fusion/docs/missions.md

Missions

← Docs index

Missions provide structured planning across multiple related tasks.

Roadmaps are a separate lightweight planning model (Roadmap → RoadmapMilestone → RoadmapFeature) used for standalone planning. Missions remain the richer execution-oriented hierarchy when you need slice activation, autopilot, and feature-to-task delivery tracking.

Mission Hierarchy

Fusion models delivery as:

Mission → Milestone → Slice → Feature → Task

Example:

Mission: Improve Reliability
  Milestone: Stabilize execution pipeline
    Slice: Retry and recovery hardening
      Feature: Stuck task recovery improvements
        Task: FN-210
        Task: FN-214

Agent task-creation admission

Mission lineage is an admission requirement only for autonomous no-task heartbeat creates and delegations. Those idle patrol calls must supply a valid active Mission → Milestone → Slice → Feature chain; an allow rule for task_agent_mutation cannot bypass this requirement. Missing or invalid lineage is rejected before a task is persisted with an explicit mission-lineage remedy.

Interactive/user-supervised, task-scoped heartbeat, executor, triage, and workflow-step calls may create or delegate freeform tasks without lineage. They remain governed by the normal task_agent_mutation permission policy, including category and exact-tool allow, approval, and block rules.

A valid active lineage may name a hand-authored defined feature only for its first task. Fusion atomically claims the feature, links that exact task, and promotes the feature to triaged; an already-linked feature rejects rather than overwriting its canonical task. This bootstrap exception does not make defined executable: later scheduler and symbol-lock admission still uses the stricter contract below.

Canonical lineage approval for autonomous symbol locks

Before autonomous scheduler work may acquire a symbol lock, it resolves the task's Mission → Milestone → Slice → Feature lineage and evaluates the single @fusion/core contract: evaluateMissionLineageApproval. Resolution and lock acquisition remain scheduler responsibilities; downstream schedulers must not redefine the approval rule.

Approval requires every one of these statuses:

• Mission: active
• Milestone: active
• Slice: active
• Feature: triaged or in-progress (never defined; defined is only allowed at the first-task bootstrap boundary)

When the scheduler passes planApprovalRequired: true, the linked task must also have an approvedPlanFingerprint that is a non-empty string after trimming whitespace. The predicate does not recompute the fingerprint; plan-approval.ts owns its generation and validation. When plan approval is not required, the fingerprint is ignored.

The predicate is pure and returns { approved, reason }. Its stable reasons are approved, missing-mission, missing-milestone, missing-slice, missing-feature, mission-not-active, milestone-not-active, slice-not-active, feature-not-implementable, and plan-not-approved. A false result is the scheduler's lineage-blocked outcome; only an approved result is eligible for symbol-lock admission.

Mission → Goal linkage

Missions and goals are stored independently, with an optional many-to-many linkage persisted in the mission_goals join table.

• Columns: missionId, goalId, createdAt
• Primary key: (missionId, goalId)
• Foreign keys: missionId → missions.id, goalId → goals.id
• Delete behavior: both foreign keys use ON DELETE CASCADE, so removing either parent deletes only the corresponding join rows
• Reverse lookups are indexed via idxMissionGoalsGoalId

MissionStore owns the persisted linkage CRUD surface:

• linkGoal(missionId, goalId) — idempotently create a link and return { missionId, goalId, createdAt }
• unlinkGoal(missionId, goalId) — remove a link and report whether anything changed
• listGoalIdsForMission(missionId) — list linked goals in deterministic creation order
• listMissionIdsForGoal(goalId) — list linked missions in deterministic creation order

No-backfill decision

Existing missions are intentionally not auto-linked to any goals. Fusion does not run a migration backfill for pre-existing missions, so a mission with no links should be treated as genuinely unlinked until an operator or agent associates it with one or more goals.

Manual linkage workflow

Mission ↔ goal links are created and removed deliberately as part of normal planning and operations work. The dashboard exposes the relationship from both directions: Mission detail has an active-goal picker plus linked-goal chips with unlink controls, and each Goals view card has a mission picker plus linked-mission chips with unlink controls. Archived goals are never offered for new links, duplicate link attempts are no-ops at the store/API layer, and removing the last link restores the empty-state copy rather than leaving an empty control shell. The workflow is intentionally manual so teams can choose the correct strategic relationship per mission instead of inheriting guessed links from older data.

Unlinked mission indicator

Mission Manager shows an Unlinked indicator on active mission cards when linkedGoalCount is zero. Linking or unlinking from either dashboard surface refreshes this count so operators can quickly find active missions that still need an explicit goal association.

The engine also emits a workflow insight with advisory key unlinked_missions_advisory when it first observes one or more active missions with zero goal links. The insight is advisory only, includes only the affected mission ids plus a count, and is deduped to one stable row so it does not spam on every scheduler heartbeat.

Task → Goal provenance

When a mission feature is linked or triaged into a task, Fusion does not copy goal ids onto the task row. Instead, task goal provenance is always derived from the mission link owned by MissionStore:

• listGoalIdsForTask(taskId) resolves the owning mission from the linked feature hierarchy first (feature -> slice -> milestone -> mission), then falls back to the live task row's missionId when needed.
• listGoalsForTask(taskId) maps those ids back to full Goal records using the same goals-table read path as getMissionWithHierarchy, so mission reads and task provenance stay in sync.
• Unknown, unlinked, or partially missing hierarchy state resolves fail-soft to [].
• Archived goals remain part of provenance; only missing goal rows are dropped.

This derived bridge lets downstream systems recover which strategic goals a task serves without duplicating mission-goal linkage during task creation.

Goal-injection diagnostics provenance field

The engine's resolveAndEmitGoalContext seam still injects only the always-on active-goal context into prompts, but diagnostics now add provenanceGoalIds: string[] alongside the existing injected goalIds / goalCount fields.

• goalIds / goalCount continue to describe the active goals injected into the prompt.
• provenanceGoalIds records which mission-linked goals the task serves.
• Diagnostics and run-audit metadata persist ids/counts only — never goal titles, descriptions, or prompt text.

Creating Missions

Mission base branch defaults

Missions support an optional baseBranch field. When set, feature triage (triageFeature) and slice triage (triageSlice) inherit this value as the task baseBranch whenever a triage request does not explicitly provide a base branch override.

