ticket: use base32 project record ids

This commit is contained in:
2026-06-09 22:10:47 +09:00
parent 0803bc3725
commit 4203988d74
798 changed files with 477 additions and 105 deletions
@@ -0,0 +1,63 @@
# Delegation intent: unify reasoning persistence with block lifecycle
Intent:
- Implement the `unify-reasoning-block-lifecycle` ticket by removing the separate `ReasoningItem` side channel and making the reasoning/thinking block lifecycle the single authoritative path for live streaming and persistence.
Requirements:
- Start by mapping the current reasoning paths in `llm-worker`: Anthropic thinking stream, OpenAI Responses reasoning items, Timeline block handlers, `ReasoningItemCollector`, Worker finalization, and request serialization round-trip tests.
- Remove the separate finalized reasoning event path:
- `ReasoningItemEvent`;
- `ReasoningItemKind`;
- `ReasoningItemCollector`;
- `Timeline::on_reasoning_item` / `dispatch_reasoning_item` / `reasoning_item_handlers`.
- Extend the reasoning/thinking block lifecycle so stop/finalization carries all provider material needed to build `Item::Reasoning`:
- text;
- reasoning item id;
- summary;
- encrypted content;
- Anthropic thinking signature;
- redacted thinking payload metadata.
- Convert Anthropic `thinking_delta` / `signature_delta` / redacted thinking handling to finalize through reasoning block stop metadata, without emitting a separate `ReasoningItem` event.
- Convert OpenAI Responses completed reasoning items into reasoning block lifecycle events, including metadata-only reasoning where there is no streamed text delta.
- Update Worker collection/finalization so `Item::Reasoning` is built from reasoning block lifecycle state.
- Preserve live streaming thinking/reasoning callbacks for UI/trace consumers.
- Preserve persisted reasoning history round-trip behavior for Anthropic and OpenAI Responses.
- Remove misleading comments that treat reasoning as meta/single-event content.
- Do not add backward compatibility aliases or keep duplicate old/new reasoning concepts.
Invariants:
- Do not drop provider material required for reasoning round-trip (`signature`, redacted thinking metadata, `id`, `summary`, `encrypted_content`).
- Do not hide model-affecting reasoning persistence in a context-only path; persisted reasoning must remain explainable through committed history items.
- Do not redesign unrelated provider request serialization.
- Do not read ignored secret-like file contents.
- Do not edit the parent workspace; work only in the delegated worktree.
- Do not close the ticket, merge the branch, delete worktrees, or push.
Non-goals:
- No E2E spawned-process test framework.
- No product-level behavior changes beyond event model cleanup.
- No dependency changes.
- No compatibility layer for the removed `ReasoningItem` event path.
Escalate if:
- A provider requires reasoning persistence material that cannot naturally fit in block lifecycle metadata.
- OpenAI Responses reasoning items cannot be represented as synthetic block lifecycle events without losing ordering or identity semantics.
- The change requires broad public API churn outside `llm-worker` and its direct consumers.
- Existing tests imply a behavior conflict between live streaming callbacks and persistence.
Validation:
- Run focused reasoning/timeline/provider tests that cover Anthropic thinking signature/redacted material and OpenAI Responses `id` / `summary` / `encrypted_content` round-trip.
- Run `cargo test -p llm-worker --lib`.
- Run `cargo check --workspace --all-targets`.
- Run `./tickets.sh doctor` and `git diff --check`.
- Run `nix build .#yoi` if feasible; record if skipped and why.
- Commit the implementation in the worktree when reviewable.
Completion report:
- investigation summary;
- implementation summary;
- changed files;
- commit hash(es);
- validation commands and results;
- unresolved risks or parent decisions needed;
- whether ready for external review.
@@ -0,0 +1,41 @@
# Implementation report: reasoning block lifecycle
## Investigation
Initial implementation unified reasoning persistence through `BlockStop.reasoning`, but OpenAI Responses text-bearing reasoning items still had two Thinking lifecycles:
1. `response.reasoning_text.delta` streamed through the real reasoning content-part block.
2. `response.content_part.done` stopped that block with no persistence metadata.
3. `response.output_item.done` emitted a second synthetic metadata-only Thinking `BlockStart`/`BlockStop` pair.
That preserved persistence but changed live callback semantics: UI/trace consumers that listen to Thinking block stop callbacks could observe an extra empty Thinking stop after the real streamed reasoning block.
## Fix summary
OpenAI Responses now defers the stop for reasoning `content_part.done` when the part is a Thinking/reasoning-text content block. At `response.output_item.done`, the provider finalizes the deferred existing block with `ReasoningBlockData` instead of creating a second synthetic live-visible block.
