ticket: route direct delegation authority split

This commit is contained in:
Keisuke Hirata 2026-06-08 14:47:22 +09:00
parent 776e7ca3e3
commit fa39f921d5
No known key found for this signature in database
2 changed files with 155 additions and 2 deletions

View File

@ -6,11 +6,13 @@ status: 'open'
kind: 'task'
priority: 'P1'
labels: ['pod', 'scope', 'delegation', 'orchestrator', 'security', 'profile']
workflow_state: 'intake'
workflow_state: 'inprogress'
created_at: '2026-06-08T01:04:42Z'
updated_at: '2026-06-08T01:04:42Z'
updated_at: '2026-06-08T05:47:15Z'
assignee: null
legacy_ticket: null
queued_by: 'workspace-panel'
queued_at: '2026-06-08T05:45:43Z'
---
## Background

View File

@ -5,3 +5,154 @@
Created by LocalTicketBackend create.
---
<!-- event: decision author: intake at: 2026-06-08T03:31:19Z -->
## Decision
## Intake classification
- readiness: requirements_sync_needed for implementation, but ready for Orchestrator routing to preflight/design synchronization.
- needs_preflight: true
- risk_flags: [authority-boundary, scope-permission, profile-manifest-resolution, pod-metadata-persistence, spawnpod-policy, backward-compatibility, diagnostics]
This Ticket touches authority boundaries and has explicit unresolved design questions, so it should not be routed directly to implementation. The next routing should be preflight/design before any coder worktree or implementation Pod is started.
---
<!-- event: intake_summary author: intake at: 2026-06-08T03:31:25Z -->
## Intake summary
Existing Ticket refined rather than duplicated. Scope is to split direct Pod/tool authority from delegation authority for child Pod spawn scopes, preserving explicit parent grants and deny behavior. Intake classification: ready for Orchestrator routing, not direct implementation; route to preflight/design first because the Ticket touches authority boundaries, profile/manifest scope resolution, Pod metadata/restoration behavior, SpawnPod policy, backward compatibility, and diagnostics. needs_preflight: true. risk_flags: [authority-boundary, scope-permission, profile-manifest-resolution, pod-metadata-persistence, spawnpod-policy, backward-compatibility, diagnostics].
---
<!-- event: state_changed author: intake at: 2026-06-08T03:31:25Z from: intake to: ready reason: intake_ready field: workflow_state -->
## State changed
Intake complete for routing. The Ticket is ready to queue for Orchestrator preflight/design routing; implementation should not start until the unresolved authority-boundary questions are settled or explicitly bounded.
---
<!-- event: decision author: intake at: 2026-06-08T05:44:05Z -->
## Decision
## Implementation direction: carrying delegation authority in Pod state
Adopt a split where direct scope and delegation scope are separate authority surfaces.
- Keep existing `scope` as direct tool/filesystem authority for the Pod itself.
- Add a separate `delegation_scope` / delegation grant surface for authority the Pod may pass to children through `SpawnPod`.
- Do not derive delegation authority implicitly from direct write scope. If `delegation_scope` is absent, the Pod cannot delegate write scope.
- Do not persist a mutable "remaining delegation scope" in Pod metadata. Store static grants in the resolved manifest snapshot, and use `spawned_children` as the durable record of currently consumed delegated scope.
- On restore, derive:
- direct effective scope = `manifest.scope` plus write-deny rules for outstanding `spawned_children` delegations, preserving the current direct-write conflict behavior;
- delegation effective scope = `manifest.delegation_scope` minus outstanding `spawned_children` delegations.
- Runtime `Pod` state should carry separate shared handles for direct scope and delegation scope. Built-in filesystem tools continue to consult direct scope; `SpawnPodTool` validates against delegation scope.
- `pod-registry::Allocation` should gain a delegation grant field so `delegate_scope()` checks subset against delegation grants rather than direct `scope_allow`. Registry conflict ownership should still prevent overlapping live write allocations.
- Existing `spawned_children` remains the durable dynamic authority-consumption record; `spawned_pods.json` remains only a runtime mirror/hint.
This supports the target role pattern: Companion/Intake/Reviewer default to direct read with no write delegation; Orchestrator can have limited/no direct workspace write while still having explicit bounded delegation grants for child worktrees; Coder receives direct child-worktree write and no further delegation unless explicitly granted.
---
<!-- event: state_changed author: workspace-panel at: 2026-06-08T05:45:43Z from: ready to: queued reason: queued field: workflow_state -->
## State changed
Ticket queued for Orchestrator routing.
---
<!-- event: decision author: orchestrator at: 2026-06-08T05:47:04Z -->
## Decision
Routing decision: implementation_ready
Reason:
- The Ticket thread contains an explicit implementation direction that resolves the authority-boundary question: every Pod has direct operation authority, while only a separately granted delegation capability authorizes `SpawnPod` scope subdivision.
- The binding decisions distinguish direct read/write scope from delegation grants, define default no-delegation behavior, preserve direct Bash/Edit/Write access, and require clear SpawnPod diagnostics.
- Remaining uncertainty is bounded implementation design: exact type names, Profile/manifest field names, metadata persistence shape, and test organization.
Evidence checked:
- Ticket body: problem, goals, requirements, non-goals/constraints, acceptance criteria, and validation expectations.
