You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Commit 4e8119b
Browse filesBrowse the repository at this point in the historyBrowse files
Copy file name to clipboardExpand all lines: AGENTS.md
+1Lines changed: 1 addition & 0 deletions
Display the source diff
Display the rich diff
Original file line number
Diff line number
Diff line change
@@ -49,6 +49,7 @@ Prefer the narrowest test layer that proves the behavior. This follows standard
49
49
50
50
- Read `docs/architecture/task-lifecycle-model.md` before changing task status, delegation, interruption, completion, abandonment, persistence ownership, or scheduler fan-out behavior.
51
51
- Keep lifecycle mutations in the shared reducers under `src/core/task-persistence/taskLifecycle.ts`. Update model actions, invariants, or named semantic landmarks for every new transition or concurrency bug class representable in the lifecycle model, then run `pnpm lifecycle:model-check`.
52
+
- The in-memory run-state submodel lives in `src/core/task/run-state/runState.ts`. It is a pure `nextRunState(state, event)` kernel, separate from the persisted reducers. Read `docs/architecture/task-lifecycle-target-architecture.md` before changing it.
52
53
- Add extension-host E2E coverage only for a boundary the reducer model cannot prove, such as restart visibility, real persistence/rehydration, delayed provider streams, scheduler permits, or webview task scoping. Do not duplicate reducer interleavings in E2E.
53
54
- Run the focused lifecycle tests and `pnpm test` before completing a Zoo Code lifecycle change.
Copy file name to clipboardExpand all lines: CHANGELOG.md
+51Lines changed: 51 additions & 0 deletions
Display the source diff
Display the rich diff
Original file line number
Diff line number
Diff line change
@@ -1,5 +1,56 @@
1
1
# Zoo Code Changelog
2
2
3
+
## [3.88.0]
4
+
5
+
- ✨ **Provider and model support:** Use Grok Build and improved custom-model support. This release also fixes Bedrock model limits and sends the Codex verbosity setting.
6
+
- 🧭 **Task reliability:** This release fixes cancellation handling, tool-turn continuation, and task-message saving.
7
+
- ⚡ **Faster editing and chat:** File writes no longer use artificial delays by default. Chat messages carry less file data, and code actions reach the last active chat.
8
+
9
+
### Provider and Model Support
10
+
11
+
- Add Grok Build model metadata for xAI (#407 by @edelauna, PR #1687 by @myk1yt)
12
+
- Honor custom model IDs for OpenAI-compatible providers (#1845 by @p12tic, PR #1846 by @p12tic)
13
+
- Update OpenAI sane-default parameter values for custom models (#1845 by @p12tic, PR #1847 by @p12tic)
14
+
- Add an `openAiStrictToolSchemas` setting for OpenAI-compatible gateways (#1827 by @easonLiangWorldedtech, PR #1828 by @easonLiangWorldedtech)
15
+
- Send the Codex verbosity setting in requests (#1958 by @WebMad, PR #1977 by @tosinxt)
16
+
- Expose the Opus 5.5 output token limit on Bedrock (PR #1903 by @PierrunoYT)
17
+
- Resolve custom ARN capabilities from the underlying Bedrock model (#1968 by @kw-bivanov, PR #1970 by @kw-bivanov)
18
+
19
+
### Task Reliability
20
+
21
+
- Recover completed tool-turn continuation (#1883 by @jaszhix, PR #1884 by @jaszhix)
22
+
- Make task-message saving stage-independent and finalize open partial tool asks (#1933 by @easonLiangWorldedtech, PR #1927 by @easonLiangWorldedtech)
23
+
- Check abort before an ask row posts (#1803 by @edelauna, PR #1895 by @edelauna)
24
+
- Add shared abort-signal cancellation-scope helpers (PR #1651 by @easonLiangWorldedtech)
25
+
- Forward the task abort signal in OpenAI and LM Studio handlers (#1450 by @allo-, PR #1686 by @myk1yt)
26
+