Precedence order during triage:

1. Explicit triage branchSelection.baseBranch / baseBranch
2. Mission baseBranch
3. Project default branch resolution

Mission task prefix override

Missions support an optional taskPrefix field. When set, feature triage (triageFeature / triageSlice) passes it as a transient minting hint on TaskCreateInput so the distributed task-id allocator issues ids under that prefix. When unset or cleared, triage inherits the project-wide settings.taskPrefix.

The Mission Manager create/edit form exposes this as Task prefix (empty = project default). Clearing a previously saved prefix on edit sends taskPrefix: null so the stored override is removed.

Mission branch strategy defaults

Missions can also persist a branchStrategy used whenever triage is triggered without explicit branch options (manual triage and autopilot triage).

Supported modes:

• project-default (or absent): shared mode; each triaged feature gets a distinct per-task working branch (for example <shared-branch>/<feature-id>) while the shared branch remains the mission group merge target
• auto-per-task: sets branchAssignment.mode = "per-task-derived" (distinct per-task working branches with no shared mission group merge target)
• existing: shared mode using branchSelection.mode = "existing" with branchName as the shared merge-target branch
• custom-new: shared mode using branchSelection.mode = "custom-new" with branchName as the shared merge-target branch

The Mission Manager create/edit form exposes this as Branch strategy plus a conditional Branch name field for existing and custom-new.

Mission auto-merge override

The Merge behavior control can inherit the project default, explicitly enable auto-merge, or select Single pull request. In-context help explains that auto-merge lands each feature individually while a single pull request retains all features on a shared branch for joint review. The latter persists autoMerge: false on the mission and stamps newly triaged feature tasks with the same false override, while preserving the mission's shared branch group. Returning the control to inherited clears the mission override. Once a shared branch has members, Mission detail displays its branch name, member count, and PR state.

Shared branch-group invariant across entry points

Across all branch entry points (planning/subtask creation, mission triage, and New Task shared-group creation), Fusion enforces one rule:

• Persist a per-task working branch on each task (the checkout branch used for execution).
• Persist the shared branch only as the group merge target via branchContext.groupId → branch_groups.branchName.
• Never persist the shared branch itself as a task working/checkout branch.

This keeps member execution isolated per task while still routing member landings into a single shared integration branch.

Dashboard

Use the Mission Manager UI to create missions and build hierarchy interactively.

On mobile, Mission Manager surfaces the primary Plan New Mission CTA at the top of the mission list for faster access, while desktop keeps the split-layout sidebar CTA anchored in the bottom action region as the primary entry point.

Mission detail refreshes now preserve expanded milestone/slice state and keep the selected milestone expanded, so persisted milestone acceptance criteria remain visible across live updates.

Mission, milestone, slice, and feature read-only text surfaces in Mission Manager render Markdown (GFM) for descriptions, verification, and acceptance criteria; edit forms continue to use raw plain-text <textarea> inputs.

CLI

fn mission create "Reliability initiative" "Reduce execution failures and improve recovery" --goal G-001 --goal G-002
fn mission list
fn mission show mission_123
fn mission goals mission_123
fn mission link-goal mission_123 G-001
fn mission unlink-goal mission_123 G-001
fn mission activate-slice slice_456
fn mission delete mission_123 --force

Mission ↔ Goal operator surfaces

Fusion surfaces the persisted mission↔goal linkage through REST, CLI, and pi-extension tools.

REST endpoints

Endpoint	Purpose
POST /api/missions	Create a mission. Optional body field goalIds: string[] links goals during creation and returns linkedGoals in the response.
PATCH /api/missions/:missionId	Update mission fields. Optional goalIds: string[] replaces the full linked-goal set; [] clears links and undefined leaves links unchanged.
GET /api/missions/:missionId	Return MissionWithHierarchy, including linkedGoals as an always-present array of Goal objects for the selected mission and optional eventCount as the authoritative unfiltered mission activity total.
GET /api/missions/:missionId/goals	List linked goals for a mission. Returns { goals }.
GET /api/goals/:goalId/missions	List linked missions for a goal. Returns { missions: [{ id, title, status }] } and skips stale links whose mission row no longer resolves.
PUT /api/missions/:missionId/goals	Replace the full linked-goal set with body { goalIds: string[] }. Duplicate ids are deduplicated before reconciliation.
POST /api/missions/:missionId/goals/:goalId	Idempotently link one goal to a mission.
DELETE /api/missions/:missionId/goals/:goalId	Idempotently unlink one goal from a mission.

The mission detail payload keeps linkedGoals separate from the milestone tree so read paths can surface strategy context without traversing slices/features. All goal-link write endpoints preserve the same invariant: missing goals on link write paths (POST /api/missions, PATCH /api/missions/:missionId, PUT /api/missions/:missionId/goals, POST /api/missions/:missionId/goals/:goalId) reject with 400 { code: "GOAL_NOT_FOUND" }, archived goals reject with 400 { code: "GOAL_ARCHIVED" }, duplicate/relinked ids are no-ops, and the DELETE /api/missions/:missionId/goals/:goalId unlink path treats unknown goals as a 404 while remaining allowed even after a goal is archived.

CLI

• fn mission create ... --goal <goal-id> [--goal <goal-id> ...] — create a mission and batch-link active goals.
• fn mission goals <mission-id> — list linked goals for a mission.
• fn mission link-goal <mission-id> <goal-id> — idempotently link a goal; archived goals reject with GOAL_ARCHIVED.
• fn mission unlink-goal <mission-id> <goal-id> — idempotently unlink a goal, including archived goals.
• Dashboard Mission detail lets operators link active goals, unlink existing goal chips, and select a chip to open the Goals view at the anchored goal card.
• Dashboard Goals cards show linked missions, let operators link/unlink missions for that goal, and select a mission chip to open Mission Manager at that mission.

Mission Planning Tools (pi extension)

The canonical per-parameter tool reference lives in packages/cli/skill/fusion/references/extension-tools.md; this section is a user-facing summary of the mission-planning tool surface.

