[ppXD]

Summary

• An Auto launch routed to plan-map or single-agent silently dropped the operator's allowed model pool, persona pool, decision critic, delivery spec and integrate opt-in (only SupervisorDefinitionBuilder read them), and plan-map dropped the acceptance floor the composer demanded. Every route-dependent control (allowedModelIds, allowedAgentDefinitionIds, acceptanceChecks, tier, decisionReviewMode, deliverySpec, integrateBranches, requirePlanConfirmation, plannerReviewMode) now gets a LaunchControlDisposition (Applied / Clamped / NotApplicable / Refused, with the reason) from one ILaunchControlResolver that TaskLaunchService and TaskRoutePreviewService share; both results carry ControlDispositions. Any Refused stops the launch before a session or run exists with task_launch_control_refused (422), listing each refused control. The preview also runs the launch's own pool checks (EnsureModelsInTeamAsync / EnsureAgentDefinitionsInTeamAsync), so a pooled row disabled or a persona deleted since the composer loaded is the same not-found on both.
• Model pool, by credentialed-model row id (model names are unique only per credential). Plan-map bounds the planner's catalog (WorkflowPlanRequest.AllowedModelIds) and every branch agent.run; single-agent bakes it on its agent; the agent.run manifest declares allowedModelIds. At dispatch HarnessModelReconciler chooses through AllowedModelPool: the pinned row, the authored model on the pinned credential, that model on any pooled credential, the pinned credential's best pooled row, then the pool default. That order now holds at the engine level: under a non-empty pool AgentDefinitionResolver leaves an unusable pinned row on the task instead of failing the node at staging (an unbounded task still fails closed). AgentRunExecutor runs and persists the chosen row and names any move on the timeline by cause, with model and credential names. A pool with nothing runnable fails the run with agent_model_pool_unresolvable (not retried, logged at Warning without a stack trace). Every escalation stays inside the pool and stamps the row it picked beside the model (every later revise round and the cold-launch record carry that row), so the output review avoids the model that did the work, live and after a re-attach. Quick makes the same choice at launch, so the frozen config shows the row that runs. A persona the pool excludes is refused on single-agent and plan-map, named by its name.
• Acceptance floor, by route and workspace. Each builder advertises whether its argv adapter needs a repository (TaskProjectionAcceptanceContract.RequiresRepository; defaults to true). The single agent grades the argv in its own workspace (the executor's local acceptance verifier), so a repository-free floor on the single-agent lane is Applied. The supervisor grades against the branch it publishes, so without a repository its floor is refused; at Delivery/Unattended the tier disposition refuses once, with the remedies the floor names (bind a repository, or launch at Quick effort; Prototype when no floor is carried), never "author a check". Plan-map refuses the floor; a malformed argv (blank executable, a null or NUL argument) is refused by name and counts as absent. On Auto, a well-formed floor keeps the route on a lane that grades it in this workspace: a repo-less floor lands on the single agent, never Deep or plan-map. The move is named in RoutePlan.DegradedReason, and the confirm card suggests the tier the route took. Grounding is resolved before routing in both launch and route snapshot, and Signals.Ambiguous asks for confirmation.
• Where this departs from the design brief (the code wins):
  • Builders stay pure (no DB), so Quick's clamp is chosen by the resolver and the builder bakes it. The dispatch choice is made in the reconciler and applied by the executor, which owns the task write. A plan-map pin outside the pool is reported Clamped without rewriting the profile: each branch is held at dispatch, and the synthesis keeps its model.
  • The repository requirement is a per-builder advertisement rather than a switch on the lane name, so a new builder states it (and is treated as needing a repository until it does).
  • IsWellFormedFloor keeps its own copy of the argv rule; the shared AcceptanceArgv normaliser is not on main yet.
• Accepted residuals:
  • The route preview resolves grounding without EnsureSessionCanContinueAsync, so a Continue preview can update the session summary and call the summarising model, as the launch always did.
  • Dispositions are returned on the launch response and kept in the route snapshot's ResultJson, not stored on the run.
  • The capability degrade runs after the floor reroute (latent): an Auto Deep route with a floor and a repository whose supervisor lane is unavailable degrades to plan-map, where the floor is refused instead of rerouted.
  • Review agents run on the launch's own ReviewerModelId, which the pool does not bound (unchanged here).
  • The TaskLaunchAuto benchmark arm now confirms an ambiguous classification, as it already did a low-confidence one.
• Follow-ups:
  • Composer: it reads the preview's verdict, so it now lets a repo-less argv floor launch on the single-agent lane. But its Delivery gate (LaunchTaskModal.tsx ~:323, the stale missing.push("an acceptance check") gate) still asks a repo-less Deep launch for a check, where the backend's remedy is to bind a repository or launch at Quick effort.
  • Deep same-name tie: RealSupervisorActionExecutor.Spawn.cs:1131 resolves a spawn's model through ModelPoolSelector.ResolveDispatchAsync, which breaks same-name ties by row id (OrderBy(m => m.Id)) and overwrites a persona's pinned credential.
  • Deep spawns choose their model inside the pool but do not stamp AllowedModelIds on the spawned AgentTask, so their revise-loop escalation is bounded by credential only.
  • Grade plan-map's integrated tree against the operator floor.
  • Have the composer render the dispositions and mark the suggested tier among the confirm card's options.

