Agent navigation: question to owner

Start with the question, open the owning heading below, then follow its source contract if the answer needs implementation detail. Cite the exact heading and source you actually consumed; distinguish a current rule, historical observation and unknown deployment fact. This is a retrieval map, not another policy document or permission to operate a project.

Choose the kind of authority

Kind	Use it for	Not evidence of
AGENTS.md and explicit owner instructions	Development/storage/safety/workflow policy	Consent inferred from a design or passing test
System specification, UI needs, UI contracts	Required behavior and presentation	Measured implementation conformance
Runbook	Actual operating procedure and decision gates	Automatic authorization to execute it
Architecture, source/schema	Ownership and exact implementation contract	A replacement for an unmet requirement
Durable format reference	Current source baseline and bounded dated change evidence	Installation inventory, supported old reader or permission to normalize
Guides, overview, field lookup	Learning/tasks/explanation/current lookup	A competing normative policy
docs/working/	Local task plan/review/provenance, when assigned	Published authority or completed implementation

Autonomous work and completion

For orientation, read the objective-to-outcome explanation. For exact behavior, start with the applicable row rather than assuming every card has the same graph or that DONE certifies the software independently.

Exact question	Owning heading	Source/contract to inspect
Who owns the objective, scope and remaining authorized work?	Planning, assessment and current status; authored records	classic-shared.ts, selected Planner/node prompt closure, authored-record-files.ts
How do Planner delegation and Executor work reach actual execution?	Activation and node execution; caller-aware admission	supervisor-runtime-api.ts, card-process-actor.ts, agent-executor.ts
Where are review acceptance and parent correction distinct?	Agent path and shared acceptance; planning and correction	node-execution.ts, supervisor-runtime-api.ts, selected workflow edges and participant grants
When can the root finish rather than replan, request review or block?	Shipped project completion and profile graphs; completion assessment	projectCompletion, selected assess-completion prompt; no runtime natural-language objective checker
What do results and tests establish, and where are action results?	Understand the outcome; action execution and stdout	node-execution.ts, card-execution-log.ts, exact cited work products and checks

Selected prompts carry role/node policy; configured graphs and tool admission bound its effects. Public documentation is a retrieval aid, not another runtime instruction source. Use prompt selection to find the actual host instead of assuming the shipped text was delivered.

Analyst, MCP and lifecycle

Exact question	Owning heading	Source/contract to inspect
Can project Stop terminate the external Analyst listener?	Project Stop and application close; Calls, disconnect and lifecycle	server-services.ts, external-analyst-operations.ts, analyst-mcp-server.ts
How do I choose internal/external/both and a tool subset?	Inbound Analyst startup selection	resolveAnalystStartup; cli.ts
How does an external client connect, initialize, get context and discover tools?	Initialize, inspect context and discover tools	analyst-mcp-server.ts, real SDK fixture
Does disconnect cancel admitted work or permit replay?	Calls, disconnect and lifecycle	external-analyst-operations.ts, invokeTool
Where does external record attribution live?	Authorship is not a verdict; Analyst operations and external clients	operation-author.ts, record-author.ts, canonical-record-artifacts.ts
Are receipts a shared chat, complete audit or delivery proof?	Control events, notifications and external receipts	external-tool-call.ts, app-log contract
Is outbound MCP the same listener?	Saivage tools to configured MCP servers	mcp-manager.ts, server-definitions.ts
Restart connection, refresh catalog, or restart remote server?	Stop, replace, restart or refresh	server-runtime.ts, streamable-http-transport.ts, stdio-transport.ts
What are invocation deadlines, roots, native text/images and result limits?	Deadlines, roots and process ownership; Native results and image limits	server-runtime.ts, mcp-native-result.ts
Pause versus Resume versus Run/Stop/close/restart?	Choose the control	supervisor-runtime-api.ts, boot/app.ts, restart-port.ts
Which lock observation allows CLI delegation?	Five ownership outcomes	readRuntimeLockStatus, cli.ts

The inbound procedure is the source for separate bearer/loopback/project-digest binding, protocol2025-11-25, both Accept types and prohibited Origin/query/session/ batch/GET-stream. Do not turn its SDK example into a generic remote deployment recipe. The outbound pinned browser recipe is a different task and is not a live launch instruction merely because a reader found it.