- Thread: Intake summary, implementation direction, binding decisions/invariants, and latest `ready -> queued` event.
- Workspace state: no matching branch/worktree exists; main workspace has unrelated Ticket-record edits and newly queued panel Ticket records outside this implementation path.
- Code map search for `SpawnPod`, scope/delegation validation, profile/manifest scope fields, pod metadata/scope persistence, and child scope allocation.
- Ticket doctor: 0 errors; existing warnings are unrelated legacy closed-Ticket diagnostics.
IntentPacket:
Intent:
- Split Pod direct operation authority from subdelegation authority so a Pod may directly read/write broad scope without automatically being allowed to spawn children over that scope.
Binding decisions / invariants:
- Direct scope authorizes the current Pod's own tools only.
- Delegation scope/capability authorizes only what that Pod may pass to `SpawnPod` children.
- Default delegation authority is none unless explicitly granted.
- Broad direct workspace/write scope must not imply broad child delegation.
- `SpawnPod` must validate requested child scope against the parent's delegation grant, not the parent's direct tool scope.
- Existing direct operations (`Read`, `Write`, `Edit`, `Bash`) must keep using direct scope and must not be reduced by this split.
- Existing/restored older Pods without delegation metadata must be unable to spawn children with delegated filesystem scope until granted by explicit new configuration/metadata.
- Profile/workspace role defaults should grant delegation intentionally only to roles that are supposed to orchestrate children.
- Do not weaken scope path validation, registry accounting, child scope reclaim, or host-authority checks.
Requirements / acceptance criteria:
- Introduce a typed representation for direct scope versus delegation/subdelegation scope.
- Update Profile/manifest/runtime config parsing so role Profiles can express broad direct scope with narrow/no delegation authority.
- Update Pod metadata/session persistence and restore so effective direct scope and delegation grant are durable/replayable.
- Update `SpawnPod` validation and diagnostics to use delegation authority and say when delegation is missing/insufficient.
- Update local role/Profile configuration for Companion/Intake/Orchestrator/Coder/Reviewer according to intended delegation behavior.
- Preserve existing scope validation for direct tools.
- Add tests covering: direct tool access allowed while SpawnPod delegation denied; explicit delegation grant permits appropriate subset; over-delegation rejected; old metadata/restored scope without delegation defaults to no delegation.
Implementation latitude:
- Coder may choose exact Rust type names and config field names if they clearly express direct vs delegation authority.
- Coder may stage internal compatibility for parsing older metadata as direct-scope-only/no-delegation, but must not create broad old-name delegation aliases.
- Coder may keep existing user-facing scope terminology where it remains clear, but diagnostics must distinguish direct authority from delegation authority.
Escalate if:
- The split requires a broad migration of stored metadata beyond defaulting missing delegation to none.
- Existing orchestrator/role Profiles cannot express needed delegation without a separate config model.
- Implementing this would require changing host authority grants, Ticket backend authority, or non-filesystem capabilities.
- Tests reveal SpawnPod registry/reclaim semantics assume direct scope and delegation scope are the same in a way that cannot be localized.
Validation:
- Focused tests for scope/delegation validation and SpawnPod over-delegation diagnostics.
- Profile/manifest parsing tests for direct/delegation fields.
- Pod metadata/session restore tests for missing delegation defaulting to none.
- `cargo test -p pod ...` / `cargo test -p manifest ...` / `cargo test -p client ...` focused equivalents selected by coder.
- `cargo check -q`.
- `cargo fmt --check`.
- `git diff --check`.
- `cargo run -q -p yoi -- ticket doctor`.
- Because scope/config/runtime/profile behavior is touched, final merge-completion must include `nix build .#yoi`.
Current code map:
- `crates/pod/src/tools/pod_management.rs`: `SpawnPod` input and requested child scope handling.
- `crates/pod/src/session_store.rs`, `crates/pod/src/pod_registry.rs`, and related metadata/session scope snapshot code.
- `crates/manifest/src/*`: scope allow parsing/profile/manifest config.
- `crates/tui/src/spawn.rs`, `crates/client/src/spawn.rs`, `crates/client/src/ticket_role.rs`: runtime launch/profile/scope propagation.
- Project role Profiles under `.yoi/profiles` / profile resources as applicable.
Critical risks / reviewer focus:
- Direct write scope must remain usable by the parent Pod after the split.
- `SpawnPod` must not fall back to direct scope when delegation grant is absent.
- Missing old delegation metadata must fail closed for child delegation.
- Orchestrator/coder/reviewer role configs must not accidentally grant broad delegation to non-orchestrator roles.
- Error messages must make the missing delegation grant understandable, not look like ordinary read/write scope denial.
---
<!-- event: state_changed author: orchestrator at: 2026-06-08T05:47:15Z from: queued to: inprogress reason: orchestrator_acceptance field: workflow_state -->
## State changed
Accepted queued implementation after reading the Ticket, implementation-direction decision, workspace state, and authority/scope code map. This acceptance precedes worktree creation and coder/reviewer Pod spawning.
---