation
There was a problem hiding this comment.
Pull request overview
Adds a new hackenproof-slop-detector skill/plugin intended to validate report citations against their sources (repo contents at an established commit, advisory databases, and on-chain RPC), and to emit per-citation verdicts that can be used as evidence for downstream reputation decisions.
Changes:
- Introduces the Slop Detector skill docs and agent configuration, including workflow steps and output rules.
- Adds reference guides for code-citation checks, advisory checks (CVE/GHSA), on-chain verification, and verdict/reputation mapping.
- Registers the new plugin in the marketplace manifest (plus a small description escape fix for an existing entry).
Reviewed changes
Copilot reviewed 8 out of 8 changed files in this pull request and generated 5 comments.
Show a summary per file
| File | Description |
|---|---|
| plugins/hackenproof-slop-detector/skills/hackenproof-slop-detector/SKILL.md | Defines the skill’s purpose, workflow, trust boundary, and output rules. |
| plugins/hackenproof-slop-detector/skills/hackenproof-slop-detector/references/verdicts-and-reputation.md | Specifies verdict taxonomy and how evidence maps (conservatively) into reputation guidance. |
| plugins/hackenproof-slop-detector/skills/hackenproof-slop-detector/references/onchain-checks.md | Documents transaction validation and RPC-based verification flow, including retention-window handling. |
| plugins/hackenproof-slop-detector/skills/hackenproof-slop-detector/references/code-citation-checks.md | Documents repo/commit establishment and file/line/symbol/dependency claim validation steps. |
| plugins/hackenproof-slop-detector/skills/hackenproof-slop-detector/references/advisory-checks.md | Documents CVE existence/product checks (NVD) and affected-range checks (GHSA), with rate-limit handling. |
| plugins/hackenproof-slop-detector/skills/hackenproof-slop-detector/agents/openai.yaml | Adds agent UI metadata and a default prompt for invoking the skill. |
| plugins/hackenproof-slop-detector/.claude-plugin/plugin.json | Declares the plugin metadata (name/version/description/author). |
| .claude-plugin/marketplace.json | Adds the new plugin to the marketplace list and fixes an existing description escaping. |
💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.
| | `path.ts:61-66` | verified | quote matches at 61-66 | `gh api …contents/…?ref=…` | | ||
| | `configs/vite/x.ts` | verified-with-drift | real path `packages/extension/configs/vite/x.ts` | `grep -i /x.ts$ tree.txt` | | ||
| | `CVE-2099-99999` | contradicted | NVD totalResults=0 | `curl …?cveId=…` | | ||
| | `3TNNto2…xY61` | pruned-inconclusive | slot predates window 441633334→442113302 (~2.2d) | `getTransaction` + `getFirstAvailableBlock` | |
| | Any `unresolvable` | No penalty, whatever else is present. The check did not run. | | ||
| | One or two `contradicted` among many `verified` | No penalty. Transcription errors. Worth a Need-more-info asking them to correct the citations. | | ||
| | `contradicted` on the claim the finding rests on — the cited code is absent from the tree, or the transaction's stated effect is absent from a transaction the node does hold | Supports Invalid (−5). State the receipts. | | ||
| | Every substantive citation `contradicted`, plus `structurally-invalid` hashes or a non-existent CVE, plus a prose PoC | Supports Spam (−20). This is the only combination that does. | |
Comment on lines
+37
to
+38
| Fails the structural check → **`structurally-invalid`**. This is the only on-chain result that | ||
| supports a fabrication finding on its own. |
Comment on lines
+107
to
+112
| - Transaction's slot or block predates the window, or its date is older than the window → **`pruned-inconclusive`**. State the window. This needs an archival provider (Helius, Triton, QuickNode with full history) or an indexer, and the report cannot be judged on this citation without one. | ||
| - Transaction falls **inside** the window and still returns null → **`not-found-in-range`**. This is a genuine red flag: the node holds that period and does not have the transaction. | ||
|
|
||
| For EVM chains, `eth_getTransactionByHash` on a full node serves all history, so `null` there is | ||
| stronger evidence than on Solana — but confirm the node is not a pruned archive-less instance before | ||
| treating it as `not-found-in-range`. |
Comment on lines
+130
to
+132
| Per hash: chain · structural result · provider that answered · retention window where relevant · | ||
| verdict · the command. Never write "the transaction does not exist" unless the result is | ||
| `not-found-in-range` with the window stated. |
dorsky
reviewed
Sep 7, 2026
| transaction — every specific-looking element a placeholder, a hypothetical, or a narrated result with | ||
| the result withheld. | ||
|
|
||
| This is the same test `hackenproof-report-quality-scorer` applies as its observed-value rule. Where |
Contributor
There was a problem hiding this comment.
