task: plan hook task state ownership

This commit is contained in:
Keisuke Hirata 2026-06-05 09:50:03 +09:00
parent 4b90fe91bf
commit d5b919e0d6
No known key found for this signature in database
10 changed files with 278 additions and 1 deletions

View File

@ -0,0 +1,14 @@
# Decision: authority handles live in Hook contexts, not Hook return effects
The internal-module and external-plugin authority split should treat host-authority APIs as handles supplied by the host, including inside Hooks.
Implications:
- Hook return values remain per-hook-point flow-control actions.
- Side effects such as durable model-visible SystemItem append are performed through typed host handles on event-specific Hook contexts.
- Built-in internal modules may receive handles according to host policy without user-facing external-plugin approval.
- Future external plugins receive only the handles allowed by their approved host authorities.
- The main API should not be “return an effect and let the host reject it at runtime.” Rejection remains defense-in-depth for malformed calls, missing handles, bounds, and policy violations.
- Do not model every authority combination as a distinct Hook context type. Use event-specific context types with authority-specific handles whose constructors are host-owned.
This preserves the clean distinction: contribution declarations are descriptor-locked; dangerous host APIs are represented by host-created handles; normal tool permission remains the per-call execution gate.

View File

@ -7,7 +7,7 @@ kind: task
priority: P1
labels: [plugin, feature-registry, permissions, architecture]
created_at: 2026-06-04T23:48:44Z
updated_at: 2026-06-04T23:50:15Z
updated_at: 2026-06-05T00:49:53Z
assignee: null
legacy_ticket: null
---

View File

@ -27,4 +27,26 @@ The external-plugin authority model remains necessary for sandbox/object-capabil
This split should be implemented separately from the Task tools extraction. The Task tools extraction should validate the contribution-only built-in module path without solving external plugin approval.
---
<!-- event: decision author: hare at: 2026-06-05T00:49:53Z -->
## Decision
# Decision: authority handles live in Hook contexts, not Hook return effects
The internal-module and external-plugin authority split should treat host-authority APIs as handles supplied by the host, including inside Hooks.
Implications:
- Hook return values remain per-hook-point flow-control actions.
- Side effects such as durable model-visible SystemItem append are performed through typed host handles on event-specific Hook contexts.
- Built-in internal modules may receive handles according to host policy without user-facing external-plugin approval.
- Future external plugins receive only the handles allowed by their approved host authorities.
- The main API should not be “return an effect and let the host reject it at runtime.” Rejection remains defense-in-depth for malformed calls, missing handles, bounds, and policy violations.
- Do not model every authority combination as a distinct Hook context type. Use event-specific context types with authority-specific handles whose constructors are host-owned.
This preserves the clean distinction: contribution declarations are descriptor-locked; dangerous host APIs are represented by host-created handles; normal tool permission remains the per-call execution gate.
---

View File

@ -0,0 +1,18 @@
# Investigation: Task state and Hook side-effect boundary
## Findings
- Hook action types are already separated per hook point after Hook hardening. The next design should preserve that: flow-control actions stay event-specific rather than becoming one global `HookDecision`.
- Hook inputs are still summary structs, not contexts with host-created handles. That is the missing abstraction for feature-owned behavior that needs durable host side effects.
- `TaskStore` is still owned by `Pod`, and `TaskReminderState`/reminder emission is still owned by `PodInterceptor`.
- The Task built-in feature module currently contributes only the four Task tools and receives `TaskStore` from Pod. This is an incomplete internal-module boundary: Task-specific state still remains on the Pod side.
- `SystemItem::TaskReminder` is currently appended through `PodInterceptor::pending_history_appends()`, which is the correct durable history direction but the wrong ownership location for Task-specific logic.
## Decision
Split follow-up into two steps:
1. Add Hook context host handles, especially a durable `SystemItem` append handle. Hook returns remain per-hook-point flow-control actions. No raw `Item` injection and no generic effect/event channel.
2. Move TaskStore and Task reminder logic into the Task feature module, implemented as Task-owned tools plus hooks that use the host-provided SystemItem append handle.
This keeps Pod responsible for generic host surfaces and history authority, while Task owns Task-specific state and policy.

