mission-orchestrator

Table of Contents

• Overview
• When to Use
• Mission Lifecycle
• Interactive Plan Review
• Mission Types
• Phase-to-Skill Mapping
• Delegation During a Mission
• Session Recovery
• Module Reference
• Related Skills
• Related Commands
• Exit Criteria

Mission Orchestrator

Overview

Wraps the entire attune development lifecycle (brainstorm → specify → plan → execute) into a single mission with automatic state detection, type selection, and phase routing. Follows the "persistent presence lens" pattern from spec-kit:speckit-orchestrator: delegates entirely to existing skills via Skill() calls, never re-implements phase logic.

When To Use

• Starting a new project from scratch (full lifecycle)
• Resuming an interrupted project workflow
• Running a focused tactical implementation from existing specs
• Quick-fixing from an existing implementation plan

When NOT To Use

• Running a single phase directly (use /attune:brainstorm, /attune:specify, etc.)
• Non-project work (code review, debugging, research)
• When you need fine-grained control over phase transitions

Mission Lifecycle

1. State Detection
   Scan for existing artifacts (project-brief.md, specification.md, etc.)
       |
2. Mission Type Selection
   Auto-detect type based on artifacts, or accept user override
       |
3. Phase Routing Loop
   For each phase in the mission type:
       a. Pre-phase validation (check prerequisites)
       b. Invoke Skill(attune:{phase-skill})
       c. Post-phase artifact check (verify output exists)
       d. Post-phase backlog triage (create GitHub issues
          for out-of-scope items after brainstorm/specify)
       e. Update mission state
       f. User checkpoint (skippable with --auto)
       g. Error handling via leyline:damage-control
       |
4. Completion
   All phases complete, final state saved

Mission Types

review is the one type selected from request intent rather than from artifacts, because a tree of build artifacts looks the same whether the ask is "ship this" or "audit this". Its check runs first. It produces reports/<topic>-<YYYY-MM-DD>.md and never enters the war-room gate, which guards a plan-to-execute transition a review mission does not have.

See modules/mission-types.md for full type definitions and custom type support.

Phase-to-Skill Mapping

Review missions route to the existing review skills rather than to attune phase skills:

The orchestrator never re-implements phase logic. Each phase is a complete Skill() invocation that handles its own workflow.

Delegation During a Mission

Missions delegate execution by default. Skill(conjure:delegation-core) governs the decision, and its default posture is on: a phase that reaches execution work hands it to an external CLI without waiting to be asked.

The split follows conjure's own line: delegate execution, retain reasoning. A mission phase that is entirely judgment stays local and does not consult the delegator at all.

Two outcomes need handling rather than reporting:

• fallback_reason is providers_exhausted: no CLI answered. Say which were tried, then do the task in the mission itself. This is the ordinary path on a machine with no CLI installed and it does not fail the phase.
• fallback_reason is delegation_disabled: the operator declined. Do the task locally and do not offer to re-enable it.

To run a whole mission without external models, set CONJURE_DELEGATION=off before invoking it.

Session Recovery

Missions persist state to .attune/mission-state.json. On resume:

1. Load mission state file
2. Validate referenced artifacts still exist on disk
3. Identify last completed phase
4. Continue from next phase in sequence

See modules/mission-state.md for the state schema and recovery protocol.

Interactive Plan Review

The plan-to-execute transition uses an interactive review loop instead of a simple checkpoint. Plans are reviewed section by section, revised based on feedback, and must pass a mandatory war-room gate before execution.

Key capabilities:

• Section-by-section terminal review (architecture first, then phases)
• Approve/revise/reject verdicts with rationale
• Plan version tracking with diff summaries
• Context improvement from structured feedback
• Additive bias scanning before user review
• Maximum 3 revision rounds before forced decision
• Mandatory war-room approval with Prosecution Counsel

See modules/plan-review.md for the full protocol.

Review Modules

• plan-review.md: Main orchestrator for the review loop
• plan-versioner.md: Version tracking and diff generation
• feedback-collector.md: Verdict capture and JSON output
• context-injector.md: Revision prompt construction
• iteration-governor.md: Round tracking and escalation

User Directive Overrides

The orchestrator parses the user's command-args and free-text at mission start for natural-language trust signals. Phrases like "ignore scope guard", "ultrathink", "don't keep asking", and "be autonomous" are recognized as directive overrides that adjust the constraint profile without requiring an explicit --constraints= flag.

