Add init, guard, runtime, advisory intelligence, and contribution prompting#2
Merged
Merged
Conversation
- opena2a init: Security posture assessment with trust score, credential scan, hygiene checks, and prioritized next steps (protect flow) - opena2a guard: ConfigGuard MVP with sign/verify/status subcommands for config file integrity via SHA-256 hash pinning - opena2a runtime: ARP wrapper with start/status/tail/init subcommands for agent runtime protection - verify: Enhanced with trust profile and oracle verdict queries from registry, showing trust score, verdict, and dependency risks - Extract credential patterns to shared module for init/protect reuse - Remove guard/runtime from adapter registry (now direct commands) - 22 new tests (init: 8, guard: 8, runtime: 6)
- Advisory check: init command fetches advisories from registry and warns about flagged tools in the project's dependencies - Advisory cache: 5-minute local cache to avoid repeated fetches - Package detection: Scans package.json, go.mod, requirements.txt for dependency names to match against advisories - Scan report submission: Utility for sharing detailed scan findings with registry when contribute mode is enabled (opt-in) - 7 new tests for advisory matching, caching, and error handling
- Track scan count in user config, only prompt to share reports after 3+ scans - Add shouldPromptContribute/dismissContributePrompt to shared config - Add open source link to init report footer and HTML security report - 5 new telemetry config tests
![github-actions[bot]](https://avatars.githubusercontent.com/in/15368?s=60&v=4){kind=small}
This was referenced Mar 2, 2026
5 tasks
6 tasks
7 tasks
thebenignhacker
added a commit
that referenced
this pull request
Apr 22, 2026
…ck (#90) * fix(cli): aggregate HMA findings into review output so Observations block reflects full scan aggregateFindings was skipping the HMA phase entirely, so `opena2a review` on a 34-critical fixture reported 1 critical while `hackmyagent secure` on the same dir reported 34. The composite score already consumed hmaData.score (weighted at 8%) so the headline grade was correct, but the text-mode totalFindings, severity counts, and the @opena2a/cli-ui Observations block saw only credential-scan + shield data -- the Categories line collapsed to "other (1 critical)" even when HMA fired on 17 distinct categories. Changes: - Expose all failed HMA findings via a new HmaPhaseData.allFailedFindings field. topFindings is preserved for the HTML report's HMA tab (deduped- by-checkId + capped at 30); allFailedFindings carries the raw set so the Observations block and overview severity counts match the HMA scan 1:1. - aggregateFindings now accepts hmaData and maps each HmaFinding to a ReviewFinding with source='hma'. checkId flows through as the ReviewFinding id so cli-ui's classifyCategory routes CRED-*, MCP-*, SKILL-*, etc. into the right buckets. - Prefer-HMA dedupe: quickCredentialScan matches at the same file:line as an HMA finding are dropped. HMA credential detection is context-gated across 200+ check IDs, which makes it the authoritative layer. - computeCompositeScore already reads hmaData.score, not findings.length, so the composite remains display-only stable. Verified on test/hma/: score stays 66/100, findings jump 1 -> 232 (56 critical, 118 high, 57 medium), Categories line breaks out credentials/MCP/prompt/skill/etc., verdict names an HMA-detected lead finding. Tests (881 total, +4 new): - review credentials smoke asserts CRED-* findings surface in the cli-ui "credentials" Categories bucket end-to-end (skipHma=true). - aggregateFindings unit tests cover the prefer-HMA dedupe, the no-overlap pass-through, and the null-hmaData backward-compat path. Closes the P1 follow-up tracked in memory/project_hma_observations_ux_shipped.md. * fix(cli): scope credential dedupe to HMA credential checks + handle line-less HMA findings Adversarial self-review surfaced two defects in the initial aggregateFindings HMA merge: 1. HMA often emits credential findings without a line number (AST-CRED-001, AST-CRED-003, ENVLEAK-*, etc. on real output routinely have file=... and line=null). The original