View File

@ -0,0 +1,73 @@
---
id: 20260605-004807-hook-context-system-item-sink
slug: hook-context-system-item-sink
title: Hook: add context handles for host-mediated SystemItem append
status: open
kind: feature
priority: P1
labels: [hooks, feature-registry, history, task-reminder]
created_at: 2026-06-05T00:48:07Z
updated_at: 2026-06-05T00:49:53Z
assignee: null
legacy_ticket: null
---
## Issue
Built-in feature modules need a closed way to influence Pod-side behavior without adding feature-specific logic back into `Pod` / `PodInterceptor`. Task reminders are the immediate example: a Task feature should be able to observe request/tool activity and append a `SystemItem::TaskReminder`, but it should not receive `Pod`, `Worker`, raw history writers, raw `llm_worker::Item` injection, or a generic effect channel.
The current Hook surface already has per-hook-point action types after `hook-public-surface-hardening`, and those action types should remain flow-control oriented. The missing piece is event-specific Hook context carrying host-owned handles for allowed side effects, especially a durable `SystemItem` append path.
## Current findings
- `crates/pod/src/hook.rs` already distinguishes hook point output types such as `HookPromptAction`, `HookPreRequestAction`, `HookPreToolAction`, `HookPostToolAction`, and `HookTurnEndAction`.
- `PreLlmRequest` / `OnTurnEnd` public actions no longer expose raw `Item` injection.
- Hook inputs are currently summary structs (`PreRequestInfo`, `ToolCallSummary`, etc.), not context objects with host handles.
- `TaskReminder` emission currently lives in `PodInterceptor::pending_history_appends()` and uses Pod-owned `TaskStore` / `TaskReminderState`.
- Existing durable model-visible path is `SystemItem` commit -> `LogEntry::SystemItem` -> `Event::SystemItem` / model context, and this path must remain the only model-visible append mechanism.
## Direction
Introduce event-specific Hook context objects that carry narrow, host-created handles. Keep Hook return values as per-hook-point flow-control actions.
- Side effects are not returned as arbitrary `HookEffect` enums.
- Side effects are requested through typed handles on the Hook context.
- Handles have private constructors and are created only by the Pod host.
- Missing authority/handle should prevent installation or make the handle unavailable; runtime rejection is defense-in-depth, not the normal API path.
## Requirements
- Define event-specific Hook context types, or evolve `HookEventKind::Input` into context values without losing the existing per-event action types.
- `PreLlmRequest` context should be able to carry request summary plus a SystemItem append handle where the host grants it.
- Tool hook contexts should continue to expose bounded call/result summaries.
- Add a host-owned `SystemItemAppendHandle` / sink concept for model-visible durable system items.
- It must not expose raw `llm_worker::Item` or raw history writer access.
- It should append/push only approved `SystemItem` request kinds, initially including TaskReminder/Notification-like system items as needed.
- Actual commit timing remains host-controlled through the existing pending history append / `LogEntry::SystemItem` / `Event::SystemItem` path.
- Preserve per-hook-point flow-control action types.
- Do not collapse every hook into one generic `HookDecision`.
- Do not reintroduce `ContinueWith(Vec<Item>)`, no-result tool skip, arbitrary `ToolResult` construction, or hidden context mutation.
- Keep internal built-in module usage distinct from external-plugin authority approval.
- Built-in modules may receive host handles according to host policy.
- Future external plugins receive model-visible append handles only when the external-plugin authority model grants them.
- Provide tests showing:
- Hook actions remain flow-control only;
- model-visible side effects use the SystemItem append handle;
- no raw `Item` / raw history writer / raw event sender is exposed through public Hook context;
- missing/unavailable handles do not produce invisible context injection.
## Non-goals
- Moving TaskStore/reminder state into the Task feature; that is tracked by `task-feature-own-store-reminder-hooks`.
- TUI dialog/custom UI work.
- Generic plugin event channels.
- External plugin loading/WASM boundary implementation.
- Changing ToolRegistry / PreToolCall permission behavior.
## Acceptance criteria
- Hook contexts can carry host-owned typed handles for allowed side effects, starting with durable SystemItem append.
- Hook return values remain event-specific flow-control actions.
- Model-visible additions through Hooks are committed through the durable SystemItem path and are user-inspectable.
- Public Hook APIs still do not expose raw `Item` injection, raw Pod/Worker/session internals, or generic UI/event payload channels.
- The resulting API is sufficient for a later Task reminder Hook to append `SystemItem::TaskReminder` without Pod-specific Task logic in `PodInterceptor`.