Tool	Purpose
fn_mission_create	Create a mission with title/description, optional baseBranch, and optional auto-advance behavior.
fn_mission_list	List missions and their current status.
fn_mission_show	Show mission details with milestone/slice/feature hierarchy, including a Linked Goals section plus milestone/feature acceptance criteria and slice verification when present.
fn_mission_list_goals	List the goals linked to a mission.
fn_mission_link_goal	Idempotently link a goal to a mission; archived goals reject with GOAL_ARCHIVED.
fn_mission_unlink_goal	Idempotently unlink a goal from a mission, including archived goals.
fn_mission_delete	Delete a mission and its hierarchy.
fn_mission_update	Update mission title/description using partial patches.
fn_milestone_add	Add a milestone to a mission.
fn_milestone_update	Update milestone fields using partial patches.
fn_slice_add	Add a slice to a milestone.
fn_slice_activate	Activate a pending slice for implementation.
fn_slice_delete	Delete a slice (with linked-task guard and optional force).
fn_feature_add	Add a feature to a slice with optional acceptance criteria.
fn_feature_delete	Delete a feature (with linked-task guard and optional force).
fn_feature_update	Update feature fields using partial patches.
fn_feature_link_task	Link a feature to a task for implementation.
fn_milestone_delete	Delete a milestone (with linked-task guard and optional force).

fn_mission_update

Updates an existing mission's title or description. Partial patches leave untouched fields intact — fields omitted from the call are not modified.

Parameter	Type	Required	Description
id	string	✓	Mission ID to update (e.g., M-001)
title	string	—	Updated mission title
description	string	—	Updated mission description

Use this to reconcile mission narrative/state text without recreating the mission.

fn_milestone_update

Updates an existing milestone's title, description, or acceptanceCriteria. Partial patches leave untouched fields intact — fields omitted from the call are not modified.

Parameter	Type	Required	Description
id	string	✓	Milestone ID to update (e.g., MS-001)
title	string	—	Updated milestone title
description	string	—	Updated milestone description
acceptanceCriteria	string	—	Updated acceptance criteria for completing the milestone

Callers can only update milestones within missions they have access to. Use fn_milestone_add to create milestones. This update behavior was introduced in FN-4578.

fn_feature_update

Updates an existing feature's title, description, or acceptanceCriteria. Partial patches leave untouched fields intact — fields omitted from the call are not modified.

Parameter	Type	Required	Description
id	string	✓	Feature ID to update (e.g., F-001)
title	string	—	Updated feature title
description	string	—	Updated feature description
acceptanceCriteria	string	—	Updated acceptance criteria for completing the feature

Use this to edit existing features without delete-and-re-add cycles.

Mission delete policy (hard delete with linked-task guard)

Mission hierarchy records (missions, milestones, slices, mission_features) use hard deletes with FK cascades and do not have deletedAt soft-delete columns.

To keep behavior consistent, Fusion uses hard delete with guard for feature/slice/milestone deletes:

• Delete is rejected when the target (or any cascading child feature) is linked to a live task (deletedAt IS NULL and not archived).
• Callers can pass force: true to override the guard. Force clears the mission linkage before deletion, then proceeds with the same hard delete.
• Linked tasks are preserved; only mission hierarchy rows are removed.

This intentionally differs from task soft-delete behavior described in docs/soft-delete-verification-matrix.md and avoids a mission-table soft-delete migration.

Mission Interview and Planning Workflow

The dashboard supports mission planning workflows where you can:

• Define mission outcomes
• Break work into milestones/slices/features
• Associate features to executable tasks
• Track progress at each layer
• Persisted missions with interviewState: "in_progress" remain visible as interview-styled mission cards in the main mission list so planning work does not disappear after reloads
• Resume in-progress mission interview sessions directly from separate transient session rows in the main missions list (mission_interview sessions in generating, awaiting_input, error, or complete) before a mission record is created; complete means the plan summary is ready for review/approval but has not been converted into a mission yet
• Mission interview closes are non-destructive: the modal now uses a single close action for header close, backdrop click, and Escape. Closing preserves the in-progress mission_interview session, and Missions re-fetches project-scoped transient rows (including on the mobile stacked Missions view) so resume/retry remains discoverable without losing persisted interviewState: "in_progress" mission cards. Resume-launched interview modals also expose a Send to background action that performs the same non-destructive park without cancelling the session. Deletion remains an explicit sidebar action.
• Mission interview, milestone interview, and slice interview agents have read-only board visibility via fn_task_list and fn_task_show, so they can reference active backlog context and avoid duplicating in-flight tasks while asking planning questions

Mission Interview Drafts

Mission interview sessions are persisted in ai_sessions before a mission row exists, so unfinished drafts stay recoverable across reloads and restarts.

• Dashboard: the Missions view shows a Drafts section for in-flight mission_interview sessions with Resume/Review and Discard actions. A complete draft is a generated-but-unapproved plan parked at the summary step.
• CLI: fn mission list shows drafts by default before normal mission status sections, including complete plan-ready drafts. Pass --no-drafts to hide them.
• pi extension: fn_mission_list includes drafts by default and accepts includeDrafts: false to suppress them; complete mission interview drafts are returned here too.
• Discarding drafts: discarding removes the ai_sessions row even for cold drafts after a server restart.

Mission interview draft endpoints:

Endpoint	Purpose
GET /api/missions/interview/drafts	List in-flight mission interview drafts
POST /api/missions/interview/drafts/:sessionId/discard	Discard a draft session

Auto-Generated Assertions

Fusion keeps a canonical per-feature assertion invariant in MissionStore:

• addFeature() creates exactly one store-managed assertion for each feature and links it.
• updateFeature() keeps that assertion synchronized when title, description, or acceptanceCriteria change.
• deleteFeature() removes the store-managed assertion to avoid orphaned rows.
• This applies to all creation paths (interview import, API, CLI, tools).

Assertion text source priority is: acceptanceCriteria → feature.description → fallback text ("Verify implementation of: {feature.title}").

Operator repair note (FN-5696): Some databases created before the feature-create-path fix could show feature acceptanceCriteria/description in the UI but still have zero mission_feature_assertions links, which caused validator auto-pass short-circuits. Use the built-in backfill operator surfaces instead of ad-hoc scripts:

• Agent/tool: fn_mission_backfill_assertions with { missionId?, dryRun? } (dry-run default)
• API: POST /api/missions/:missionId/backfill-assertions with body { dryRun?: boolean }

Run dry-run first, then apply (dryRun=false) when the report looks correct. Scope by mission id for targeted repair (for example Goals mission M-MP32KU9Y-0001-2ADN).

