Skip to content

Latest commit

 

History

History
281 lines (142 loc) · 12.7 KB

File metadata and controls

281 lines (142 loc) · 12.7 KB

Chapter 17 — Glossary

Alphabetical reference for all key terms. Each entry includes a definition and chapter cross-references.


A

ABSTAIN — A reviewer or Planner output status meaning "I cannot assess with sufficient confidence." Does not block the pipeline. → Ch.07, Ch.06

Acceptance criteria — A measurable condition that must be true for a phase to be considered complete. Minimum 1 required per phase. → Ch.06, Ch.08, planner.plan.schema.json

ACTING — An Orchestrator workflow state indicating active phase execution by subagents. → Ch.05

Agent — A specialized AI system defined by a *.agent.md file with P.A.R.T. structure. 13 agents in the ControlFlow system. → Ch.03

Agent grantsgovernance/agent-grants.json — the tool allowlist for each agent. → Ch.10

AssumptionVerifier — An adversarial reviewer that detects 17 types of plan mirages (claims not supported by the codebase). → Ch.07, assumption-verifier.plan-audit.schema.json


B

Backbone pattern — See docs/agent-engineering/MIGRATION-CORE-FIRST.md. A shared implementation pattern for multi-agent consolidation.

Batch approval — One approval request per wave (not per phase). Exception: destructive waves require per-phase approval. → Ch.08

Behavior contractdocs/agent-engineering/PROMPT-BEHAVIOR-CONTRACT.md. Behavioral invariants complementing P.A.R.T. structural rules. → Ch.04

BrowserTester — A subagent that runs E2E browser tests and validates UI/accessibility. → Ch.03, browser-tester.execution-report.schema.json

Budget tracking — A skill that monitors token/retry budgets per phase. → Ch.11


C

Clarification policydocs/agent-engineering/CLARIFICATION-POLICY.md. Rules for when to use vscode/askQuestions vs NEEDS_INPUT. → Ch.05, Ch.06

Clarification request — A structured payload (schema: clarification-request.schema.json) containing a question, options, and recommendation. → Ch.13

CodeMapper — A subagent for read-only codebase exploration. Returns a discovery report. → Ch.03, code-mapper.discovery.schema.json

CodeReviewer — A post-execution review subagent. Never owns a fix cycle. Returns validated_blocking_issues. → Ch.03, Ch.08, code-reviewer.verdict.schema.json

Cold start — The condition in which an executor arrives at a phase with only the repository and the plan description, without additional context. ExecutabilityVerifier tests exactly this. → Ch.07

COMPLETE — The final Orchestrator workflow state. → Ch.05

Completion gate — The final task phase: consistency review, todo verification, session-outcome flush, final summary. → Ch.08

Complexity tier — TRIVIAL / SMALL / MEDIUM / LARGE. Determined by Planner; drives review pipeline and model routing. → Ch.06, Ch.07, Ch.08

Confidence — A numeric value (0–1) in Planner output and gate events reflecting how certain the agent is in its assessment. → Ch.06

ControlFlow — A prompt/governance/eval repository for an agent orchestration system. 13 agents, 20 schemas, 7 governance files, 20 skills. → Ch.00

Convergence detection — See stagnation. → Ch.07

CoreImplementer — A subagent for backend implementation (code, tests, migrations). → Ch.03, core-implementer.execution-report.schema.json


D

Definition of done — A list of conditions that must be true for a phase to be declared complete. Matches quality_gates. → Ch.08

Delegation protocol — The payload the Orchestrator sends to a subagent. Schema: orchestrator.delegation-protocol.schema.json. → Ch.05, Ch.09

Drift check — A test in evals/drift-checks.mjs that verifies agent files haven't gone out of sync with contracts. → Ch.04, Ch.14


E

Eval harness — The offline test suite in evals/ (~410 checks). No live agents, no network. → Ch.14

Escalate — A failure classification: security/data risk or unresolvable blocker. Immediate STOP + WAITING_APPROVAL. → Ch.13