View File

@ -0,0 +1,33 @@
<!-- event: create author: tickets.sh at: 2026-06-05T00:48:07Z -->
## Created
Created by tickets.sh create.
---
<!-- event: decision author: hare at: 2026-06-05T00:49:53Z -->
## Decision
# Investigation: Task state and Hook side-effect boundary
## Findings
- Hook action types are already separated per hook point after Hook hardening. The next design should preserve that: flow-control actions stay event-specific rather than becoming one global `HookDecision`.
- Hook inputs are still summary structs, not contexts with host-created handles. That is the missing abstraction for feature-owned behavior that needs durable host side effects.
- `TaskStore` is still owned by `Pod`, and `TaskReminderState`/reminder emission is still owned by `PodInterceptor`.
- The Task built-in feature module currently contributes only the four Task tools and receives `TaskStore` from Pod. This is an incomplete internal-module boundary: Task-specific state still remains on the Pod side.
- `SystemItem::TaskReminder` is currently appended through `PodInterceptor::pending_history_appends()`, which is the correct durable history direction but the wrong ownership location for Task-specific logic.
## Decision
Split follow-up into two steps:
1. Add Hook context host handles, especially a durable `SystemItem` append handle. Hook returns remain per-hook-point flow-control actions. No raw `Item` injection and no generic effect/event channel.
2. Move TaskStore and Task reminder logic into the Task feature module, implemented as Task-owned tools plus hooks that use the host-provided SystemItem append handle.
This keeps Pod responsible for generic host surfaces and history authority, while Task owns Task-specific state and policy.
---