Test plan

• Unit, full suite: 13,241 passed, 1 skipped (the environment-gated cross-device case in LocalRwxAtomicCreateTests), 0 failed. Focused classes: LaunchControlResolverTests 30, AllowedModelPoolTests 10, LlmEffortClassifierTests 42, TaskRoutePreviewServiceTests 24, TaskLaunchServiceQualityTierTests 6, EffortRouterTests 19, EffortRouterDegradeTests 1, EffortRouterGenericityTests 4, EffortPolicyTests 24, DeploymentAutonomyCeilingTests 38, FailureTaxonomyTests 8, SingleAgentDefinitionBuilderTests 38, PlanMapSynthDefinitionBuilderTests 34, SupervisorDefinitionBuilderTests 30, AgentCodeNodeTests 151, PlanAuthorNodeTests 13, AgentRunExecutorEscalationTests 10, AgentRunExecutorOutputReviewTests 70, AgentDefinitionResolverTests 22

• Integration (real Postgres):

  • Launch side: TaskLaunchFlowTests 59, TaskRoutePreviewFlowTests 8, TaskLaunchContractFlowTests 11, TaskRouteSnapshotFlowTests 11, EffortRouterFlowTests 3, PlannerModelPoolFlowTests 1, PlannerCassetteDriftTests 2 (ExpectedPlannerKey unchanged)
  • Dispatch side: HarnessModelReconcilerFlowTests 19, AgentDefinitionResolverCredentialedModelFlowTests 9, AgentDefinitionResolverFlowTests 22, AgentRunExecutorTests 152 (engine-staged pool and escalation-row tests included), AgentRunReviseLoopFlowTests 17, AgentNodeFlowTests 26, LocalAcceptanceExecutorFlowTests 14, AgentReviewerLoopFlowTests 2, CriticObservedModelFlowTests 20, AgentRunReattachFlowTests 15, SupervisorRichSpawnFlowTests 13
  • All green

• E2E: TaskRouteSnapshotEndpointTests 20, TaskLaunchEndpointE2ETests 6, SpecPreviewArgvVocabularyTests 10, SupervisorWholeLoopE2ETests 14, PlanMapSynthFanoutFlowTests 2, PlanMapDynamicFanoutFlowTests 1, all green

• 41 mutations. Each reverts one decision's key line, and the named test went red; every file was restored by copy and checked by hash.