Thinking block handler scopes are also keyed by block index, so a deferred reasoning-text stop still uses its original streamed buffer even if another Thinking block (for example a reasoning summary block) starts and stops before `output_item.done`.
Metadata-only reasoning items with no reasoning content-part still emit a synthetic metadata-bearing Thinking block so encrypted/id-only reasoning can be persisted and round-tripped.
The fix preserves:
- live `reasoning_text.delta` Thinking deltas;
- OpenAI Responses `id`, `summary`, and `encrypted_content` persistence;
- a single Thinking lifecycle for text-bearing reasoning items;
- metadata-only reasoning coverage.
## Validation
Passed:
- `cargo test -p llm-worker openai_responses::events::tests::reasoning --lib`
- `cargo test -p llm-worker --lib`
- `cargo check --workspace --all-targets`
- `./tickets.sh doctor`
- `git diff --check`
- `nix build .#yoi`
## Residual risk
The provider delays the stop event for OpenAI Responses reasoning text blocks until `response.output_item.done` so final encrypted/summary metadata can be attached to the same block. This avoids duplicate live stops but means the block stop is slightly later than the raw `content_part.done` SSE for reasoning text. This is intentional for the unified persistence model and covered by focused provider tests for the reviewed sequence.
@@ -0,0 +1,57 @@
# Review: unify reasoning block lifecycle
Reviewer Pods:
- Initial review: `reasoning-block-lifecycle-reviewer-20260603`
- Re-review: `reasoning-block-lifecycle-rereviewer-20260603`
## Result
Approved after fixes. No remaining blockers.
## Initial blocker
The initial reviewer found that OpenAI Responses text-bearing reasoning items produced two live-visible Thinking lifecycles:
1. streamed `reasoning_text.delta` used the real Thinking block;
2. later `response.output_item.done` emitted a second synthetic empty Thinking block to carry persistence metadata.
That preserved persistence but changed live callback/trace semantics beyond the intended event-model cleanup.
## Fix verification
Coder added commit `abb6adb fix: preserve openai reasoning live stops`.
The re-review confirmed:
- text-bearing OpenAI reasoning now attaches `ReasoningBlockData` to the deferred existing Thinking block stop;
- no second synthetic empty Thinking start/stop is emitted for text-bearing reasoning;
- metadata-only OpenAI reasoning still persists through a synthetic metadata-bearing Thinking block when no live reasoning text block exists;
- Anthropic signature/redacted material and OpenAI id/summary/encrypted_content remain carried through `ReasoningBlockData`;
- the old `ReasoningItem` event/collector/Timeline side channel was not reintroduced.
## Validation evidence
Coder reported:
- `cargo test -p llm-worker --lib`
- `cargo test -p llm-worker --test reasoning_round_trip_test`
- `cargo check --workspace --all-targets`
- `./tickets.sh doctor`
- `git diff --check`
- `nix build .#yoi`
Parent/orchestrator reran after merge:
- `cargo test -p llm-worker openai_responses::events::tests::reasoning --lib`
- `cargo test -p llm-worker --lib`
- `cargo test -p llm-worker --test reasoning_round_trip_test`
- `cargo check --workspace --all-targets`
- `./tickets.sh doctor`
- `git diff --check`
- `nix build .#yoi`
- `./result/bin/yoi pod --help`
## Residual risk
Low. OpenAI Responses reasoning-text block stop is intentionally delayed until `response.output_item.done` so provider persistence metadata can be attached to the same Thinking block lifecycle. The remaining edge risk is unusual ordering/error behavior around multiple concurrent Thinking scopes, but the implementation now keys Thinking scopes by block index and focused tests cover the reviewed duplicate-stop sequence.
+56
View File
@@ -0,0 +1,56 @@
---
title: "Unify reasoning persistence with block lifecycle"
state: "closed"
created_at: "2026-06-03T00:11:24Z"
updated_at: "2026-06-03T02:15:02Z"
---
## Background
Yoi currently has two reasoning-related event paths in `llm-worker`:
- streaming thinking/reasoning content is represented as block lifecycle events (`BlockStart` / `BlockDelta` / `BlockStop`);
- finalized reasoning persistence is represented separately as `ReasoningItemEvent` / `ReasoningItemCollector` and later converted into `Item::Reasoning`.
This duplicates the same conceptual model. Thinking and reasoning are provider vocabulary for the same class of generated model content, and generated content should be represented through the block lifecycle. `ReasoningItem` exists because the current reasoning block stop event does not carry the provider material needed for persistence, not because it is a distinct domain concept.
Backward compatibility is not required for this cleanup. Prefer a clear event model over compatibility aliases or parallel old/new APIs.
## Requirements
- Make the block lifecycle the authoritative path for reasoning/thinking generation and persistence.
