Skip to main content
Status: draft - recovered from Donna’s donna-rescue-uncommitted-2026-05-08T215800Z stash for Candi governance review. Treat as docs-mirror placeholder pending typed-artifact promotion under SPEC-078.

​
method.agent-operating-doctrine

schema_version: 1
version: 2
id: method.agent-operating-doctrine
title: "Prism Agent Operating Doctrine"
status: draft
source_of_record: prism_typed_artifact
generated_mirror_paths:
  - docs/method-fragments/method.agent-operating-doctrine.md
authority_layer: project
owning_specialization: governance
owner_identity: Candi
applies_to_specializations:
  - engineering
  - install
  - rte
  - dashboard
  - architecture
  - governance
  - docs
  - security
applies_to_surfaces:
  - claude_code
  - codex
  - claude_desktop
  - cursor
  - vscode
use_when:
  - "A new or cross-machine agent starts work on a Prism-tracked project."
  - "An agent is about to solve a local symptom without first checking recorded Prism context."
  - "A task touches methodology, lifecycle, memory, signaling, architecture, incident handling, or cross-agent handoff."
  - "An agent is preparing to mark work complete, hand off work, or close a session."
skip_when:
  - "The turn is a maintenance tick and prism_signals_pending returns no work."
  - "The operator explicitly requests a narrow read-only command output with no Prism project work."
  - "A narrower method fragment governs the exact case and does not conflict with this doctrine."
risk_tier_default: R1
risk_tier_override_allowed: true
risk_tier_escalation_triggers:
  - "Work affects architecture, governance, methodology, lifecycle, memory, or cross-surface behavior."
  - "Work changes shared runtime behavior or another lane's owned files."
  - "Work involves an incident, RCA, postmortem, migration, deployment, or session/signal failure."
required_context:
  - rule
  - decision
  - task_state
  - prior_postmortems
expected_outputs:
  - "Relevant Prism context retrieved before acting."
  - "Materially used recall results cited with prism_memory_cite when citation is ergonomic on the active surface."
  - "Source-of-truth architecture or governance artifact identified when the task depends on it."
  - "Peer review or lane-owner signal for R2/R3 work."
  - "Incident/RCA artifact when an incident-class failure is named."
  - "Completion evidence recorded before work is declared done."
required_prism_verbs:
  - prism_start
  - hd_exec
  - prism_signals_pending
  - prism_signal_ack
  - prism_signal
  - semantic_recall
  - prism_governance_lookup
  - prism_note
  - prism_checkpoint
  - prism_wrap
review_gates:
  - governance
review_required: true
memory_writes:
  - task_state
  - validation_result
  - incident_rca
conflict_policy: "Narrower active method fragments win for their exact trigger. Architecture conflicts route to Texi; methodology/process conflicts route to Candi; implementation feasibility conflicts route to Donna."
telemetry_key: method.agent-operating-doctrine
supersedes: []
superseded_by: []
source_refs:
  - AGENTS.md v2.4.1
  - PRISM.md Session Protocol
  - SPEC-021 Session Bootstrap Enforcement
  - SPEC-041 Lifecycle vs Messaging
  - SPEC-063 RCA persistence rule
  - SPEC-076 Citation-Discipline Intervention
  - SPEC-078 Method Fragments
  - docs/research/memory-learning-loop-acceptance-checklist-2026-05-14.md
  - Journal #195 A+B memory learning-loop baseline rerun

​
Purpose

Train a new or cross-machine agent to operate inside Prism methodology before it starts solving local symptoms. This fragment is broader than a lane checklist. It covers the behaviors that must activate on every Prism-tracked project: memory-first work, architecture-first judgment, peer review, incident capture, completion discipline, and proper use of Prism verbs.

​
Operating Rule

Prism work starts from recorded context and source-of-truth architecture, not from a convenient guess. If an agent is missing context, it retrieves context. If the architecture is unclear, it asks the owning peer or records a decision path. If the work exposes an incident, it records the incident immediately. If completion cannot be proven, it is not done.

​
Bootstrap and Signal Discipline

Every substantive session starts with:
  1. Read local methodology files, especially AGENTS.md or CLAUDE.md and PRISM.md.
  2. Call prism_start or hd_exec.
  3. Display the Prism status card on substantive responses.
  4. Drain pending signals when required.
  5. ACK and act on non-empty signal drains before unrelated work.
Signals are work. A drained signal is not background noise once observed.