  Mutation	Test that went red
  single-agent advertises it needs a repository	LaunchControlResolverTests.A_repo_less_argv_floor_on_the_single_agent_lane_is_applied_…; TaskLaunchFlowTests.A_repo_less_argv_floor_on_the_single_agent_lane_launches_and_its_agent_carries_the_check
  verdict ignores the repository	LaunchControlResolverTests.A_repo_less_argv_floor_on_the_supervisor_lane_is_refused_…
  router grades as if a repository were bound	LlmEffortClassifierTests.A_repo_less_operator_floor_moves_a_deep_route_to_the_single_agent_that_grades_it
  route request always says "repository"	TaskLaunchFlowTests.An_auto_route_with_a_repo_less_floor_lands_on_the_single_agent_that_grades_it_never_on_deep
  tier tells a repo-less Deep launch to author a check	LaunchControlResolverTests.A_repo_less_supervisor_launch_at_delivery_has_one_remedy_…; TaskLaunchFlowTests.A_launch_at_delivery_quality_without_an_acceptance_check_is_refused_as_its_preview_predicts (deep, no repo)
  engine fails the node for a pooled pin	AgentRunExecutorTests.An_engine_staged_run_pinned_to_a_row_whose_credential_died_runs_another_pooled_row_and_says_so and …whose_pool_has_nothing_runnable_fails_typed_once_and_is_not_respawned
  escalation keeps the old row (task / persisted / live review)	AgentRunExecutorTests.An_escalated_run_is_persisted_and_reviewed_as_the_row_that_did_the_work (three separate mutations)
  preview skips the model guard / the persona guard	TaskRoutePreviewFlowTests.A_pool_member_gone_since_the_composer_loaded_… (the matching row each time)
  persona refused by id	TaskLaunchFlowTests.An_auto_route_to_quick_refuses_a_persona_the_allowed_pool_excludes_before_any_run
  unresolvable pool logged at Error	AgentRunExecutorTests.An_engine_staged_run_whose_pool_has_nothing_runnable_…
  manifest drops allowedModelIds	AgentCodeNodeTests.The_manifest_declares_the_allowed_model_pool_the_node_reads
  escalation escapes the pool	AgentRunReviseLoopFlowTests.A_run_bound_to_an_allowed_model_pool_never_escalates_outside_it
  chooser matches by name before the pinned row	AllowedModelPoolTests.A_pinned_row_is_honoured_over_a_same_named_row_on_another_credential
  dispatch pins by model name only	HarnessModelReconcilerFlowTests.The_same_model_on_another_pooled_credential_is_named_as_a_credential_move
  untyped pool failure / node retries it	AgentRunExecutorTests.A_bounded_run_whose_pool_resolves_nothing_fails_rather_than_run_outside_it; AgentCodeNodeTests.A_resumed_failure_carries_the_retry_verdict_for_the_engine (that row)
  malformed floor counts as a floor	LaunchControlResolverTests.A_malformed_floor_is_refused_by_name_…; LlmEffortClassifierTests.A_malformed_floor_is_no_routing_signal
  repo-less supervisor floor applied	TaskLaunchFlowTests.A_repo_less_argv_floor_on_the_supervisor_lane_is_refused_with_the_verdict_its_preview_gives
  tier left out of the dispositions	TaskLaunchFlowTests.A_launch_at_delivery_quality_without_an_acceptance_check_is_refused_as_its_preview_predicts
  preview / launch result drops dispositions; refusal not thrown	TaskRoutePreviewServiceTests.Preview_reports_the_same_control_dispositions_…; TaskRoutePreviewFlowTests.The_preview_reports_the_same_control_dispositions_the_launch_then_applies; TaskLaunchFlowTests.An_explicit_standard_launch_with_an_operator_floor_is_refused_…
  Quick clamp skipped; planner sees the whole pool	TaskLaunchFlowTests.An_auto_route_to_quick_clamps_a_pinned_model_…; PlannerModelPoolFlowTests.The_planners_catalog_lists_only_the_allowed_pool_…
  pool never baked / empty pool baked; node drops it	SingleAgentDefinitionBuilderTests.The_allowed_model_pool_rides_…, PlanMapSynthDefinitionBuilderTests.The_allowed_model_pool_bounds_…; AgentCodeNodeTests.The_allowed_model_pool_a_projection_baked_is_carried_onto_the_task
  reconciler never consults the pool; executor ignores the pooled row	HarnessModelReconcilerFlowTests.A_bounded_task_naming_a_model_outside_its_pool_runs_the_pools_default_row_and_says_so; AgentRunExecutorTests.A_bounded_run_naming_a_model_outside_its_pool_runs_the_pooled_row_…
  excluded persona applied; supervisor-only controls dropped	TaskLaunchFlowTests.An_auto_route_to_quick_refuses_a_persona_…; LaunchControlResolverTests.Every_control_the_launch_carries_gets_exactly_one_disposition_per_lane
  plan-map adapter reports a floor compatible; auto floor not rerouted	TaskLaunchContractFlowTests.A_standard_launch_carrying_an_operator_floor_is_refused_and_records_nothing; TaskLaunchFlowTests.An_auto_route_with_an_operator_floor_never_lands_on_plan_map_…
  launch or snapshot routes before grounding; ambiguity ignored	TaskLaunchFlowTests.A_continue_is_routed_on_the_threads_grounding_by_both_the_preview_and_the_launch; LlmEffortClassifierTests.The_router_STILL_confirms_an_AMBIGUOUS_task_…

• Pins changed deliberately:

  • TaskRoutePreviewServiceTests.Adapter_compatibility_comes_from_the_actual_builder_and_workspace: repo-less Quick is now Compatible, and RequiresRepository is true only on Deep.
  • The four mandate pins in TaskLaunchServiceQualityTierTests moved to LaunchControlResolverTests.The_tiers_floor_mandate_is_a_disposition_the_preview_can_predict.
  • In TaskLaunchFlowTests, the Deep ArgumentException pin and A_quick_launch_at_delivery_quality_is_never_rejected_for_a_missing_acceptance_check became one Theory: Deep with a repository and Quick without one are told to author a check; Deep without a repository is told to bind one.
  • The persona refusal is pinned by name instead of id.
  • TaskLaunchContractFlowTests.Every_launch_lane_records_original_controls_in_the_frozen_run_detail binds a repository only on the Deep row (Deep's capability key becomes GitBranch); its Standard row sends no floor, with that refusal pinned in its own test.
  • Fixtures that send Deep an argv floor bind a repository. The single-agent fixtures are back to main's repo-less form.

• Added in the final round (each mutation of its line went red): later revise rounds and the cold-launch record keep the escalated row (AgentRunReviseLoopFlowTests.A_later_round_that_does_not_escalate_…, AgentRunExecutorTests.A_cold_launch_that_escalated_at_dispatch_…, the stored row in …A_round_that_over_claims_…); the persona tenancy filter (TaskLaunchFlowTests.A_persona_of_another_team_outside_the_pool_…); the dispatch team barrier (AgentDefinitionResolverCredentialedModelFlowTests Theory over revoked / disabled / foreign-team / missing, HarnessModelReconcilerFlowTests.A_pinned_row_of_another_team_…).