- Remove the separate finalized reasoning side channel:
- `ReasoningItemEvent`;
- `ReasoningItemKind`;
- `ReasoningItemCollector`;
- `Timeline::on_reasoning_item` / `dispatch_reasoning_item` and related handler storage.
- Extend the reasoning/thinking block representation so block completion can carry all provider material required to build `Item::Reasoning`, including where applicable:
- reasoning item id;
- text;
- summary;
- encrypted content;
- Anthropic thinking signature;
- redacted thinking payload metadata.
- Update Anthropic streaming conversion so `thinking_delta` / `signature_delta` are accumulated in the block lifecycle and finalized through reasoning block stop metadata, without emitting a separate `ReasoningItem` event.
- Update OpenAI Responses conversion so completed reasoning items are represented as reasoning block lifecycle events, including metadata-only reasoning where there is no text delta.
- Update Worker collection/finalization so `Item::Reasoning` is built from reasoning block lifecycle state rather than `ReasoningItemCollector`.
- Preserve live streaming thinking/reasoning callbacks for UI/trace consumers.
- Preserve persisted reasoning history round-trip behavior for Anthropic and OpenAI Responses.
- Remove or rename misleading comments that imply reasoning is a meta/single event rather than generated content.
- Do not add backward-compatible aliases or keep the old `ReasoningItem` event path.
## Non-goals
- Do not redesign provider request serialization beyond what is needed to preserve existing reasoning round-trip behavior.
- Do not add a new E2E spawned-process test framework.
- Do not change public product behavior except for the internal event model cleanup.
- Do not hide missing provider material by dropping signatures, summaries, ids, or encrypted content.
## Acceptance criteria
- `ReasoningItemEvent` / `ReasoningItemCollector` and their Timeline handler path are removed.
- Anthropic thinking stream still produces live reasoning/thinking block callbacks and persists `Item::Reasoning` with signature/redacted material intact.
- OpenAI Responses reasoning items still persist and round-trip `id`, `summary`, and `encrypted_content` as before.
- Existing reasoning persistence tests are updated to assert the block-lifecycle path rather than the removed side channel.
- No compatibility aliases or duplicate reasoning concepts remain unless justified in the implementation report.
- Validation includes focused `llm-worker` reasoning tests, `cargo test -p llm-worker --lib`, `cargo check --workspace --all-targets`, `./tickets.sh doctor`, `git diff --check`, and `nix build .#yoi` unless explicitly deferred with rationale.
+1
View File
@@ -0,0 +1 @@
Unified reasoning persistence with the Thinking block lifecycle. Removed the separate ReasoningItem event/collector path; Anthropic and OpenAI Responses reasoning metadata now persist via block stop metadata; reviewer blocker fixed; focused tests and nix build passed.
+152
View File
@@ -0,0 +1,152 @@
<!-- event: create author: tickets.sh at: 2026-06-03T00:11:24Z -->
## Created
Created by tickets.sh create.
---
<!-- event: plan author: hare at: 2026-06-03T00:51:09Z -->
## Plan
# Delegation intent: unify reasoning persistence with block lifecycle
Intent:
- Implement the `unify-reasoning-block-lifecycle` ticket by removing the separate `ReasoningItem` side channel and making the reasoning/thinking block lifecycle the single authoritative path for live streaming and persistence.
Requirements:
- Start by mapping the current reasoning paths in `llm-worker`: Anthropic thinking stream, OpenAI Responses reasoning items, Timeline block handlers, `ReasoningItemCollector`, Worker finalization, and request serialization round-trip tests.
- Remove the separate finalized reasoning event path:
- `ReasoningItemEvent`;
- `ReasoningItemKind`;
- `ReasoningItemCollector`;
- `Timeline::on_reasoning_item` / `dispatch_reasoning_item` / `reasoning_item_handlers`.
- Extend the reasoning/thinking block lifecycle so stop/finalization carries all provider material needed to build `Item::Reasoning`:
- text;
- reasoning item id;
- summary;
- encrypted content;
- Anthropic thinking signature;
- redacted thinking payload metadata.
- Convert Anthropic `thinking_delta` / `signature_delta` / redacted thinking handling to finalize through reasoning block stop metadata, without emitting a separate `ReasoningItem` event.
- Convert OpenAI Responses completed reasoning items into reasoning block lifecycle events, including metadata-only reasoning where there is no streamed text delta.
- Update Worker collection/finalization so `Item::Reasoning` is built from reasoning block lifecycle state.
- Preserve live streaming thinking/reasoning callbacks for UI/trace consumers.
- Preserve persisted reasoning history round-trip behavior for Anthropic and OpenAI Responses.