• Verification fields: Milestone and slice verification criteria from the interview are stored in dedicated verification fields rather than concatenated into descriptions
• Milestone acceptanceCriteria derivation: explicit milestone.acceptanceCriteria from interview output is authoritative. When omitted/blank, Fusion derives a deterministic bulleted summary from child features after creation: prefer feature.acceptanceCriteria, fall back to feature.description, skip empty contributors, and leave milestone acceptance empty when nothing contributes
• Partial plans handled: Auto-generation is robust to partial plans (missing slices/features or empty criteria) without throwing errors

Milestone Text Field Semantics

Milestones now carry three complementary free-text fields:

• description — narrative scope of the phase
• verification — informal "how to confirm" notes
• acceptanceCriteria — structured acceptance/assertion text (the canonical pass/fail bar), parallel to feature-level acceptanceCriteria

Slice Activation and Progress

Slices represent staged execution windows.

• Pending slices remain inactive
• Active slices are currently allowed to progress
• Completion rolls up through feature → slice → milestone → mission

Manual activation is available through fn mission activate-slice <slice-id>.

Mission Autopilot

Missions are always created stopped (status: "planning", autopilotEnabled: false, autoAdvance: false). Autopilot must be enabled explicitly after creation (for example via start/update actions). When autopilotEnabled is on, Fusion can watch completion events and progress missions automatically.

State machine:

• inactive
• watching
• activating
• completing

Typical flow:

1. Mission is watched (missions updated with autopilotEnabled: true or explicitly started are watched)
2. Task completion updates feature status
3. If a slice is complete, autopilot activates next pending slice
4. When milestones are all complete, mission transitions to complete

If validation cannot run (unexpected loop state, duplicate trigger, blocked validation, or validator error), Fusion logs a mission warning/error event with structured metadata so the stuck state is visible in mission events.

Mission status and autopilotEnabled transitions are atomically written with a mission activity event. The event records stable actor type/id, optional display name, source, and before/after values; unchanged values create no transition event. Dashboard controls identify an operator, tools identify an agent when they expose a sensitive mutation, and autonomous engine paths identify the system/autopilot.

autopilotEnabled vs autoAdvance

• autopilotEnabled: primary control for autopilot behavior — enables background monitoring, orchestration, and automatic slice activation when a slice completes. Also triggers auto-planning (converting features to tasks) when a slice is activated.
• autoAdvance: legacy fallback for backward compatibility with existing mission data. Kept for compatibility — new missions should use autopilotEnabled.

Auto-planning behavior:

• autopilotEnabled=true → features in activated slices are automatically planned (converted to tasks)
• autopilotEnabled=false, autoAdvance=true → features are planned (legacy compat)
• Active autopilot slices are continuously reconciled on startup recovery and periodic maintenance: stranded features (taskId == null) are re-triaged idempotently, title-matched tasks are linked first, and successful link/triage repairs emit mission:stranded-feature-triaged run-audit events.
• autopilotEnabled=false, autoAdvance=false → manual slice activation only

Slice progression (on slice completion):

• autopilotEnabled=true → next pending slice is automatically activated
• autopilotEnabled=false, autoAdvance=true → next pending slice is activated (legacy compat)
• autopilotEnabled=false, autoAdvance=false → manual activation required

Dashboard UI: The Mission Manager groups mission run settings together: explicit Start mission / Stop mission / Resume mission actions control mission run-state, while the Autopilot toggle controls automatic slice advancement and feature planning. The autopilot badge uses human-readable states (Off, Watching, Activating slice, Completing). When enabling autopilot on an already-active mission, the system automatically checks whether recovery is needed (no active slice or completed active slice) and progresses accordingly.

Autopilot API Endpoints

Endpoint	Purpose
GET /api/missions/:missionId/autopilot	Get autopilot status for mission
PATCH /api/missions/:missionId/autopilot	Enable/disable autopilot ({ enabled: boolean })
POST /api/missions/:missionId/autopilot/start	Start watching manually
POST /api/missions/:missionId/autopilot/stop	Stop watching manually

Feature Reconciliation API Endpoint

Use this endpoint when a feature's delivery task has already shipped and is now terminal (done or archived), but the feature status still needs to be reconciled to done.

POST /api/missions/features/:featureId/reconcile-done

Request body:

{ "taskId": "FN-123" }

Safety gate behavior:

• Validates featureId and requires a non-empty string taskId.
• Resolves the feature and terminal delivery evidence in the request's scoped PostgreSQL project.
• Accepts either a live, non-deleted task in done, or a supported archived task represented by both its retained soft-deleted column=archived project-task tombstone and its project-scoped cold archive snapshot. A deletion without both archive representations is not delivery evidence.
• Atomically validates conflicts, writes the canonical feature→task link, marks the feature done, updates the live task's reverse mission/slice link when applicable, and recomputes slice/milestone rollups. Any validation or write failure rolls back the complete operation.
• Archived evidence remains archived and read-only: reconciliation intentionally leaves its retained task tombstone without a writable reverse backlink rather than resurrecting the task.
• Repeating the same terminal task against the same completed feature is idempotent. A feature linked to another task, or a task linked to another feature, returns 409 without mutation.
• The repair path never enters ordinary triage/implementation state, changes loop attempts, creates or moves a task, activates/watches the mission, or changes mission status, autopilotEnabled, or autoAdvance.

Safe duplicate cleanup: preserve the first 409; verify through supported APIs that the current linked task is generated duplicate work with no unique delivery or lineage value; call POST /api/missions/features/:featureId/unlink-task; archive only the proven duplicate through the supported task archive API; then call reconcile-done with the canonical terminal task. Never overwrite a mismatched link, unarchive/move canonical delivery evidence, or use direct storage edits. If the duplicate is ambiguous, leave it untouched and escalate for evidence.