ExecutabilityVerifier — A reviewer that simulates cold-start executability of the first 3 tasks per phase. Active on LARGE tier. → Ch.07, executability-verifier.execution-report.schema.json

executor_agent — A required phase field (enum: 8 values). Orchestrator must not infer it heuristically. → Ch.08, Ch.06


F

Failure classification — A required field when status is FAILED/NEEDS_REVISION/NEEDS_INPUT/REJECTED. Values: transient/fixable/needs_replan/escalate. → Ch.13

Final review gate — An optional final CodeReviewer pass after all phases complete. Configured in governance/runtime-policy.json. → Ch.08

Fixable — A failure classification: small correctable error. Retry once with a fix hint. → Ch.13

Frontmatter — YAML metadata at the top of *.agent.md files (e.g., agents:, tools:). → Ch.04


G

Gate event — An Orchestrator transition event with a fixed field contract. Schema: orchestrator.gate-event.schema.json. → Ch.05, Ch.09

Governance — Configuration files in governance/ that define agent permissions, model routing, and operational policies. → Ch.10

Handoff — The Planner output field handoff: {target_agent, prompt} that routes the plan to Orchestrator. → Ch.06

HIGH_RISK_APPROVAL_GATE — A gate event type for high-risk operations requiring user approval. → Ch.05


H

Idea interview — The first Planner phase: converting a vague idea into concrete requirements through clarifying questions. → Ch.06


I

Iteration indexiteration_index — the current iteration number in the PLAN_REVIEW loop. Passed to all reviewers. → Ch.07


L

LARGE — The highest complexity tier. Full PLAN_REVIEW pipeline: PlanAuditor + AssumptionVerifier + ExecutabilityVerifier. → Ch.07


M

Memory architecture — The three-layer memory model (session / task-episodic / repo-persistent). → Ch.12, MEMORY-ARCHITECTURE.md

Mirage — A plan claim not supported by the actual codebase. Detected by AssumptionVerifier. → Ch.07

Model role — The model routing field per task type. Configured in governance/model-routing.json. → Ch.10


N

NEEDS_INPUT — A subagent status indicating that a clarifying question must be asked of the user. Routes through vscode/askQuestions, not through failure_classification. → Ch.13

needs_replan — A failure classification: architecture mismatch requiring a phase replan via Planner. Max 1 retry. → Ch.13

NOTES.md — The repo-persistent active-objective state file. Updated at each phase boundary. → Ch.12


O

Observability — Gate event logging to plans/artifacts/<task-id>/observability/<task-id>.ndjson. → Ch.05, OBSERVABILITY.md

Orchestrator — The conductor agent. Runs the state machine, dispatches subagents, enforces gates. → Ch.03, Ch.05


P

P.A.R.T. — The mandatory section order for every *.agent.md: Prompt → Archive → Resources → Tools. → Ch.04, PART-SPEC.md

Phase — A plan unit with an executor_agent, acceptance criteria, and quality gates. → Ch.06, Ch.08

PHASE_REVIEW_GATE — A gate event after phase completion. → Ch.05

PlanAuditor — An adversarial reviewer of architecture, security, risks, and rollback. → Ch.07, plan-auditor.plan-audit.schema.json

Planner — The planning agent. Runs an idea interview, produces phased plans, selects skills. → Ch.03, Ch.06

PLANNING — The initial Orchestrator workflow state during plan creation. → Ch.05

Plan path — The plan_path field in Planner output — the path to the plan file. A reviewable input, not an implicit approval. → Ch.05, Ch.06

Plan Review Gate — The PLAN_REVIEW stage between plan arrival and execution start. Conditional. → Ch.07

PlatformEngineer — A subagent for CI/CD, containerization, and infra deployment. → Ch.03, platform-engineer.execution-report.schema.json

PreFlect — A mandatory gate before each action batch. Uses preflect-core skill. 4 risk classes: destructive, scope-drift, assumption, dependency. → Ch.05, Ch.11

Prompt — The first section of P.A.R.T. Contains the agent's mission, scope, and invariants. → Ch.04

Quality gate — A phase readiness condition. Enum: tests_pass, lint_clean, schema_valid, safety_clear, human_approved_if_required. → Ch.08