- Remove misleading comments that treat reasoning as meta/single-event content.
- Do not add backward compatibility aliases or keep duplicate old/new reasoning concepts.
Invariants:
- Do not drop provider material required for reasoning round-trip (`signature`, redacted thinking metadata, `id`, `summary`, `encrypted_content`).
- Do not hide model-affecting reasoning persistence in a context-only path; persisted reasoning must remain explainable through committed history items.
- Do not redesign unrelated provider request serialization.
- Do not read ignored secret-like file contents.
- Do not edit the parent workspace; work only in the delegated worktree.
- Do not close the ticket, merge the branch, delete worktrees, or push.
Non-goals:
- No E2E spawned-process test framework.
- No product-level behavior changes beyond event model cleanup.
- No dependency changes.
- No compatibility layer for the removed `ReasoningItem` event path.
Escalate if:
- A provider requires reasoning persistence material that cannot naturally fit in block lifecycle metadata.
- OpenAI Responses reasoning items cannot be represented as synthetic block lifecycle events without losing ordering or identity semantics.
- The change requires broad public API churn outside `llm-worker` and its direct consumers.
- Existing tests imply a behavior conflict between live streaming callbacks and persistence.
Validation:
- Run focused reasoning/timeline/provider tests that cover Anthropic thinking signature/redacted material and OpenAI Responses `id` / `summary` / `encrypted_content` round-trip.
- Run `cargo test -p llm-worker --lib`.
- Run `cargo check --workspace --all-targets`.
- Run `./tickets.sh doctor` and `git diff --check`.
- Run `nix build .#yoi` if feasible; record if skipped and why.
- Commit the implementation in the worktree when reviewable.
Completion report:
- investigation summary;
- implementation summary;
- changed files;
- commit hash(es);
- validation commands and results;
- unresolved risks or parent decisions needed;
- whether ready for external review.
---
<!-- event: review author: hare at: 2026-06-03T02:15:01Z status: approve -->
## Review: approve
# Review: unify reasoning block lifecycle
Reviewer Pods:
- Initial review: `reasoning-block-lifecycle-reviewer-20260603`
- Re-review: `reasoning-block-lifecycle-rereviewer-20260603`
## Result
Approved after fixes. No remaining blockers.
## Initial blocker
The initial reviewer found that OpenAI Responses text-bearing reasoning items produced two live-visible Thinking lifecycles:
1. streamed `reasoning_text.delta` used the real Thinking block;
2. later `response.output_item.done` emitted a second synthetic empty Thinking block to carry persistence metadata.
That preserved persistence but changed live callback/trace semantics beyond the intended event-model cleanup.
## Fix verification
Coder added commit `abb6adb fix: preserve openai reasoning live stops`.
The re-review confirmed:
- text-bearing OpenAI reasoning now attaches `ReasoningBlockData` to the deferred existing Thinking block stop;
- no second synthetic empty Thinking start/stop is emitted for text-bearing reasoning;
- metadata-only OpenAI reasoning still persists through a synthetic metadata-bearing Thinking block when no live reasoning text block exists;
- Anthropic signature/redacted material and OpenAI id/summary/encrypted_content remain carried through `ReasoningBlockData`;
- the old `ReasoningItem` event/collector/Timeline side channel was not reintroduced.
## Validation evidence
Coder reported:
- `cargo test -p llm-worker --lib`
- `cargo test -p llm-worker --test reasoning_round_trip_test`
- `cargo check --workspace --all-targets`
- `./tickets.sh doctor`
- `git diff --check`
- `nix build .#yoi`
Parent/orchestrator reran after merge:
- `cargo test -p llm-worker openai_responses::events::tests::reasoning --lib`
- `cargo test -p llm-worker --lib`
- `cargo test -p llm-worker --test reasoning_round_trip_test`
- `cargo check --workspace --all-targets`
- `./tickets.sh doctor`
- `git diff --check`
- `nix build .#yoi`
- `./result/bin/yoi pod --help`
## Residual risk
Low. OpenAI Responses reasoning-text block stop is intentionally delayed until `response.output_item.done` so provider persistence metadata can be attached to the same Thinking block lifecycle. The remaining edge risk is unusual ordering/error behavior around multiple concurrent Thinking scopes, but the implementation now keys Thinking scopes by block index and focused tests cover the reviewed duplicate-stop sequence.
---
<!-- event: close author: hare at: 2026-06-03T02:15:02Z status: closed -->
## Closed
Unified reasoning persistence with the Thinking block lifecycle. Removed the separate ReasoningItem event/collector path; Anthropic and OpenAI Responses reasoning metadata now persist via block stop metadata; reviewer blocker fixed; focused tests and nix build passed.
---