dedupe keyed credData matches against `file:line` only, so the "prefer HMA" path was a no-op on the canonical credential scenario. Users saw both the credential-scan CRITICAL and the HMA HIGH/ CRITICAL at the same file, inflating totalFindings. Now we key by BOTH `file:line` AND `file` alone, so credData is suppressed when HMA already fired on the same location — with or without a line. 2. The file-only key was unscoped: a non-credential HMA finding on a file (e.g. GIT-002 on .gitignore, DEP-004 on package.json) would have silently masked a credential that quickCredentialScan found in that same file. Dedupe is now scoped to HMA credential-category check prefixes (CRED, AST-CRED, WEBCRED, SEM-CRED, AGENT-CRED, ENVLEAK, CLIPASS, DRIFT) so non-credential HMA findings on a file never suppress a credential finding in that same file. 3. MEDIUM bug: when dedupe did fire, the surviving HMA finding kept its own severity. If credData flagged CRITICAL (e.g. sk-ant-*) and HMA flagged HIGH (SEM-CRED-002), the user-visible severity silently narrowed. The HMA finding's severity is now upgraded to max(hmaSeverity, credSeverity) so the verdict and category counts reflect the worst case. Three new unit tests cover: the line-less HMA dedupe path (regression for #1), the non-credential-prefix bypass (regression for #2), and the severity upgrade path (regression for #3). Verified: `opena2a review --ci test/hma/` still reports 232 findings / 66/100 composite (unchanged — credData returns 0 on that fixture, so the dedupe behavior isn't exercised there). The regressions addressed are path-dependent on credData producing overlap with HMA, which the new unit tests exercise directly. Full test suite: 884/884 passing. * fix(cli): guard against out-of-targetDir paths in review credData aggregation Auto-reviewer on #90 flagged the credData loop for lack of path boundary validation. credData.matches is sourced from walkFiles(targetDir, ...) in credential-patterns.ts, which already constrains traversal to within targetDir — so this is defense in depth rather than a missing security control. If a symlink escape or future upstream contract change ever leaked an absolute or parent-traversal path into aggregation, we now drop it rather than render a misleading ".." row in the review output. Also adds a JSDoc note on HmaPhaseData.topFindings clarifying it is the display-only, deduped-by-checkId + slice(0,30) list for the HMA tab, distinct from allFailedFindings which is the source of truth for counts. Addresses the dual-data-path architecture concern in the same review. New unit test covers the boundary: three credData matches (abs path, parent traversal, in-scope) → only the in-scope one survives. 885 tests passing. * fix(cli): strengthen review path boundary check with resolve+startsWith Auto-reviewer round 2 correctly pointed out that `rel.startsWith('..') || path.isAbsolute(rel)` doesn't catch Windows-drive-letter paths on Unix (e.g. `C:\\etc\\passwd` is not absolute under POSIX `path.isAbsolute`). Switch to the stronger pattern: resolve both sides, require the credential file to be the targetDir itself or start with `resolvedTarget + path.sep`. Still defense in depth — credData.matches is still produced by walkFiles under targetDir — but now survives the specific bypass the reviewer named.
4 tasks
This was referenced Apr 29, 2026
5 tasks
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.
Summary
opena2a init: Security posture assessment with trust score, credential scan, hygiene checks, and prioritized next steps (screenshot moment)opena2a guard sign/verify/status: ConfigGuard MVP -- SHA-256 hash pinning for config file integrityopena2a runtime start/status/tail/init: ARP runtime protection wrapper with auto-config generationopena2a verify: Enhanced with trust profile queries and Ed25519-signed oracle verdicts from registry--passWithNoTests)Stats
Test plan
npx vitest run)npx turbo run build)npx turbo run lint)opena2a initin project with credentials shows protect as first next stepopena2a guard signthen modify file thenopena2a guard verifydetects tamperingopena2a verify --package hackmyagentshows trust score from registry