R

Reflection loop — A skill for agents with a revision step: produce output, self-evaluate, refine. → Ch.11

Reliability gatesdocs/agent-engineering/RELIABILITY-GATES.md. Verification gate requirements. → Ch.08

REPLAN_REQUIRED — A Planner output status indicating requirements need clarification before planning can proceed. → Ch.06

Repo memory/memories/repo/ — durable codebase facts. Create-only (no edits). → Ch.12

Repo-persistent — The third memory layer: NOTES.md + /memories/repo/. Survives context resets. → Ch.12

REVIEWING — An Orchestrator workflow state during code review after a phase. → Ch.05

Researcher — A subagent for deep research. Returns findings with citations. → Ch.03, researcher.research-findings.schema.json

risk_review — A planner.plan field with 7 semantic risk categories, dispositions, and applicability. → Ch.06, Ch.07


S

Schema — A schemas/*.json file defining an agent output contract (JSON Schema draft 2020-12). 20 schemas total. → Ch.09

Scope drift — Executing actions beyond the declared plan scope. Detected by PreFlect and llm-behavior-guidelines. → Ch.05, Ch.11

Scoring specdocs/agent-engineering/SCORING-SPEC.md. Quantitative scoring formula for reviewer verdicts. → Ch.07

Semantic risk taxonomy — 7 risk categories in risk_review: data_volume, performance, concurrency, access_control, migration_rollback, dependency, operability. → Ch.06

Session memory — Layer 1: /memories/session/. Conversation-scoped scratch. → Ch.12

Skill — A reusable Markdown pattern file in skills/patterns/. Loaded just in time. → Ch.11

Skill indexskills/index.md. The registry from which Planner selects ≤3 skills per phase. → Ch.11

SMALL — A complexity tier. PLAN_REVIEW runs PlanAuditor only. → Ch.07

Stagnation — When score improvement over 2 iterations < 5% at iteration ≥ 3. Triggers WAITING_APPROVAL. → Ch.07

Subagent — Any agent that executes as a phase executor or reviewer. Not an entry point for direct user interaction. → Ch.03


T

Task-episodic — Layer 2: plans/artifacts/<task-slug>/. Per-task revision history and deliverables. → Ch.12

TDD — Test-driven development. Applied via the tdd-patterns skill. → Ch.11

TechnicalWriter — A subagent for documentation generation and code-doc parity. → Ch.03, technical-writer.execution-report.schema.json

Tool grantsgovernance/tool-grants.json. Properties and allowed callers for each tool. → Ch.10

Tool routingdocs/agent-engineering/TOOL-ROUTING.md. Rules for when to use external tools vs subagents. → Ch.05

Trace ID — A UUID v4 generated at task start and propagated to all gate events and delegation payloads. Enables log correlation. → Ch.05

Transient — A failure classification: temporary error (timeout, rate limit). Retry identically, max 3. → Ch.13

TRIVIAL — The lowest complexity tier. PLAN_REVIEW is skipped entirely. → Ch.07


U

UIImplementer — A subagent for frontend implementation (UI, styling, responsive, accessibility). → Ch.03, ui-implementer.execution-report.schema.json


V

Validated blocking issuesvalidated_blocking_issues in CodeReviewer verdict. Only these block phase progression — not raw CRITICAL/MAJOR findings. → Ch.08, Ch.09

Verified items — A list of items confirmed as correct in the previous iteration. Passed to reviewers for regression tracking. → Ch.07

vscode/askQuestions — The tool used by Orchestrator for mandatory clarification classes and NEEDS_INPUT routing. → Ch.05, Ch.13


W

WAITING_APPROVAL — An Orchestrator workflow state requiring user confirmation before proceeding. → Ch.05

Wave — A group of parallel plan phases. Wave N+1 starts only after all phases of wave N complete. → Ch.08

Wave-aware execution — Phase execution strategy with wave grouping and parallelism. → Ch.08

Workflow state — The current Orchestrator state machine node. Enum in orchestrator.gate-event.schema.json: PLANNING/WAITING_APPROVAL/ACTING/REVIEWING/COMPLETE. → Ch.05


See Also