Evidence and uncertainty

Exact question	Owning heading	Source/contract to inspect
Is document0/draft/accepted number the head revision?	Baseline0, drafts and accepted documents	authored-record-files.ts, canonical-record-artifacts.ts
Does a recorded image establish provider perception or current source pixels?	Current file versus recorded image; Explicit image observations	conversation-image.ts, tool-result-settlement.ts, openai-content-serializer.ts
Current loaded instructions, historical preparation, or exact provider request?	Prepared, current and historical instructions	activation-instructions.ts, agent-operator-read-model.ts, bootstrap-context.ts
Missing process output: running, completed or unavailable?	Process state and missing output	process-view.ts, operator-api-processes.ts
Where is action/hook evidence rather than a conversation?	Action execution and stdout	node-execution.ts, card-execution-log.ts, operator-api-executions.ts
Does a shell report author mean semantic approval?	Authorship is not a verdict	record-author.ts, node-execution.ts
Why did startup/provider/auth/context fail?	Diagnosis and its symptom heading	environment.ts, actual named failing consumer, not only a dashboard
May I retry/read/clean up after uncertain publication?	Publication uncertainty	publication-outcome.ts, replace-file.ts

Public evidence and sensitive private failed-request captures have different owners. Missing evidence stays unknown. No row authorizes secret disclosure, provider perception claims, or a universal “everything recorded” ledger.

Configuration, extension and retained state

For commissioning, first resolve command-specific inputs and browser access/setup. Before offline mutation, use the effective-input gate, not Start's selectors. Current workflow declarations use format3; the literal-key adoption boundary explains why historical workflow2 examples are not current inputs.

Exact question	Owning heading	Source/contract to inspect
Which YAML/default/capability is current?	Configuration reference, grouped field heading	saivage-config.ts, resolved-config-authority.ts, provider.ts
How do I add an action and validate it?	Add an action without inventing a plugin service; Change a workflow	operation-contract.ts, shell-operation.ts, workflow-action-compilation.test.ts
Where do tool grants/binding/projection belong?	Add or change a tool at its existing binder	runtime-tool-catalog.ts, invocation.ts, owning provider/projection
Where should reusable prompt method versus task delta live?	Author prompts at the effective delivery owner	prompt-api.ts, AgentExecutor; mandatory prompt-authoring skill
Can a previous selector repair old-format state?	Choose the operation; Current selections, previous slots and orphans	repair-target.ts, publish-head.ts, exact current schema
Which source owns current conversation format, and which adoption is permitted?	Current source baseline; Explicit external offline migration	canonical-conversation-artifacts.ts, conversation-file.ts, AGENTS Storage Policy
Upgrade failure, current damage or intentional reset?	Choose the safe next task, then state-adoption decision	Exact source/target contracts and named failing consumer; readiness alone does not certify compatibility
Which check and CI evidence applies?	Choose validation and deliver the documentation, Validation	package.json, validation workflow

For adoption questions, retrieve the complete selected procedure, not only a table row: current-format repair is not migration; reset has separate destructive consent; external migration needs the explicit identified-project request and full offline backup/candidate/validation/fidelity/loss/cutover gates. Start and Run remain separate decisions. Historical equal-version incompatibilities remain historical, not implicit normalization permission.

Retrieve enough, not everything

Use local search or file search for the exact heading/symbol. Read its surrounding contract, not a matching line alone; follow any required precondition link before an action. If source and a requirement genuinely disagree, report the conflict instead of weakening the requirement or improvising a product fix. Record the pages/source retrieved and evidence limits when handing off an answer.

Public documentation is not automatically loaded into runtime agents. A static composition or current-instructions view does not prove delivery or understanding. Host development instructions, shipped role prompts, target project overrides and conversation evidence are distinct inputs with distinct owners.