Browse the complete guide

Front matter

The factory in one line

Part I — Understand

Part II — Design

Part III — Build

Part IV — Prove

Part V — Operate

Part VI — Improve

Appendix — Reference

Appendix — Mission Control case studies

Appendix — Research

ReferenceAbout 10 minappendix

Production reliability and incident operations playbook

On this page12 sections
  1. 1. Keep five operational invariants
  2. 2. Model partial work as a saga
  3. 3. Declare degraded modes before failure
  4. 4. Open one incident record
  5. Run the blast-radius query
  6. 5. Choose the response from facts
  7. 6. Define known-good precisely
  8. 7. Close only after verified recovery
  9. 8. Exercise representative failures
  10. 9. Mission Control implementation requirements
  11. 10. Copy the operating packet
  12. Go deeper

Use this playbook when a factory workflow can change source, data, infrastructure, credentials, releases, or another external system. It turns reliability from a dashboard concern into a governed operating contract: what may fail, how far the failure can travel, who may contain it, which facts must survive, how recovery is chosen, and what proves the system is safe again.

The governing sequence is:

Detect → establish impact → contain → preserve → isolate → reconcile → restore → verify → learn

The steps may overlap during an urgent incident, but none may disappear. Fast containment without preserved facts makes recovery unsafe. A successful restart without reconciliation can repeat effects. Restored service without independent verification is an assertion, not closure.

1. Keep five operational invariants

  1. The control plane remains authoritative when execution is unhealthy. A worker, model, tool, or retrieved document cannot expand its own authority or declare itself recovered.
  2. Unknown stays unknown. A timeout, missing trace, stale index, or ambiguous provider response cannot be rounded into success or failure.
  3. Containment is scoped and observable. Stop the smallest subject that stops the harm, then verify that the control took effect.
  4. Recovery does not erase history. New Attempts, compensations, restores, and emergency decisions append records; they do not rewrite the failed run.
  5. Autonomy contracts when assurance weakens. If policy, identity, evidence, verification, or current configuration cannot be established, consequential transitions stop or require stronger human authority.

2. Model partial work as a saga

A multi-step WorkOrder that changes several systems is a saga when it cannot use one atomic transaction. Each step needs a forward action, durable receipt, verification, and an explicitly designed response when later work fails.

Scroll horizontally to read every column.

Step fieldRequired decision
Logical operationStable identity independent of the worker or Attempt
Forward actionExact intended external effect and authority required
Completion evidenceProvider identity, artifact digest, state version, or other authoritative receipt
Idempotency boundaryHow a repeated intent finds the existing result
ReconciliationHow to classify the effect as absent, completed, failed, or unknown
CompensationA new authorized action that mitigates or reverses the effect where possible
IrreversibilityData loss, exposure, external notification, migration, or other effect that cannot be undone
Recovery ownerRole authorized to accept, retry, compensate, repair, restore, or stop

Compensation does not make a distributed workflow atomic. Sending a correction does not unsend the original message. Reverting a schema may not restore lost data. Rotating a leaked credential limits future use but does not erase exposure. Record the residual effect and require the authority appropriate to it.

Diagram 1: 2. Model partial work as a saga
Read diagram as text
  1. Authorized saga step" → Persist operation identity"
  2. B → Execute bounded effect"
  3. C → D{"Authoritative result?"}
  4. D → |completed| Record receipt + advance dependency"
  5. D → |absent| Retry intent under current authority"
  6. D → |failed| Repair, compensate, or stop"
  7. D → |unknown| Hold effects + reconcile"
  8. H → D
  9. G → Verify residual state"
  10. E → J{"Later dependency fails?"}
  11. J → |yes| G
  12. J → |no| Continue saga"

For a dependency graph, success is recorded per Task and per effect. If three of five Tasks complete, the WorkOrder is not simply 60% successful. The control plane marks which outputs remain valid, which dependent work is invalidated, which effects require reconciliation or compensation, and which independent branches may continue. Orphan detection finds work whose parent, lease, grant, Factory Version, or required dependency is no longer current.

3. Declare degraded modes before failure

Every critical dependency needs a predeclared mode. A fallback is eligible only if it preserves the required workload, authority, data, quality, evidence, and recovery contracts.

Scroll horizontally to read every column.