How this differs from PATCH /api/missions/features/:featureId:

• PATCH keeps the execution-status guard and rejects done/triaged/in-progress/blocked when no linked task exists.
• reconcile-done is a dedicated, evidence-gated transaction for already-shipped work. It is not a shortcut for active work or mission execution.

Error responses:

• 400 — invalid feature ID format or missing/empty taskId.
• 404 — feature not found or no task/archive evidence exists for the supplied task ID.
• 409 — feature/task mismatch, task already linked to another feature, nonterminal task, or a deleted/archived task lacking the supported retained tombstone plus cold snapshot. Every 409 leaves feature, task, rollups, and mission controls unchanged.

Validation Contract Lifecycle

Fusion's validation contract lifecycle is the structured feature delivery system for missions. It combines validation contracts, AI validation, and bounded retries to provide systematic, auditable feature completion. The lifecycle covers the full end-to-end path from clarification through blocked handoff.

End-to-End Flow

Clarification → Validation Contract → Feature Execution → Validator Loop
      ↑                                                         ↓
      │    Fix-Feature Retry ←─ (budget exhausted?) ←───────────┘
      │
Blocked Handoff ←── (budget exhausted, root cause unresolvable)

Phase 1: Clarification

The clarification phase occurs during mission interview and planning. Operators define:

• Milestone outcomes and slice verification criteria stored in dedicated verification fields
• Feature descriptions and acceptance criteria

These inputs flow directly into assertion auto-generation in the next phase.

Phase 2: Validation Contract

Contract assertions (MissionContractAssertion) formalize what must be true for a feature to be considered complete:

interface MissionContractAssertion {
  id: string;              // e.g., "CA-A3B7CD-E9F2"
  milestoneId: string;     // Parent milestone
  sourceFeatureId?: string;// Store-managed feature assertion owner
  scope: "feature" | "milestone";
  origin: "authored" | "imported" | "derived_milestone_acceptance";
  title: string;           // Human-readable title
  assertion: string;       // Behavioral plan
  status: AssertionStatus; // pending | passed | failed | blocked
  orderIndex: number;      // Sort order within milestone
  featureIds: string[];    // Linked features (many-to-many)
}

Assertion text source priority:

1. acceptanceCriteria (from feature planning)
2. feature.description (fallback)
3. Fallback text: "Verify implementation of: {feature.title}"

Coverage tracking: MilestoneValidationRollup computes per-milestone coverage:

interface MilestoneValidationRollup {
  milestoneId: string;
  totalAssertions: number;
  passed: number;
  failed: number;
  blocked: number;
  pending: number;
  unlinked: number;
  state: MilestoneValidationState;
}

Validation state precedence (highest priority wins):

1. not_started — no assertions exist
2. needs_coverage — assertions exist but some are not linked to features
3. ready — assertions exist and are linked, but not all have passed
4. passed — all assertions have passed
5. failed — at least one assertion failed
6. blocked — at least one assertion is blocked

Completion Gate Contract

Canonical authored feature criteria live on MissionFeature.acceptanceCriteria, and each feature validator derives its verdict only from its linked feature-scoped assertions. Validator prompts list each authoritative assertion ID in brackets; responses must return exactly one result keyed by each listed ID. To recover older model output safely, only an exact-count response with zero recognized IDs is matched positionally and recorded in diagnostics. Partial matches, duplicate IDs, and count mismatches remain fail-closed. Model summary prose, milestone prose, and behavioral results that are not mapped to a linked behavioral assertion cannot override that verdict.

Milestone prose is synchronized to one canonical milestone-scoped assertion with origin: "derived_milestone_acceptance". PostgreSQL restricts uniqueness to that derived origin per project/milestone; authored, imported, and migrated legacy milestone assertions stay independent, are never inferred from title/text, and require no feature links. The rollup evaluates all milestone-scoped assertions after feature coverage and feature assertion passes are ready; unmet parent criteria therefore block milestone completion without failing an already-passing feature. See Mission Completion Gate Contract.

Phase 3: Feature Execution Loop

Features track their implementation state via FeatureLoopState separate from task status:

type FeatureLoopState =
  | "idle"         // Not yet started
  | "implementing" // Tasks are in-flight
  | "validating"   // Awaiting AI validation
  | "needs_fix"    // Validation failed, retry in progress
  | "passed"       // All assertions passed
  | "blocked";     // Retry budget exhausted, cannot proceed

State transitions:

idle → implementing → validating → passed (all assertions pass)
                          ↓
                   needs_fix → implementing (retry feature created)
                          ↓
                      blocked (budget exhausted)

When a feature enters the implementing state, implementationAttemptCount is initialized and incremented on each retry.

Phase 4: Validator Loop

On task completion, the scheduler calls MissionExecutionLoop.processTaskOutcome() to run AI validation:

1. Find the feature linked to the completed task
2. If assertions are linked, keep feature completion gated until validation passes
3. Transition feature to validating state
4. Fire the AI validator agent (read-only judge) against contract assertions
5. Apply the behavioral-verification posture (see below): static assertions keep the judge's verdict; behavioral/bug assertions default to fail until a bounded, non-mutating verification run confirms them
6. Record MissionValidatorRun metadata for the validation attempt (per-assertion failures are stored separately in MissionAssertionFailureRecord rows)

For a linked task with a recorded mergeDetails.commitSha, the read-only judge runs from a disposable detached checkout of that landed merge revision rather than the ambient project checkout. If that checkout cannot be materialized, the judge falls back to the project root; a fail is deferred to inconclusive when the landed commit is not reachable from that same inspected root, or when the landed revision/its ancestry cannot be verified, preventing a branch-divergence false failure. The task worktree fork point (baseCommitSha) is never used as an inspection revision.

Behavioral-verification posture (adversarial default-to-fail). A Contract Assertion now carries a type (static | behavioral). The validator no longer grades a Feature "done" purely from the diff's apparent intent:

• Static assertions (e.g. "documented in README") keep today's read-only static judging — no added cost or strictness.
• Behavioral / bug-fix assertions default to fail. The read-only judge's "pass" on a behavioral assertion is advisory, not authoritative; an authoritative pass requires a separate, bounded verification run that exercises the implemented code (running the test suite / an agent-supplied regression test against a disposable checkout) and confirms the observable behavior. An agent's narrative claim is not evidence on its own.

