nerymurillohnd/agent-context-audit
Parallel, cross-domain audits of the operating context that governs coding agents.
Changelog
0.1.0 — unreleased candidate
-
Allow the exact trusted native Codex executable to re-exec inside the deny-by-default Linux sandbox, without granting its parent directory or audited sources.
-
Align the canonical/native listing summary with the 30-character field limit.
-
Preserve hybrid observation provenance and adjudicated finding IDs in the strict persisted run contract, with malformed-field and legacy-readability regressions.
-
Hybrid review: automatic observations, coherent file assignments, independent discovery, semantic verification and accountable final adjudication.
-
Run schema 1.1 with observation/assessment ledgers, legacy report reading and scope-aware baseline comparisons.
-
Broader embedded-credential redaction and line-preserving Claude MCP views that exclude unrelated application state.
-
Synthetic semantic acceptance runner for five cases with three repetitions each; live acceptance is separate from controller and native transport tests.
-
Initial user-owned Codex and Claude operating-context inventory.
-
Complete redacted source chunks, explicit directive dispositions and strict finding contracts.
-
Isolated native Codex workers, parallel specialists, central reconciliation and bounded challenges.
-
Citation/coverage validation, immutable run directories, Markdown reports, canonical status and baseline tracking.
-
One launcher for direct, interactive and scheduled usage; recipe generation has no scheduler side effects.
-
Public release remains gated on the owner-reviewed package and recorded verification matrix.