"observed-value rule" on hackenproof-report-quality-scorer doesnt exists.
This needs to be fixed regardless of #18's merge, either rename to match Dimension 3 or drop the specific rule-name and just say "defer to its Impact & Evidence scoring."
dorsky
reviewed
Sep 7, 2026
dorsky
left a comment
dorsky
left a comment
Contributor
There was a problem hiding this comment.
@m0erbye for your info
No cross-reference either direction between this skill and hackenproof-poc-grader, even though they overlap: poc-grader already treats "fabricated txhash" and "code citations that don't exist" as negative signals driving its own Invalid/OOS verdict, while slop-detector does deeper per-citation verification of the same facts feeding a separate reputation table. Not a bug, but a triager could run one and not the other and get differently-grounded conclusions on the same evidence. Worth a short "relationship to other skills" note (like #18 has for poc-grader) rather than a functional fix.
| - Dependency-and-version claims ("ships lodash 4.17.x") | ||
| - Repository issue and pull-request numbers | ||
|
|
||
| A report with no citations at all is not slop by that fact alone — it is a report to hand to `hackenproof-report-quality-scorer`, whose observed-value rule covers prose-only submissions. |
Contributor
There was a problem hiding this comment.
"observed-value rule" on hackenproof-report-quality-scorer doesnt exists.
This needs to be fixed regardless of #18's merge, either rename to match Dimension 3 or drop the specific rule-name and just say "defer to its Impact & Evidence scoring."
| { | ||
| "name": "hackenproof-all-reports-export", | ||
| "description": "Export all reports of a HackenProof program into a structured file (Markdown, JSON, or CSV — user's choice)", | ||
| "description": "Export all reports of a HackenProof program into a structured file (Markdown, JSON, or CSV \u2014 user's choice)", |
Contributor
There was a problem hiding this comment.
no need to do any changes here
Closes hackenproof-public#13. Checks a report's cited facts against the thing each one cites: files, lines and symbols against the in-scope commit; CVEs and advisories against NVD and GHSA including product and version-range matching; transaction hashes against the chain; and whether the PoC is prose. Every threshold, endpoint and failure mode here was measured against live reports before being written, and the measurements are in the references so a verdict can be re-derived rather than trusted. The governing rule is resolve-before-accuse. Three ordinary citation mistakes are indistinguishable from fabrication under a naive check, and all three were found on a report that turned out to be sound: an annotated tag whose ref SHA is not the commit, a quote sitting at line 67 inside a correct 62-74 range, and a path missing its packages/extension prefix. So every check runs a resolution step — basename search over the tree, tag dereference, a +/-30 line quote search — before it may return a negative. Verdicts are graded per citation and never averaged: verified, verified-with-drift, citation-without-commit, unresolvable, contradicted. The load-bearing distinction is unresolvable versus contradicted — could not look versus looked and it is absent. A public Solana node holds about 2.2 days of history, so a null result on a three-month-old signature is pruning, not invention; the retention window is checked with getFirstAvailableBlock before any such verdict, and several endpoints commonly recommended for this are dead, including the one a live report told the triager to use. Only contradicted findings on a load-bearing claim can support the -5/-20 reputation policy the issue asks this to feed, receipts printed, and never while an unresolvable verdict stands on the same claim. No clone anywhere: repositories in scope reach several gigabytes, so files come from contents?ref= and the tree is fetched once per commit.
m0erbye
force-pushed
the
feat/slop-detector
branch
from
September 7, 2026 18:04
0a9d886 to
c5470a1
Compare
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
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.
Summary
Implements the Slop Detector requested in #13: a skill that checks each fact a report cites against the source it cites, and returns a verdict per citation rather than a judgement on the report.
verified,verified-with-drift,citation-without-commit,unresolvable,contradicted. The load-bearing pair isunresolvablevscontradicted— could not look vs looked and it is absent.contradictedon a load-bearing claim can support a penalty, receipts printed, and never while anunresolvableverdict stands on the same claim.target.typeisgithub_repository, and only against a commit that could actually be established — a citation pinned to a moving branch is reported as its own defect, not silently checked againstHEAD.contents?ref=, the tree is fetched once per commit. Repositories in scope reach several gigabytes.Structure
Follows the existing plugin layout (mirrors
hackenproof-poc-grader):Plus a
marketplace.jsonentry.claude plugin validatepasses for both the plugin and the marketplace manifest.Notes for review
hackenproof-poc-grader's evidence tiers; it defers to that skill rather than reimplementing it, and comes out cleanly if you would rather this skill not touch PoC form.file:line, not dependency advisories — happy to split them into a dependency-class skill instead.pruned-inconclusiveis honest but terminal for any on-chain proof older than a couple of days. If the team has an archival RPC key, tell me where the skill should read it from and the verdict becomes decidable.Closes #13