Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion docs/adaptive-context.md
Original file line number Diff line number Diff line change
Expand Up @@ -6,7 +6,7 @@ After collection passes, the trusted Actions host writes a private `review-plan.

The orchestrator is immutable application code, not a model-provided plan. It checks hashes of host-prepared inputs and referenced text/image files before delivery. The actual pinned app-server processes groups sequentially in one ephemeral thread, with one required successful inspection command per turn and matching thread/turn identity. It allows at most two targeted continuations **across the whole review**, including synthesis. Explicit missing mandatory or visual evidence cannot be cleared by a later model claim. For multiple groups, a separate final synthesis reconciles all unit results and cross-group behavior; it cannot waive unfinished units. A sole group instead receives an explicit final whole-PR/cross-file check in its own response, avoiding a redundant turn. Its verified evidence, inspection and completeness gates still apply. No model field such as “all groups inspected” controls scheduling or coverage.

One credential proxy and one engine timer cover all turns: the existing 100 inference-request / 20-minute ceiling is not multiplied per group. The same 24 context requests, 768 GitHub calls, text/image byte limits and model/effort selection remain in force. Protocol receipts and scope checks establish supplied evidence and completed scheduled responses, not proof of a model's semantic reasoning. More turns can change actual consumption within those ceilings; no real-provider cost or quality improvement is claimed from synthetic tests.
One credential proxy and one engine timer cover all turns: the existing 100 inference-request / 20-minute ceiling is not multiplied per group. The context socket allows 24 distinct uncached snapshot/path lookups and 100 total requests; the 768 GitHub-call, text/image-byte and model/effort limits remain in force. Protocol receipts and scope checks establish supplied evidence and completed scheduled responses, not proof of a model's semantic reasoning. More turns can change actual consumption within those ceilings; no real-provider cost or quality improvement is claimed from synthetic tests.

Valid incomplete results return to the trusted runner as data. It independently checks every expected unit ID/hash, bounded attempts and synthesis status against its original plan, drains context lookups, and rechecks host and model completeness. Incomplete work fails the Actions review job and may publish only COMMENT, preserving validated summary and findings. Invalid or quarantined output remains rejected. Later execution failures preserve earlier safe findings and the primary execution diagnosis; generic workflow failure does not overwrite it. Findings are deduplicated by original path/side/line and normalized title; more than 40 findings or other output loss is explicitly incomplete. The model cannot delete earlier valid findings during synthesis.

Expand Down
2 changes: 1 addition & 1 deletion docs/architecture.md
Original file line number Diff line number Diff line change
Expand Up @@ -39,7 +39,7 @@ The manifest distinguishes `diffComplete`, `requiredContextComplete` and supplem

## Isolated inference

Nonroot Docker containers use read-only root filesystems, resource caps, no capabilities, no-new-privileges and private temporary homes. The reviewer has network=none and only context/socket read-only mounts. It has no Docker socket, GitHub/OIDC token, API key, refresh token or vault key. A separate audited proxy owns the access credential, only permits stateless streaming Responses inference at one HTTPS route, rejects hosted network tools, pins model/effort/service-tier selection and replaces provider errors. Default limits are 100 requests and a 20-minute turn. The host owns CLI-session refresh and all context GitHub access. Approved hosted-session refresh stays on the Worker behind OIDC and leases.
Nonroot Docker containers use read-only root filesystems, resource caps, no capabilities, no-new-privileges and private temporary homes. The reviewer has network=none and only context/socket read-only mounts. It has no Docker socket, GitHub/OIDC token, API key, refresh token or vault key. A separate audited proxy owns the access credential, only permits stateless streaming Responses inference at one HTTPS route, rejects hosted network tools, pins model/effort/service-tier selection and replaces provider errors. Default inference limits are 100 provider requests and 20 minutes for the entire isolated execution, shared by all groups, continuations and synthesis. The host owns CLI-session refresh and all context GitHub access. Approved hosted-session refresh stays on the Worker behind OIDC and leases.

Codex 0.169.0 uses `approvalPolicy: never` for unattended inspection. The thread's read-only default remains, and each review turn explicitly selects `externalSandbox` with restricted network, relying on the unchanged Docker boundary. `never` is not a client approval grant: all server-initiated approval/tool/permission requests still receive JSON-RPC -32601 and cause the analysis to fail. Native declined commands also fail. At least one completed command with exit 0 is necessary before a turn can produce a review; this is execution evidence, not proof of full semantic coverage. Both the isolated entrypoint and trusted host validate completeness, exact head and changed-line references. Incomplete/quarantined/truncated/inconclusive output fails the review job; the protected publisher can still publish a non-approving fallback.

Expand Down