​
Memory First

Before guessing, retrieve prior state. Use the narrowest relevant retrieval tool:
  • semantic_recall for prior incidents, plans, decisions, RCA, and method notes
  • prism_diff for recent deltas and exact persistence checks
  • prism_governance_lookup for active governance constraints
  • prism_method_fragment_recall for methodology guidance when available
  • prism_memory_cite when a memory result materially changes or supports the work
Use distinctive content terms, not only UUIDs. If recall fails by id, query by meaningful terms from the artifact. Citation discipline is advisory until Recallgate precision has a real traffic window. When semantic_recall returns a recall_event_id, cite a recalled memory if it materially supports:
  • an answer to the operator or another agent
  • a decision, ADR, SPEC, plan, governance change, or method change
  • an incident, postmortem, RCA, or safety/security judgment
  • an implementation plan, review finding, or release/deploy disposition
Do not cite every recall result. Cite the memory actually used. If no recalled memory materially affects the work, no citation is required. If the active surface cannot expose recall_event_id or cannot call prism_memory_cite, record that tooling gap in the handoff instead of pretending compliance. Hard enforcement remains deferred. Recallgate warnings are advisory until a post-Slice-B real-traffic measurement window proves warning precision and the governance lane ratifies an enforcement change. The minimum memory-first questions are:
  1. What did Prism already know about this issue?
  2. Which artifact or memory result supports the next action?
  3. Is there a newer decision, spec, postmortem, or handoff that changes the answer?
  4. Did I cite the memory I used?

​
Architecture First

Architecture takes precedence over the easy fix. Do not hack to a passing result when the result conflicts with the design, hides the root cause, bypasses Prism lifecycle, or creates a local-only workaround for a system problem. After two failed probes, stop blind debugging and identify the source of truth:
  • docs before folklore
  • API and structured state before UI edge behavior
  • logs and code before assumptions
  • ADR/SPEC before convenience
  • owning peer review before architectural shortcuts
Architecture-impacting work needs the appropriate artifact: SPEC, ADR, plan update, decision record, or postmortem.

​
Peer Work and Four-Eyes

Prism parallelism is consensus-first. Use four-eyes review when work is high risk, architectural, governance-changing, cross-lane, incident-class, or likely to affect other agents. Standing lane defaults:
  • Donna: engineering execution and implementation feasibility
  • Texi: architecture and design authority
  • Candi: governance, methodology, and continuous improvement
  • Lafonda: install lane and environment setup
  • Samantha: runtime environment and release/test execution
  • Desiree: documentation
  • Porsche: dashboard and operator surfaces
Peer claims need evidence. Prefer artifact ids, branch/SHA, diff, test output, exact persistence checks, and semantic recall checks over conversational assertions.

​
Risk Tiers

Classify before acting:
  • R0: local, reversible, narrow; single-agent execution is usually fine
  • R1: affects another lane or user-visible behavior; at least one peer review or clear handoff evidence
  • R2: architecture, governance, methodology, lifecycle, or cross-surface behavior; consensus path required before implementation
  • R3: program-level, migration, multi-stack, or irreversible changes; explicit owner direction and staged verification
When in doubt between tiers, choose the higher tier long enough to get the right review.

​
Incident and RCA Discipline

All incidents and RCA findings are recorded immediately once the failure class can be named. This includes startup, bootstrap, MCP, launcher, install, signal, session, cross-machine, and methodology failures. A recurrence gets a new record even if a prior postmortem exists. Required incident sequence:
  1. Name the failure class.
  2. Create the incident or RCA artifact, usually prism_postmortem.
  3. Verify exact persistence.
  4. Verify semantic recall with distinctive terms.
  5. Continue or finish the fix.
  6. Cite the artifact in checkpoint, wrap, signal, PR, or handoff.

​
Completion Discipline

Completion means evidence, not confidence. For implementation work, done generally means:
  1. designed
  2. built
  3. tested
  4. committed
  5. deployed when deployment is part of the assignment
  6. smoke-verified when runtime behavior matters
For incident-class work, done also requires:
  1. incident/RCA artifact id
  2. exact persistence verification
  3. semantic recall verification
  4. citation in the final handoff
If any gate is intentionally deferred, record who deferred it and why.

​
Prism Verbs

Use the right verb for the job. Lifecycle:
  • prism_start
  • hd_exec
  • prism_checkpoint
  • prism_wrap
  • prism_journal
Coordination:
  • prism_signal
  • prism_signals_pending
  • prism_signal_ack
Memory and retrieval:
  • semantic_recall
  • prism_diff
  • prism_note
  • store_session_delta
  • prism_memory_cite
Governance and method:
  • prism_governance_lookup
  • prism_method_fragment_recall
  • prism_spec
  • prism_plan
  • prism_todo
  • prism_decide
  • prism_postmortem
  • prism_postmortem_resolve
Verification:
  • prism_test when available and appropriate
If a verb is unavailable on the current surface, record the gap and use the nearest durable Prism artifact. Do not silently skip the obligation.

​
Cross-Machine Handoff

For immediate cross-machine learning, use Prism memory as the primary channel:
  1. Put the full retrain bundle in a Prism artifact.
  2. Verify semantic recall.
  3. Signal the target identity with the artifact id and expected response.
  4. Use Git for canonical files after commit/pull exists.
  5. Use a server bucket only for large attachments; cite the bucket link and hash in Prism memory.
A file path alone is not a cross-machine handoff unless it is committed and the target machine has pulled it.

​
Self-Test

Before declaring work complete, answer:
  1. What did memory say before I acted?
  2. Which source of truth governed the work?
  3. What risk tier applied?
  4. Which peer or lane owner needed to review it?
  5. Which Prism artifact records the decision, incident, or handoff?
  6. What exact evidence proves persistence and recall?
  7. What remains deferred, and who accepted that deferral?
Last modified on May 18, 2026