- Fix permanent DeepSeek and OpenAI request failures from lone UTF-16 surrogates in task history (#461 by @yoloseeyou, PR #1831 by @app/zoomote)
27
+
- Reject unsupported MIME types in image blocks (PR #1961 by @WebMad)
28
+
29
+
### Faster Editing and Chat
30
+
31
+
- Remove artificial write delays by default (part of #1375) (PR #1381 by @easonLiangWorldedtech)
32
+
- Omit `originalContent` from webview messages and fetch it on demand (#1885 by @daewoongoh, PR #1943 by @daewoongoh)
33
+
- Route code actions to the last active chat (#1944 by @WebMad, PR #1946 by @WebMad)
34
+
- Preserve Unicode when clipping file lines (PR #1960 by @WebMad)
35
+
- Keep the streamed read path empty until content is available (PR #1962 by @WebMad)
36
+
37
+
### Safety, Code Index, and Maintenance
38
+
39
+
- Add blanket auto-deny for unapproved commands in hands-free mode (#1569 by @DaubnerF, PR #1760 by @DaubnerF)
40
+
- Remove automatic startup embedder validation (PR #1879 by @WebMad)
41
+
- Refresh code-index status routing on workspace-folder changes (#1857 by @WebMad, PR #1957 by @tosinxt)
42
+
- Parse MCP tool names with double underscores in server names (#1731 by @myk1yt, PR #1737 by @myk1yt)
43
+
- Add the pure run-state kernel module (RSK-01) (#1791 by @edelauna, PR #1942 by @edelauna)
44
+
- Remove the dead Task.isPaused field (RSK-12) (#1793 by @edelauna, PR #1894 by @edelauna)
45
+
- Disable extensions when launching the Extension Host for debugging (#1859 by @daewoongoh, PR #1860 by @daewoongoh)
46
+
- Revert the `apply_diff` line-number prompt guidance (PR #1908 by @app/zoomote)
47
+
- Add packages/types to changed-code mutation testing (#1892 by @edelauna, PR #1893 by @app/zoomote)
48
+
- Wait for the resolved context window in the visual chat scenario (PR #1941 by @p12tic)
49
+
- Remove the stray tab before the Tool Use Guidelines heading (#1415 by @daewoongoh, PR #1440 by @easonLiangWorldedtech)
50
+
- Update posthog-js to v1.435.3 (PR #1586 by @app/renovate)
51
+
- Update posthog-node to v5.55.0 (PR #1587 by @app/renovate)
52
+
- Merge the **previous v3.86.0 release preparation** after its publication; this is prior-release housekeeping, not a v3.88.0 feature (PR #1897 by @taltas)
53
+
3
54
## [3.86.0]
4
55
5
56
- ✨ **New model support:** Use GPT-6.1 Sol, get the corrected Claude Opus 5.5 output budget, and select DeepSeek V4.1 Flash on OpenCode Go.
Copy file name to clipboardExpand all lines: README.md
+4-4Lines changed: 4 additions & 4 deletions
Display the source diff
Display the rich diff
Original file line number
Diff line number
Diff line change
@@ -46,11 +46,11 @@ Zoo Code builds on the foundation created by Roo Code and continues to expand it
46
46
-**More dependable terminal and editing workflows** — fixes for premature terminal completion, task-state races, context management, diff editing, and provider-specific tool use.
47
47
-**More control over your workspace** — rules management, per-mode MCP restrictions, multi-root path controls, model reasoning options, and completion change review actions.
48
48
49
-
## What's New in v3.86.0
49
+
## What's New in v3.88.0
50
50
51
-
- ✨ **New model support:** Use GPT-6.1 Sol, get the corrected Claude Opus 5.5 output budget, and select DeepSeek V4.1 Flash on OpenCode Go.
52
-
- 🧭 **Reliable task aborts and interruptions:**Preserve the first abort reason, reset stream-abort state for each request, and stop interrupted tasks from replaying rejected subtasks.
53
-
-🛠️**Dependable tools, prompts, and UI streaming:**Keep streamed tool arguments intact, enforce MCP tool policy, prevent invented line numbers, and smooth chat updates.
51
+
- ✨ **Provider and model support:** Use Grok Build and improved custom-model support. This release also fixes Bedrock model limits and sends the Codex verbosity setting.
