Agent Workflow Proof

The workflow-proof example is an explicit, controlled comparison for three small synthetic, project-safe scenarios. It is not a normal tree-ring subcommand, background process, CI job, or universal benchmark.

Run it deliberately from a source checkout with a locally available Codex CLI:

cargo run --locked -p tree-ring-memory-cli --example workflow_proof -- \
  fixtures/workflow-proof target/tree-ring-certification/workflow-proof \
  --model <model-id>

--model <model-id> is required for an evidence-producing run; an omitted or blank value is a usage error before a Codex adapter is constructed. The adapter also validates its model and always passes that recorded ID to Codex. The report records the requested value as agent_identity: "codex:<model-id>" in JSON and as an agent-identity line in the Markdown summary.

The command runs the same task in three retained workspaces for every fixture:

• no_memory receives no memory context.
• raw_memory receives the visible normal, non-superseded seed memories in fixture order.
• tree_ring receives the normal, non-superseded memories returned by local Tree Ring retrieval.

The fixture pack exercises constraint recall (no-background-writer), current rules over superseded rules (stale-cli-contract), and a failure scar changing the recovery decision (scar-recovery). Each workspace materializes a decision-format.json action enum. The agent task asks it to inspect that workspace and write decision.json with an action and rationale; it does not name the expected action. Deterministic validators remain outside the agent request and compare only the exact decision.json /action value.

Evidence and Reproducibility

The selected output directory retains all trial workspaces at trials/<scenario>/<arm>/workspace/, plus a machine-readable workflow-proof-report.json and a readable workflow-proof-summary.md. For evidence integrity, artifacts use the resolved output directory, and descriptor-relative validation accepts only regular unlinked workspace files, rejecting symlink and hard-link outputs; non-Unix evaluation currently fails closed until an equivalent descriptor-safe implementation exists. Treat workflow-proof-report.json as observed paired evidence for these specific controlled fixtures: inspect the retained workspaces, memory context, agent response, and deterministic JSON field checks before drawing a conclusion.

For every run, record alongside the output:

• the Tree Ring commit (git rev-parse HEAD);
• the Codex CLI version (codex --version) and the required --model value;
• the agent_identity recorded in both reports (normally codex:<model-id> for this example);
• the complete command, timestamp, and any non-default Codex binary path.

No unit test, normal certification command, or CI job invokes Codex automatically. A real model run happens only when an operator explicitly runs the example above, and a failed control arm remains evidence rather than a runner failure.

This pack does not establish a universal model score, a general claim that memory improves all workflows, or a replacement for external evaluations. The next validation step is to run external benchmark adapters and preserve their native reports beside this controlled evidence.