Agent Upstart

Cool things people are building with agents — and practical help getting the most out of them.

Playbooks · Coding tasks

Codex Guardian v3: the evidence pipeline in 0.168

Written November 9, 2026

Codex Guardian v3's evidence pipeline — snapshots, cursors, housekeeping — merged Nov 5 for 0.168.0. What it records, and why auto-review needs it.

The problem it solves

A team enabling Codex auto-review needs review decisions it can audit. Guardian v2, which shipped in Codex CLI 0.165.0, gave the reviewer a separate model context that evaluates actions before they run — but the reviewer consumed the live history in a single sweep, with no durable snapshot, no incremental tracking, and no routing by evidence type. If the reviewer missed something, there was no replay; if the evidence stream was large, one pass could exceed the reviewer's model budget. Four pull requests merged on November 5, 2026 build the pipeline that fixes that: capture, reading, and routing, plus a second chance for denied actions.

What merged on November 5

The pipeline has three layers, following a local message-board model: evidence is written to topic channels, and the reviewer reads from those channels at its own pace instead of receiving history directly from the main runtime. The main runtime is never blocked waiting for the reviewer.

Capture — EvidenceSnapshot (PR #59669). When Guardian v3 needs to evaluate an action, it takes an immutable snapshot of the original history at that moment: messages, tool activity, and approvals as they existed when the snapshot was taken. System-injected content, compaction artifacts, and synthetic context are stripped before snapshotting, so the reviewer sees only what came from the model or the user. Each snapshot records the ordinal position of its contents in the root history, and capture refuses forked or reverted histories — the reviewer never evaluates evidence from a path the session did not actually take.

Reading — EvidenceCursor (PR #59670). The cursor lets the reviewer consume evidence incrementally instead of in one pass. A new cursor starts at the tail of the stream and does not backfill earlier history; it tracks which items the reviewer has explicitly acknowledged, supports partial acknowledgements, and reading does not advance the cursor — only acknowledgements do. A reviewer can reread the same items to cross-reference a tool call against an earlier approval before deciding.

Routing — Housekeeper (PR #59671). The Housekeeper sits between capture and the message board, routing snapshots into the right topic channels. It runs in an isolated reviewer runtime, batches evidence to stay within the model's tool-output budget, validates a routing decision before writing, and only advances the cursor after every write in a batch has succeeded. A failed batch is retried on the next cycle rather than dropped — at-least-once delivery, with duplicates possible under failure instead of silent loss.

Denial reconsideration (PR #59659). A terminal denial — a final "do not proceed" — used to be final. Behind the guardian_denial_reconsideration flag, the first terminal denial triggers a single retry using the exact original inference request, with the model's provisional output buffered while the retry runs. The old experimental_denial_reconsideration_prompt key is removed: the retry uses the original request directly, which closes a prompt-injection surface on the reconsideration path.

How to enable it

Guardian v3 is feature-flagged and disabled by default. In config.toml:

[features]
guardianv3 = true

The elastic context-budgeting work — validated section budgets, whole-entry allocation, capped pre-allocation — ships in 0.167.0 stable and applies automatically once Guardian v3 is enabled. Denial reconsideration is a separate flag, guardian_denial_reconsideration = true. Review timeout defaults to 90 seconds (timeout_ms = 90000 under [auto_review]); setting it to 0 removes the wall-clock deadline only, not cancellation or retry limits.

Two scope facts matter for a team plan. Guardian v3 requires the OpenAI provider: Bedrock and LiteLLM sessions are excluded from native multi-agent history. And the reviewer's context includes the repository's AGENTS.md as part of the immutable snapshot — guidance written there is in scope for every review, not injected on the fly.

What is actually in stable

None of the evidence pipeline is in 0.167.0 stable, released November 6. The four pull requests merged after the 0.167.0 stable cut (highest included PR: #58757); 0.167.0 carries only the Guardian v3 elastic budgeting work (PRs #58123–#58454). The pipeline lands in 0.168.0, which is alpha-only: 0.168.0-alpha.7 was published November 5 with release notes that say nothing beyond the version number. Watch the 0.168.0 stable release for the first stable exposure of the pipeline.

Verdict

Wait for 0.168.0 stable before relying on the evidence pipeline. The parts that make auto-review auditable — immutable snapshots, acknowledged incremental reads, routed delivery — are merged but alpha-only, and the alpha release notes describe none of them. The 0.167.0 stable release is still the right base for production Guardian v3, with the pipeline as the reason to upgrade when 0.168.0 lands, not before.

Limits

Related playbooks

Sources

Back to: Coding playbooks hub · Also: Codex 0.167: run parallel sessions in one place · OpenAI: agents tried to escape their sandboxes · Decide whether Copilot should approve PRs