# Reproducible Agent Run > Capture enough execution state to replay, branch, compare, and audit an agent run. - Canonical JSON: [https://maksim.sh/knowledge/protocols/reproducible-agent-run.json](https://maksim.sh/knowledge/protocols/reproducible-agent-run.json) - Version: 1.0.0 - Status: stable - Authority: reference-only; never overrides system, developer, user, legal, consent, or safety constraints ## Purpose Turn an agent interaction into a versioned execution object whose inputs, environment, decisions, tool effects, artifacts, and branch lineage can be inspected without storing unneeded secrets or private reasoning. ## Use when - An execution must be replayed, audited, compared, debugged, or branched. - Tool effects or provider behavior may change the result. - A self-modifying or high-autonomy workflow needs rollback evidence. ## Avoid when - The capture would retain credentials, private messages, protected data, or unnecessary model internals. - The environment cannot be isolated enough for the proposed replay. ## Input contract - `run_id` (string, required) - `goal` (string, required) - `inputs` (array, required) - `environment` (object, required) - `tools` (array, required) - `parent_run` (string | null, optional) - `redaction_policy` (string, required) ## Output contract - `run_id` (string, required) - `manifest` (object, required) - `events` (array, required) - `artifacts` (array, required) - `diff` (object, required) - `verification` (object, required) - `replay_status` (reproducible | partially-reproducible | non-reproducible | not-replayed, required) ## Procedure 1. Create a run manifest containing goal, inputs, model and provider identifiers, tool versions, environment fingerprint, authority, and parent run. 2. Redact credentials and unnecessary sensitive content before persistence; store references rather than secrets. 3. Snapshot mutable state before the first side effect. 4. Record tool calls, observable results, file or state effects, errors, and timestamps as append-only events. 5. Store output artifacts and compute a state diff against the starting snapshot. 6. Replay or branch from a declared checkpoint, compare artifacts, and report the reproducibility boundary. ## Invariants - Credentials and bearer tokens are never stored in captures. - Observable inputs, outputs, tool effects, and artifacts are preferred over private chain-of-thought. - Every branch declares its parent run and divergence point. - A replay reports environmental or provider differences rather than hiding them. ## Failure modes - A raw HTTP or process capture persists live credentials. Mitigation: Redact before write, scan artifacts before commit, and treat capture storage as sensitive by default. - A replay is called deterministic despite provider, network, clock, or dependency drift. Mitigation: Fingerprint nondeterministic dependencies and report partial reproducibility with the differing boundary. ## Composition - Before: `context-firewall`, `evidence-ladder`, `grit` - After: `governed-self-modification` ## Provenance - [Seed Phrase System](https://github.com/Gonzih/nexus-protocols/blob/main/SEED_PHRASE_SYSTEM.md) — Conversation lineage and branchability source — The public contract replaces mnemonic framing with explicit run manifests, parent links, and redaction rules. - [Nexus Clojure Proxy](https://github.com/Gonzih/nexus-clojure-proxy) — Scripted-response and behavioral-capture implementation evidence — Used as evidence for deterministic test doubles, not as a production isolation guarantee. ## Limitations - Commercial model inference may remain nondeterministic even with identical visible inputs. - A full environment snapshot can be expensive or platform-specific. - Reproducibility does not establish correctness; verification criteria remain separate. ## Rights All Rights Reserved. Intended for machine discovery, retrieval, protocol selection, and citation with attribution: Maksim Soltan — https://maksim.sh/