The verification run is not read-only and is not part of the judge session. The AI judge session stays tools: "readonly" (no bash/edit/write/task-mutation). The verification run is a separate, side-effecting execution that runs against an isolating sandbox backend (fail-closed when none is available) and a disposable checkout at a trusted revision — never the live worktree, never the repo root. Its effects are confined to that disposable surface: it creates no board task, mutates no mission/board row, and leaves the source tree that feeds diff/merge byte-identical (git-clean) after the run. Verification is therefore no longer "purely read-only/static" — but it is non-mutating to mission/board state, which is the invariant the recovery sweep and reaper depend on (see Surface Enumeration).

Inconclusive is a first-class verdict, distinct from fail. Verification yields pass / fail / inconclusive. A real behavioral failure (fail) spawns a Fix Feature with a recorded observed-vs-expected reason. An inconclusive verdict — verification could not run or conclude (no isolating backend, timeout, isolation-setup failure, rejected proof, detected flakiness) — routes the feature to a blocked/needs-attention state with a persisted verification_inconclusive mission event and spawns no Fix Feature, so a fragile verification surface cannot manufacture remediation churn. A non-passing verification never resolves to a default pass.

Mission validation resolves its model from the validator lane before session creation: assigned agent runtime model (when the linked task has an assigned durable agent) → per-task validatorModelProvider/validatorModelId → project validatorProvider/validatorModelId → global validatorGlobalProvider/validatorGlobalModelId → project defaultProviderOverride/defaultModelIdOverride → global defaultProvider/defaultModelId. In testMode, validation is forced to mock/scripted instead of falling through to provider auto-detection.

Validation runs are internal mission-loop operations: Fusion does not create visible 🔍 Validate: board tasks for single-feature validation.

interface MissionValidatorRun {
  id: string;
  featureId: string;
  milestoneId: string;
  sliceId: string;
  status: "running" | "passed" | "failed" | "blocked" | "error";
  triggerType?: string;
  implementationAttempt: number;
  validatorAttempt: number;
  taskId?: string;
  summary?: string;
  blockedReason?: string;
  startedAt: string;
  completedAt?: string;
  createdAt: string;
  updatedAt: string;
}

Validation timeout: 10 minutes (VALIDATION_TIMEOUT_MS = 10 * 60 * 1000). If session creation, auth/credit checks, prompting, or timeout fails, the run is marked error and emits a surfaced validation_error mission event instead of silently spawning a fix feature.

Stale validator-run reaper: startup recovery and periodic self-healing also sweep MissionValidatorRun rows stuck in status="running" longer than VALIDATOR_RUN_STALE_MAX_AGE_MS (currently 6 hours). Runs still owned by the live process (tracked in activeValidations) are skipped, so a slow-but-legitimate verification is never reaped while its session is in-flight. Ownerless stale runs are reaped to terminal status="error", their reap reason is stored in summary, and live (non-done) mission features are moved to loopState="needs_fix" with lastValidatorStatus="error" so the loop can re-trigger. A done feature's loop state is intentionally left untouched (it keeps the loopState="validating" set when the run started) so the reaper does not rewrite a feature that already finished its task. Runs whose parent mission is already complete/archived are likewise terminated without touching feature state. Each successful reap emits a run-audit event with mutationType: "mission:validator-run-reaped".

Verification wall-clock is bounded under the reaper window. The aggregate verification budget — checkout materialization plus the test-suite command (VERIFICATION_COMMAND_TIMEOUT_MS, 10 min), including the optional pre-fix baseline run — is provably far shorter than the 6-hour reaper stale window, so a legitimate verification run completes long before it would be eligible for reaping. The reaper's activeValidations skip is the second line of defense: an in-flight run is never reaped regardless of wall-clock.

Phase 5: Fix-Feature Retries

Validation always records failed runs and diagnostics. Remediation is separately opt-in: Fusion creates and auto-triages a Fix Feature only when autopilotEnabled === true || autoAdvance === true. With both flags false or unset, validation is report-only: the failed validator run and validation_failed/validation_report_only mission events remain visible, but Fusion does not mint a Fix Feature, create or plan a task, triage work, or dispatch remediation. This same predicate applies to completion, startup recovery, and periodic recovery, so a restart cannot bypass supervised mode.

autoAdvance remains the legacy compatible opt-in. New missions should prefer autopilotEnabled.

When the opt-in is enabled and validation fails, MissionStore.createGeneratedFixFeature() creates a fix feature with lineage tracking:

interface MissionFixFeatureLineage {
  sourceFeatureId: string;      // Original feature being remediated
  fixFeatureId: string;         // New fix feature
  runId: string;                // Validator run that triggered this fix
  failedAssertionIds: string[]; // Assertions that failed
}

The fix feature is auto-planned (converted to tasks) for immediate execution. Each fix increments the canonical root feature's implementationAttemptCount; generated fixes never receive a fresh budget. With the default retry budget of 3 (DEFAULT_IMPLEMENTATION_RETRY_BUDGET), requests 1–3 mint one remediation each and request 4 mints nothing, records budget-exhausted, and blocks the root.

Missing, cyclic, or legacy blocked lineage fails closed. A pre-migration blocked root without an explicit stop reason cannot mint remediation and cannot be implicitly resumed.

Phase 6: Blocked Handoff

A feature transitions to blocked when:

1. All retry budget is exhausted (implementationAttemptCount >= maxRetryBudget)
2. Validation continues to fail
3. Root cause cannot be resolved through iteration

Blocked semantics:

• Autopilot stops advancing the slice containing the blocked feature
• MilestoneValidationRollup.state reflects blocked assertions
• The feature remains in blocked state until operator intervention
• Deleting a generated fix, or archiving/deleting its generated task, records a durable root-scoped operator-intervention stop in the same transaction as unlink/removal. Recovery, duplicate delivery, unarchive, task/root recreation, and relinking cannot mint a sibling. The stop remains even if a hierarchy cascade removes root and lineage rows.
• POST /api/missions/:missionId/resume is the sole resume seam. It atomically clears only operator-intervention stops, preserves attempt counts, moves extant roots to needs_fix, and activates the mission. If any root is budget-exhausted or legacy-unknown, it returns a typed MISSION_RESUME_CONFLICT with canonical root IDs/reason categories and changes no root, tombstone, counter, or mission state.