Directive overrides win over mission-type defaults but never bypass the Safety Floor (pre-commit hooks, proof-of-work evidence, destructive-operation confirmation, external-facing actions). When a directive is detected, the orchestrator acknowledges it once at mission start and stops asking for the corresponding checkpoints. Repeated approval-seeking after a directive override is itself a workflow bug.

See modules/adaptive-constraints.md "User Directive Override" section for the parsing table.

Mission Charter

Define mission boundaries using the structured template from references/mission-charter.md. A Mission Charter specifies:

• Outcome: What success looks like
• Success metric: Measurable completion criteria
• Deadline: Time boundary (session, date, or duration)
• Constraints: Token/time budgets, forbidden actions
• Scope: In-scope and out-of-scope areas
• Stop criteria: Conditions that halt the mission

See references/mission-charter.md for the full template and examples.

Progress Reports

Track progress with structured checkpoints using references/progress-report.md. Generate reports at:

• Phase boundaries (between brainstorm→specify→plan→execute)
• Blocker identification
• Risk escalation
• Budget thresholds (50%, 75%, 90%)

See references/progress-report.md for the template and checkpoint rhythm guidance.

Module Reference

Core modules (always loaded)

• mission-types.md: Type definitions, auto-detection logic, custom types
• state-detection.md: Artifact existence checks, quality validation, staleness
• phase-routing.md: Phase execution protocol, transition hooks, error handling
• mission-state.md: State schema, persistence, recovery protocol

Plan-review modules (load when plan phase runs)

• plan-review.md: Interactive section-by-section review with bias scanning
• plan-versioner.md: Version tracking and diff summaries
• feedback-collector.md: Verdict capture and feedback files
• context-injector.md: Revision prompt construction from feedback
• iteration-governor.md: Round tracking, cap enforcement, escalation

Conditional modules (load only when triggered)

• reflexion-buffer.md: Cross-session learning buffer; load when iteration count > 1 or after a failed revision round.
• trust-tier.md: Constraint-profile classifier; load when a user directive override is detected at mission start.
• adaptive-constraints.md: Constraint adaptation rules; load alongside trust-tier.md when directive overrides are active.

Module Loading by Mission Type

This skill declares progressive_loading: true. To keep the orchestrator's resident token cost minimal, load only the subset of modules each mission type actually needs. The orchestrator itself loads only the four core modules at mission start; the rest are loaded on-demand when their phase runs.

Token cost (approximate, computed from wc -w on hub + loaded modules and converted at ~1.3 tokens per word):

The previous load-all pattern brought in roughly 10,100 tokens for every mission, including quickfix runs that only need the execute phase. With per-type loading, quickfix is ~60% lighter and the standard / tactical / full paths save 22-32%.

When a directive override fires, the trust-tier + adaptive-constraints pair adds ~2,200 tokens on top of the mission-type baseline.

Reference Modules

• mission-charter.md: Structured mission definition template (load only when defining a charter)
• progress-report.md: Checkpoint status report template (load only when emitting a progress report)

Related Skills

• Skill(attune:project-brainstorming) - Brainstorm phase
• Skill(attune:project-specification) - Specify phase
• Skill(attune:project-planning) - Plan phase
• Skill(attune:project-execution) - Execute phase
• Skill(attune:war-room-checkpoint) - Risk assessment for RED/CRITICAL tasks
• Skill(leyline:risk-classification) - Task risk classification
• Skill(leyline:damage-control) - Error recovery during phases
• Skill(conjure:delegation-core) - Default-on delegation of execution work

Related Commands

• /attune:mission - Invoke this skill
• /attune:mission --resume - Resume from saved state
• /attune:mission --type tactical - Override mission type

Exit Criteria

• All phases in mission type completed successfully
• Artifacts exist for each completed phase
• Mission state saved to .attune/mission-state.json
• Risk summary generated (tier counts across all tasks)
• Execute-phase work either delegated or held back by a named Keep Local clause, never skipped for want of a delegation decision
• Any providers_exhausted result completed locally and reported with the providers it tried
• No unresolved errors or blockers