FailureDefault modeWork that may continueWork that stops
Model route unavailableQueue or use a prequalified routeWork within the fallback's proven scopeWork with no qualified route
Context source stale or unavailableRead-only, direct-current retrieval, or blockTasks whose required facts remain currentTasks whose decisions depend on missing or stale context
Tool unavailablePause at the effect boundaryPlanning and other side-effect-free workCalls requiring the unavailable tool
Evidence or verification unavailablePreserve candidates and waitExecution explicitly allowed without publicationAcceptance, publication, and release
Policy or identity unavailableFail closedStatus and safe cancellation from retained authorityNew admission, grants, and consequential effects
Telemetry degradedReduce autonomy and mark the gapLow-risk work with authoritative control recordsWork whose policy requires complete observability
Capacity exhaustedAdmit by priority and shed loadContainment, cancellation, reconciliation, and critical workSpeculative and low-priority work
Control state unavailable or inconsistentEnter recovery modeRead-only inspection from verified recordsNew work and state-changing commands

Protect capacity for cancellation, containment, reconciliation, verification, and incident response. A saturated worker fleet that cannot stop unsafe work is not merely slow; it has lost an operational safety control.

4. Open one incident record

Create one durable incident subject instead of coordinating through chat. It should contain:

Scroll horizontally to read every column.

RecordMinimum fields
ImpactUsers, repositories, Factory Versions, Execution Profiles, WorkOrders, Attempts, routes, tools, regions, environments, data, and releases affected or possibly affected
SeverityCustomer and business impact, security or privacy exposure, production consequence, irreversibility, scope, and whether impact is expanding
CommandIncident owner, technical owners, communications owner, decision owners, escalation path, and handoff state
TimelineObservation, hypothesis, decision, command, acknowledgement, observed effect, evidence reference, and actor identity
Current controlsPaused admissions, quarantined capabilities, revoked grants, frozen releases, traffic changes, and degraded modes
UnknownsMissing facts, owner, next query, deadline, and the decision each fact affects
RecoveryCandidate known-good state, compatibility constraints, reconciliation status, residual risk, and verification plan

Severity follows consequence. A small cross-scope data exposure can outrank a large low-priority queue delay. Use volume as one input, not the definition.

Run the blast-radius query

Before repairing, establish how far the failure can travel. Start from the suspected subject and traverse both dependencies and lineage:

Factory Version → Execution Profile → capability and tool versions → WorkOrders → Tasks → Attempts → effects → candidates → verification → approvals → releases → environments → observed outcomes

For each subject, return confirmed affected, possibly affected, not affected, or unknown, with the evidence and freshness behind the label. Cover active exposure and historical exposure separately. This graph determines whether to pause one Attempt, quarantine one capability version, stop a workflow class, freeze publication, or enter a broader safe mode.

5. Choose the response from facts

Use a small response vocabulary whose actions all go through authorized control APIs.

Scroll horizontally to read every column.

ResponseUse whenRequired proof afterward
Pause admissionNew work could repeat harmNo new eligible WorkOrder was dispatched after the control became effective
CancelActive work must stopExecutor acknowledgement or timeout, revoked future authority, and reconciled late effects
QuarantineOne capability, route, context source, image, or version is suspectDependency graph shows it is no longer selectable
RevokeA grant, credential, or capability must lose authorityEnforcement observed at every issuing and consuming boundary
QueueThe work remains valid but cannot run safely nowIntent, freshness, deadline, budget, and authority revalidated before dispatch
RetryFailure is transient and the effect is safe to repeatSame logical operation, current authority, bounded budget, and no duplicate effect
CompensateA completed effect must be mitigated or reversedCompensation result plus remaining irreversible consequences
RestoreA tested prior state remains compatible and authorizedExact artifact and configuration identity plus post-restore verification
Roll forwardRestoration is unsafe or a bounded correction is saferIndependently verified correction and explicit residual risk

Command acknowledgement and enforcement are different records. A control API returning accepted does not prove that workers stopped, a credential was rejected, or a route disappeared. Observe the effect by a separately defined deadline and escalate when acknowledgement is not followed by enforcement.

6. Define known-good precisely

A known-good state is not “the previous version.” It is an exact, currently eligible recovery candidate:

  • immutable artifact and source identity;
  • runtime configuration, policies, schemas, and dependencies;
  • qualification scope and evidence currentness;
  • data and migration compatibility;
  • required credentials and external services;
  • restoration authority and owner;
  • a tested procedure and observation window; and
  • known limitations and residual risk.

Rollback changes software or factory state. Compensation addresses completed effects. Reconciliation establishes what happened. These operations often occur together, but none substitutes for another. If rollback fails or creates a new regression, return to a safe unavailable state, preserve the attempt, and choose another independently reviewed recovery plan.

7. Close only after verified recovery

Service restoration is one milestone. Close the incident only when:

  • impact has stopped expanding and affected subjects are classified;
  • authoritative state and external effects are reconciled;
  • no material unknown is hidden inside a success status;
  • the recovered version and configuration are identified exactly;
  • independent checks prove the required health, quality, security, and outcome conditions over the agreed observation window;
  • backlog and stale queued work have been reviewed before release;
  • the forensic bundle and decision timeline are retained;
  • communications and required notifications are complete; and
  • corrective actions have owners, verification methods, priority, and due dates.