View File

@ -0,0 +1,84 @@
---
id: 20260605-004807-task-feature-own-store-reminder-hooks
slug: task-feature-own-store-reminder-hooks
title: Task: move TaskStore and reminders into Task feature
status: open
kind: task
priority: P1
labels: [tasks, hooks, feature-registry, history]
created_at: 2026-06-05T00:48:07Z
updated_at: 2026-06-05T00:49:53Z
assignee: null
legacy_ticket: null
---
## Issue
`TaskCreate` / `TaskUpdate` / `TaskGet` / `TaskList` have been extracted into a built-in internal feature module, but Task state is still Pod-owned. `Pod` owns `TaskStore`, `PodInterceptor` owns `TaskReminderState`, and `PodInterceptor::pending_history_appends()` contains Task-specific reminder logic.
For the feature/module boundary to be meaningful, Task-specific state and behavior should live in the Task feature module. Pod should provide generic host surfaces: tool registration, hook dispatch, and durable `SystemItem` append. Pod should not know about TaskStore or Task reminder rules.
## Current findings
- `crates/pod/src/feature/builtin/task.rs` currently constructs the Task tool feature with a host-provided `tools::TaskStore`.
- `crates/pod/src/pod.rs` still stores `task_store: tools::TaskStore` and `task_reminder_state: Arc<TaskReminderState>`.
- `crates/pod/src/ipc/interceptor.rs` still stores `TaskStore` and `TaskReminderState`, detects task-management tool calls, and emits `SystemItem::TaskReminder` from `pending_history_appends()`.
- `crates/pod/src/pod.rs` also uses TaskStore snapshots for session restore/compaction/reminder context. Those uses must be audited before removing Pod ownership.
- TUI-facing Task visibility is intentionally out of scope here; if TUI needs a compatibility/status surface, create a follow-up instead of keeping Task ownership in Pod.
## Direction
Move TaskStore and Task reminder state into the Task feature module after `hook-context-system-item-sink` provides a host-mediated SystemItem append handle for Hook contexts.
Target shape:
- `TaskFeature` owns:
- `tools::TaskStore`
- Task reminder/cooldown state
- Task tool contribution construction
- hooks that track Task tool usage and decide when to emit reminders
- Pod owns:
- feature registry
- Hook dispatch
- ToolRegistry integration
- durable SystemItem append/commit authority
- generic history/session machinery
- Pod does not own or special-case TaskStore / TaskReminderState.
## Requirements
- Depend on or first implement `hook-context-system-item-sink`.
- Move TaskStore construction/ownership from `Pod` into the built-in Task feature module.
- Preserve session restore behavior by giving Task feature whatever history snapshot or restore input it needs, rather than keeping TaskStore in Pod.
- Preserve one session-lifetime TaskStore shared by all Task tools and reminder hooks.
- Move TaskReminderState and reminder decision logic out of `PodInterceptor` and into Task feature-owned hooks.
- A tool hook records Task tool usage.
- A `PreLlmRequest` hook evaluates inactivity/cooldown and appends `SystemItem::TaskReminder` through the host-provided SystemItem append handle.
- Remove Task-specific checks from `PodInterceptor` such as direct `is_task_management_tool` handling for reminder state.
- Preserve current observable behavior:
- Task tool names/schemas/descriptions/outputs;
- TaskStore snapshot/restore semantics;
- task reminder threshold/cooldown/body/source behavior;
- model-visible history path via `LogEntry::SystemItem` / `Event::SystemItem`;
- normal ToolRegistry / PreToolCall permission path.
- Audit compaction/session snapshot code that currently reads `self.task_store`.
- Either route this through a Task feature status/snapshot service, or leave a documented temporary compatibility façade with no Task ownership in Pod.
- Do not silently drop TaskStore snapshot/resume behavior.
- Keep external-plugin authority model out of this ticket except insofar as the Hook SystemItem append handle is used by a trusted built-in module.
## Non-goals
- TUI UI changes.
- External plugin loading or package approval.
- Generic event channels or dialogs.
- Changing Task tool behavior or adding/removing Task tools.
- Reworking Memory/WorkItem/Pod-management modules.
## Acceptance criteria
- Task feature module owns TaskStore and Task reminder state.
- Pod and PodInterceptor no longer contain Task-specific store/reminder state or task-tool special-casing.
- Task reminder emission is implemented as Task feature Hook logic using host-mediated SystemItem append, not raw `Item` injection.
- TaskStore snapshot/restore/compaction behavior is preserved or explicitly routed through a new feature-owned status/snapshot surface.
- Existing Task tool and Task reminder tests are moved/updated and pass.
- Workspace validation passes.

View File

@ -0,0 +1,33 @@
<!-- event: create author: tickets.sh at: 2026-06-05T00:48:07Z -->
## Created
Created by tickets.sh create.
---
<!-- event: decision author: hare at: 2026-06-05T00:49:53Z -->
## Decision
# Investigation: Task state and Hook side-effect boundary
## Findings
- Hook action types are already separated per hook point after Hook hardening. The next design should preserve that: flow-control actions stay event-specific rather than becoming one global `HookDecision`.
- Hook inputs are still summary structs, not contexts with host-created handles. That is the missing abstraction for feature-owned behavior that needs durable host side effects.
- `TaskStore` is still owned by `Pod`, and `TaskReminderState`/reminder emission is still owned by `PodInterceptor`.
- The Task built-in feature module currently contributes only the four Task tools and receives `TaskStore` from Pod. This is an incomplete internal-module boundary: Task-specific state still remains on the Pod side.
- `SystemItem::TaskReminder` is currently appended through `PodInterceptor::pending_history_appends()`, which is the correct durable history direction but the wrong ownership location for Task-specific logic.
## Decision
Split follow-up into two steps:
1. Add Hook context host handles, especially a durable `SystemItem` append handle. Hook returns remain per-hook-point flow-control actions. No raw `Item` injection and no generic effect/event channel.
2. Move TaskStore and Task reminder logic into the Task feature module, implemented as Task-owned tools plus hooks that use the host-provided SystemItem append handle.
This keeps Pod responsible for generic host surfaces and history authority, while Task owns Task-specific state and policy.
---