On engine restart, recoverActiveMissions() re-enqueues features in validating or needs_fix states, ensuring no validation work is lost. It also re-triggers implementing features whose linked task is already done/archived and whose assertion validation has not passed yet. When the stale-run reaper has already converted an abandoned validator run into needs_fix, processTaskOutcome() promotes the feature back through implementing and re-validates instead of skipping it. The same recovery path is replayed during periodic self-heal maintenance, so historically stranded implementing features can self-heal without requiring an engine restart.

Reaper → slice deadlock closure (P0). A task-less, done, assertion-linked feature is the dangerous case: it carries no board task to re-drive from, and computeSliceStatus refuses to count it complete until its validator passes. When the reaper terminates such a feature's stale run, the feature is left stranded in loopState="validating" (the reaper's done-guard, above) — a state the validating/needs_fix recovery branches (which only re-drive features that carry a taskId) never re-validate, while default-to-fail would otherwise re-drive it forever to a non-terminal error. recoverActiveMissions() closes this with a stranded-done catch-all: any task-less, done feature in loopState implementing or validating (or needs_fix + lastValidatorStatus="error") that has not reached a passing validator status and is not currently being validated is re-driven directly through runFeatureValidation(). Because the verification run is bounded and non-mutating, this reaches a terminal pass / fail / inconclusive (and the slice can finally resolve) instead of livelocking on validating/error.

Surface Enumeration — validation re-drive entry points (R15)

Now that the verification step has side effects (on a disposable, isolated surface — never mission/board state), every site that re-drives validation must remain correct: after a run the source tree feeding diff/merge is git-clean, no duplicate Fix Feature is minted, and a terminal verdict is reached without an error-state slice deadlock. The complete set of re-drive entry points, each gated by an adversarial reliability test in packages/engine/src/__tests__/reliability-interactions/mission-verification-redrive-surface.test.ts:

Entry point	Trigger	Post-conditions asserted
processTaskOutcome()	Normal task-completion validation	terminal verdict; one Fix Feature on fail (idempotent on re-drive); no validation-created board task
recoverActiveMissionValidations → validating branch	Restart with a feature stranded mid-validation (has taskId)	re-driven to terminal verdict; git-clean; no duplicate Fix Feature
→ needs_fix branch	Reaped/abandoned run on a feature with a taskId	promoted via processTaskOutcome; terminal verdict
→ implementing + taskId branch	Feature left implementing while its task already finished	re-triggered to terminal verdict
→ stranded-done catch-all (implementing/validating/needs_fix+error, no taskId)	Orphaned or reaped task-less done feature (the P0 deadlock)	re-driven directly; terminal verdict, never indefinitely re-driven error; slice resolves
reapStaleMissionValidatorRuns	Stale ownerless run	run → terminal error; live feature → needs_fix; done feature loopState untouched; in-flight runs skipped

Each path is verified to leave zero mission/board residue from the verification run itself — the only board task a failed verdict legitimately creates is the auto-triaged Fix Feature, and an inconclusive verdict creates none.

For features with missing linked assertions, the completion path is now validator-first: the loop lazily restores the store-managed per-feature assertion just before validation, then runs the AI validator instead of auto-passing. Milestone acceptanceCriteria is threaded into the validator prompt for every feature in that milestone, so all mission criteria are AI-evaluated. Contract details are defined in Mission Completion Gate Contract.

Autopilot / Scheduler Interplay

The scheduler and autopilot collaborate through a carefully ordered call sequence:

1. Task completes → scheduler detects completion
2. scheduler.missionExecutionLoop.processTaskOutcome() — validation FIRST
   - Finds linked feature, runs AI validation, records MissionValidatorRun
3. autopilot.handleTaskCompletion() — feature status sync SECOND
   - Syncs feature status from task state, advances slice if complete
4. scheduler filters blocked missions from further advancement (line ~532)

Autopilot vs Execution Loop retry tracking:

• Autopilot: Per-task retry tracking for slice/feature completion events
• Execution Loop: implementationAttemptCount for retry budget enforcement (default: 3)

These are independent tracking mechanisms — autopilot monitors mission progress while the execution loop manages feature-level retry budgets.

Telemetry and Observability

MissionHealth snapshot fields:

• activeSlices, activeFeatures, blockedFeatures
• validationState, validationRollup
• inProgressCount, passedCount, failedCount, blockedCount

MissionEvent audit types:

• slice_activated, feature_planned, feature_completed
• validation:started, validation:passed, validation:failed, validation:blocked
• milestone_missing_structured_assertions (legacy-data warning surface; enforcement still lazy-restores managed assertions at runtime)
• fix_feature:created, feature:blocked

Validator run telemetry:

• triggerType — free-form trigger source (manual, task_completion, auto, etc.)
• implementationAttempt — which retry attempt this was
• validatorAttempt — how many validator runs for this implementation
• status — running | passed | failed | blocked | error
• summary — natural language summary of results

Assertion failure records:

interface MissionAssertionFailureRecord {
  assertionId: string;
  assertionTitle: string;
  expected: string;
  actual: string;
  message: string;
}

Full state snapshots: MissionFeatureLoopSnapshot captures complete loop state including all validator runs and lineage chains for post-mortem analysis.

Validation failure diagnostics

A validation_failed Mission activity event includes metadata.validationDiagnostics, the typed source of truth for failure reporting. It contains the validator runId, sourceFeatureId, overall outcome, next action, and ordered per-assertion verdicts with expected, observed, message, and evidence references. The visible event text is derived from this object—not an AI summary—so a failed event always names failed assertion IDs and labels any separately blocked assertion IDs as blocked (never as failed).

Evidence is secret-redacted before persistence. Each assertion retains at most 16 evidence entries and every message, expected, observed, and evidence text field is capped at 4,096 UTF-8 bytes. Bounded fields carry truncated: true, excess evidence is reported as omittedEvidenceCount, project paths become project-relative, and external or disposable absolute paths become [external path omitted].