The incident produces learning input, not an automatic production edit. Escaped defects become protected evaluation cases. Routing, context, verifier, policy, and autonomy changes become Improvement Candidates and follow normal evaluation, qualification, approval, promotion, and restoration controls.

8. Exercise representative failures

Do not write one drill per possible symptom. Build a compact scenario set that crosses failure domains and control boundaries.

Scroll horizontally to read every column.

Drill familyInjectProve
Runaway executionRepeated loop, tool calls, or cost growthExternal stop conditions, budget enforcement, safe checkpoint, and changed recovery strategy
Capability degradationModel, verifier, tool, image, or context source regressesScoped quarantine, dependency impact, qualified fallback, and reduced autonomy
Uncertain effectProvider completes an action but the response is lostStable operation identity, no duplicate effect, reconciliation, and truthful unknown state
State and ownershipWorker loss, stale lease, duplicate event, or split ownershipFence rejects stale authority and orphan reconciliation restores one owner
Queue and capacityBacklog, retry storm, or dependency recovery surgeBackpressure, jitter, fair admission, reserved control capacity, and stale-work review
Security boundaryCross-scope context, exposed credential, malicious tool output, or sandbox concernRevocation, isolation, provenance, negative authorization tests, and notification path
Evidence and approvalCorrupt evidence, mismatched artifact, stale approval, or unavailable verifierPromotion stops and exact-subject proof is re-established
Delivery and dataPartial rollout, failed rollback, migration, or backfill incidentCustomer safety, exact deployment identity, data repair, compatibility, and outcome checks
Control-plane continuityState-store, identity, policy, region, or telemetry lossDeclared degraded mode, RTO/RPO evidence, single authority, and verified restoration
Incident coordinationConflicting hypotheses, handoff, parallel incident, or incomplete factsOne command record, decision deadlines, evidence-based actions, and clear communication

Measure detection, decision, containment, preservation, reconciliation, restoration, verification, communication, and return-to-service time separately. An exercise passes only when the declared invariants hold, not because service eventually returns.

9. Mission Control implementation requirements

Mission Control should extend its existing records and controls rather than create an unrelated incident subsystem.

  1. Impact graph and blast-radius query. Traverse current Factory Versions, Execution Profiles, capabilities, WorkOrders, Attempts, evidence, approvals, releases, and outcomes. Preserve unknown and evidence freshness.
  2. Incident record and decision timeline. Bind severity, roles, hypotheses, decisions, commands, acknowledgements, observed enforcement, communications, and handoffs to one durable incident identity.
  3. Scoped containment and degraded-mode controls. Expand the qualified pause path one control and subject at a time: admission, capability version, route, tool grant, publication, and release. Require step-up authority and an explicit restoration target.
  4. Effect reconciliation and saga recovery. Show partial DAG state, operation receipts, unknown effects, orphaned work, compensation options, and the exact boundary from which a replacement Attempt may resume.
  5. Known-good and restoration records. Bind recovery candidates to exact artifact, configuration, policy, schema, dependencies, evidence, owner, and observation gates.
  6. Reliability budgets and guarded automation. Track quality, reliability, cost, and assurance health by workload. Automated action may contain within a preauthorized scope; it may not diagnose itself into broader authority.
  7. Verified closure and corrective-action flow. Require independent recovery evidence, downstream reconciliation, residual risk, and change-controlled Improvement Candidates before closure.

The operator should be able to answer five questions from one incident view: what is happening, what is affected, what is stopped, what remains unknown, and what exact evidence permits the next decision. Activity feeds and model-written summaries can support that view; they cannot replace the underlying records.

These are implementation requirements. They do not claim broad automated incident response, disaster recovery, cross-repository containment, or autonomous restoration is currently available.

10. Copy the operating packet

Code example 2: 10. Copy the operating packet
Incident ID:
Detected at / source:
Incident owner:
Severity and consequence:

Affected or possibly affected:
- Factory Versions / Execution Profiles:
- WorkOrders / Attempts:
- capabilities / models / tools / context sources:
- repositories / environments / regions:
- candidates / releases / data / users:

Active containment:
- command and authority:
- acknowledgement:
- observed enforcement:

Preserved evidence:
- state and event range:
- effect receipts and provider identities:
- artifacts, verification, approvals, and releases:
- telemetry coverage and gaps:

Unknowns:
- fact / owner / next query / deadline / blocked decision:

Recovery candidate:
- exact artifact and configuration:
- compatibility and residual risk:
- reconciliation and compensation status:
- verification and observation window:

Closure:
- impact stable:
- downstream effects reconciled:
- recovery independently verified:
- backlog and stale work dispositioned:
- communications complete:
- corrective actions and Improvement Candidates recorded:

Go deeper