Run Ledger adds execution continuity to Gate. It seals checkpoints at safe boundaries, preserves tracked run state across interruption, and blocks replay of side effects that already happened.
npm install zehrava-gateShipped in zehrava-gate@0.3.0 with CLI commands and SDK methods.
Not at the beginning. Right after approval. Right after execution request. Right after the side effect. That is where replay becomes dangerous.
This is the entire point of Run Ledger: continue from the last safe boundary instead of rebuilding from transcript and hoping nothing fires twice.
Agent runs do not fail cleanly. They fail mid-approval, mid-execution, mid-retry, or after the side effect but before the receipt. Starting over from the top is not recovery. It is replay with worse odds.
Agent enriches 847 leads and calls Salesforce. Connection drops halfway. Agent restarts. No checkpoint. Tries again — same 847 records, some already imported.
Replayed writeAgent submits high-risk write. Waits for human approval. Worker crashes. Approval comes through. No record of where the run was. Agent restarts from the top.
Lost run stateExecution order issued. Worker fetches it. Calls Stripe. Payment succeeds. Worker crashes before reporting. Restart → no receipt. Did it execute? Try again?
Unknown execution stateAgent halfway through a multi-step CRM sync. Deploy rolls out. Process terminates. Agent restarts. No saved state. Starts over. Some records already written.
No checkpointAgent calls webhook. Times out. No response. Retries. Webhook actually fired twice — first one just took 90 seconds. Now two notifications sent.
Duplicate side effectAgent crashes. Transcript has 200 messages. Replay from the top? Which tool calls already executed? Which approvals already granted? Transcript doesn't know.
Transcript ≠ execution historyTranscript replay is not execution continuity.
A transcript can tell you what the agent said. It cannot prove what executed, what was approved, what artifact was created, or which side effect must not run again. Run Ledger records receipts. That is the difference.
Four components. Ordered event history, sealed resumable state, continue from last safe point, block replayed side effects. That's it.
Ordered event history for a tracked run. Every meaningful boundary is recorded: plan locked, tool finished, intent proposed, policy checked, approval received, execution requested, execution succeeded, interruption detected.
Sealed resumable state captured at a safe boundary. Includes receipts, artifacts, unresolved approvals, blocked side-effect keys, and suggested next action.
Load the latest valid checkpoint, verify integrity, rebuild resume context, and continue from the current step instead of from the beginning.
External mutations, payments, and notifications get side-effect keys. If a key already exists in the ledger, the replay is skipped.
No workflow engine. No transcript replay. No UI-heavy abstraction layer. Just continuity.
Run Ledger is narrower than workflow orchestration. It lives at the governed write boundary inside Gate: approvals, execution boundaries, checkpoint integrity, and non-replayable side effects.
| Feature | Temporal | LangGraph | Restate | Inngest | Zehrava |
|---|---|---|---|---|---|
| Durable execution | ✓ | ✓ | ✓ | ✓ | ✓ |
| Agent-aware | — | ✓ | — | — | ✓ |
| Approval gates | — | — | — | — | ✓ |
| Side-effect dedupe | — | Graph nodes | — | Steps | Receipts |
| Governed writes | — | — | — | — | ✓ |
Execution continuity. Sealed checkpoints. Side-effect deduplication. CLI inspection. Opt-in integration — backward compatible with existing Gate deployments.
Resume interrupted tracked runs from the last safe checkpoint.
Sealed checkpoints fail verification if tampered with.
Skip non-replayable side effects that already happened.
Inspect, verify, checkpoint, and resume from terminal.
Opt-in per run. Existing Gate flows still work.
Intent and execution lifecycle recorded in one run ledger.
Run Ledger is live in the package. Start a tracked run, record execution boundaries, seal a checkpoint, and resume from the last safe point.
import { Gate } from "zehrava-gate" const gate = new Gate({ endpoint: "https://gate.yourco.com", apiKey: "gate_sk_..." }) const run = await gate.startRun({ agentId: "enrichment-agent", intentSummary: "Enrich leads and sync approved changes to Salesforce" }) await gate.recordEvent({ runId: run.runId, eventType: "plan_locked", actorId: "enrichment-agent", stepName: "enrich", payload: { steps: ["fetch", "enrich", "review", "sync"] }, sideEffectClass: "none" }) const intent = await gate.propose({ runId: run.runId, payload: "./leads.csv", destination: "salesforce.import", policy: "crm-low-risk", recordCount: 847, idempotency_key: "batch-2026-03-22" }) // interruption happens here await gate.createCheckpoint({ runId: run.runId, reason: "interruption", suggestedNextAction: "resume_then_execute" }) const ctx = await gate.resumeRun({ runId: run.runId }) if (!ctx.blockedSideEffectKeys.has("salesforce.import:batch-2026-03-22")) { await gate.execute({ intentId: intent.intentId }) }
Tracked runs attach continuity to the real Gate lifecycle: propose, approve, execute, report.
No dashboard in v1. Everything via CLI. Inspect run status, list events, verify checkpoint integrity, resume from last safe point.
The CLI is the proof surface in v1. If a run cannot be inspected, verified, and resumed from terminal, the system is not trustworthy yet.
View pending approvals, audit trails, and run history in the web UI. No terminal required.
The Dashboard provides a visual interface for monitoring tracked runs, inspecting checkpoints, and reviewing execution history. Approved intents, blocked side effects, and resumable runs — all visible without touching the CLI.
Open Dashboard →These are not bugs. These are conscious decisions about what ships in v1 versus what waits for v2. If you need one of these, self-host and build it — or wait for the roadmap.
Resume packets are Zehrava-specific.
Checkpoints are local SQLite. Multi-node needs external coordination.
Failed verification requires operator investigation. No auto-repair.
Old runs accumulate. Operator cleans up manually.
Events recorded individually. Rare partial state possible.
CLI-only inspect/resume. Dashboard ships in v2.
Checkpoint triggers are manual or event-based.
Parent/child integrity checks are roadmap.
Self-host free. MIT license. Full docs, SDK, CLI, and integration guide available now.