Generated fix features and their triaged tasks include the same Validation cause section with source feature, validator run, failed assertion IDs, bounded observations, and evidence. SQLite MissionStore and PostgreSQL AsyncMissionStore use the shared renderer, so a retry does not produce backend-specific causes or duplicate sections. A fix that is already linked to a canonical task is an idempotent race; otherwise Mission activity tells the operator to inspect and retry triage rather than exposing internal exception/loop-state prose.

The loop state is internal scheduling context, not an operator diagnosis. Its public meanings and actions are:

Public state	Meaning	Operator action
validating	A validator run is evaluating the landed implementation.	Inspect the run only if it remains active beyond the stale-run window.
needs_fix	A validator found a remediable assertion failure.	Review the event’s Validation diagnostics and triage the generated Fix feature/task.
blocked	Validation could not obtain sufficient proof, or retry budget is exhausted.	Resolve the stated external constraint or root cause, then retry/triage the feature.
implementing	A task is carrying out the feature or its generated remediation.	Follow the linked task; duplicate validator triggers with a canonical task are ignored.

Operator Troubleshooting

Symptom	Diagnosis	Resolution
Feature stuck in "validating"	Validator owner may have died, leaving a stale MissionValidatorRun in status="running"	Check mission-loop/self-healing logs; the startup or maintenance reaper should terminate runs older than VALIDATOR_RUN_STALE_MAX_AGE_MS (6h) and emit mission:validator-run-reaped
Fix feature not auto-planning	planFeature() may have errored; check logs	Manual planning via fn mission plan-feature <id>; investigate planFeature() errors
Budget exhaustion loop	implementationAttemptCount >= maxRetryBudget (default: 3)	Increase maxRetryBudget in mission settings or fix root cause
Blocked mission not advancing	MilestoneValidationRollup.state shows blocked	Identify blocked assertions; operator must resolve root cause
Validation agent errors	AI session creation failed or VALIDATION_TIMEOUT_MS (10 min) exceeded	Check model configuration and logs; verify AI provider auth
No validation runs after task completion	processTaskOutcome() not called; check scheduler logs	Verify mission linkage on feature → task mapping; check scheduler event handlers
Recovery after engine restart	Features in validating/needs_fix/stalled implementing state may not re-enqueue	recoverActiveMissions() should run on startup; check recovery log count and mission-loop logs

Parity Verification Tests

This lifecycle is validated by integration tests in two dependent tasks:

FN-1571 — Core parity tests:

• packages/core/src/mission-factory-parity.integration.test.ts — MissionStore rollups, assertion persistence, validator run records, fix feature lineage
• packages/engine/src/mission-factory-parity.integration.test.ts — Scheduler/autopilot/runtime parity with the validation loop

FN-1572 — Dashboard parity tests:

• packages/dashboard/src/mission-e2e.test.ts — API contract telemetry round-trip (MissionContractAssertion → validator run → MissionHealth)
• packages/dashboard/app/components/__tests__/MissionManager.test.tsx — UI blocked/iterating state rendering

Screenshot

See also: Multi-Project and Task Management.

Agent and dashboard-chat tools

Mission hierarchy operations are available with the same project-scoped MissionStore contract in the pi extension, engine-managed executor/triage/heartbeat agents, and provider-backed dashboard chat. The surface is fn_mission_list, fn_mission_show, fn_mission_create, fn_mission_update, fn_mission_delete, fn_milestone_add, fn_milestone_update, fn_milestone_delete, fn_slice_add, fn_slice_activate, fn_slice_delete, fn_feature_add, fn_feature_update, fn_feature_delete, and fn_feature_link_task.

fn_mission_list and fn_mission_show are positively classified read-only. All other hierarchy operations mutate persisted project data and remain subject to the engine action gate and permanent-agent permission policy; they are never treated as unknown or exempt tools.

For example, activate a ready work unit with fn_slice_activate({ id: "SL-…" }). Link it to live work with fn_feature_link_task({ featureId: "F-…", taskId: "FN-…" }). Linking delegates to MissionStore.linkFeatureToTask(): it verifies the task is a live row in the same project, changes the feature to triaged, and records the mission/slice linkage on the task. Archived, deleted, missing, and other-project tasks are rejected.

Ideation handoff

Persisted ideation converges a selected candidate into this canonical hierarchy. It atomically creates or attaches a Mission and persists that linkage, rather than maintaining a parallel roadmap document.

Research-derived features

A completed cited research finding may become a normal Mission Feature. Its feature retains research run, stable finding, and source-URL provenance; optional triage uses the normal feature task flow. Linked task changes reconcile through the existing feature → slice → milestone → mission rollups, and task completion remains subject to assertion validation.

Autonomous mission admission

Autonomous no-task heartbeat agents may create or delegate implementation work only with an approved Feature → Slice → Milestone → Mission lineage. Interactive and task-scoped calls remain governed by task_agent_mutation policy as described in Agent task-creation admission. The created task stores that lineage as task metadata; it does not replace the canonical feature taskId link except at the documented defined-feature first-task bootstrap. Missing or invalid autonomous lineage is rejected before a task is persisted. Roadmap reconciliation marks done tasks done, returns cancelled/requeued tasks to triaged, keeps failed work non-complete, and treats archives as non-promoting no-ops.

Validator memoization and failure budget (FN-8694)

Automatic feature validation is content-addressed by landed SHA, resolved judge provider/model, and exact built prompts. Admission is atomic per project, feature, and fingerprint: a matching running run is not duplicated; the latest terminal history is selected deterministically; static-only passes are reused; and matching failures permit at most three dispatched runs before the feature is blocked. Behavioral or mixed assertions never reuse a pass, but failures are still budgeted.

Every automatic suppression appends one visible validation memoized activity event (running, reuse-pass, or budget-exhausted) with fingerprint and referenced run ID where available. Initial exhaustion additionally appends one validation-stuck event; later unchanged sweeps append only their memoized event. No synthetic validator run or verdict is created for reuse or exhaustion. Missing landed SHA, fallback checkout, unknown judge identity, and preparation failures fail open to ordinary validation; error/blocked outcomes are transient. Manual validation bypasses memoization and the budget. Recovery revisits only a feature bearing FN-8694's budget-block provenance: unchanged inputs remain blocked, while a changed prepared fingerprint can be admitted; unrelated blocked/remediation/operator states stay closed.