-⚡**Faster editing and chat:**File writes no longer use artificial delays by default. Chat messages carry less file data, and code actions reach the last active chat.
Copy file name to clipboardExpand all lines: docs/architecture/task-lifecycle-model.md
+38-3Lines changed: 38 additions & 3 deletions
Display the source diff
Display the rich diff
Original file line number
Diff line number
Diff line change
@@ -6,15 +6,16 @@ Zoo Code checks task lifecycle protocols through one umbrella command for indepe
6
6
pnpm lifecycle:model-check
7
7
```
8
8
9
-
The baseline command runs seven independent bounded checks in sequence:
9
+
The baseline command runs eight independent bounded checks in sequence:
10
10
11
11
1. the persisted task delegation lifecycle;
12
12
2. shared-store concurrency across task-history hosts;
13
13
3. production-backed handoff reducers with an abstract provider/scheduler protocol;
14
14
4. the task cleanup protocol;
15
15
5. request-stream parser scoping;
16
-
6. completion persistence; and
17
-
7. delegated-mode reader refinement.
16
+
6. completion persistence;
17
+
7. delegated-mode reader refinement; and
18
+
8. tool-turn continuation.
18
19
19
20
The planned two-sibling fan-out protocol is intentionally outside the baseline and CI umbrella. Run it explicitly with `pnpm fanout-protocol:model-check`; it describes optional future functionality, not current production coverage.
20
21
@@ -91,6 +92,40 @@ The known-unsafe witnesses currently compare exact shortest action sequences. Th
91
92
92
93
The umbrella command also runs a separate bounded child model for in-memory abort, disposal, and provider-shutdown ordering. It models cleanup settlement and rejection as environment transitions and makes no filesystem, editor Promise, fairness, or timing-liveness claim. See [Task cleanup protocol model check](./task-cleanup-protocol-model.md).
93
94
95
+
## Tool-turn continuation model
96
+
97
+
`scripts/check-tool-turn-continuation.ts` is a separate bounded child model for the in-memory handoff between the presenter and the request loop. It also runs directly with `pnpm tool-turn:model-check`. It models one assistant turn with two tool blocks and one trailing text block. It makes no claim about tool handlers, approvals, or timing liveness.
98
+
99
+
The model maps its state to production fields as follows:
|`phase`, `index`|`Task.assistantMessageContent` partial flags, matching `tool_result` blocks in `userMessageContent`, and `currentStreamingContentIndex`|
104
+
|`stream`|`didCompleteReadingStream`|
105
+
|`lock`, `owner`, `pending`|`presentAssistantMessageLocked` and `presentAssistantMessageHasPendingUpdates`, owned by the `presentAssistantMessage` wrapper |
106
+
|`latch`|`userMessageContentReady`|
107
+
|`epoch`| the per-request reset in the request loop (`Task.ts`) |
108
+
|`poll` and `post-wait`| the `pWaitFor` condition and the abort check that follows it |
109
+
110
+
The checker enforces these invariants in every reachable state:
111
+
112
+
1. The next request never starts while a tool block has no result.
113
+
2. The next request never starts while a presenter pass is live.
114
+
3. The presenter lock is never held without a live pass, including after a presenter failure.
115
+
4. A released lock never hides a pending update, unless the task aborted or a pass failed.
116
+
5. The next request never starts after abort.
117
+
6. Every reachable state either can continue without abort or has started the request.
118
+
7. A live pass always holds the lock.
119
+
120
+
Five injected policies must each violate one named invariant: a latch-only wait (the #1883 hang), a readiness check with no lock gate, a check on the first tool result only, a presenter that keeps the lock after a failure, and a wait with no abort check. Six landmarks keep the intended paths reachable, including recovery from a lost latch, continuation after a presenter failure, a two-tool turn, abort on a ready turn, and a stale pass that unwinds after a retry.
121
+
122
+
The model has two documented gaps. The checker asserts that both still reproduce, so a fix must update the model and this section:
123
+
124
+
- A presenter failure on a tool block before it pushes a result releases the lock but leaves no result. The derived check stays false and the wait has no timeout. This hang existed before the derived check, when the lock stayed held. The change deliberately does not synthesize a tool result.
125
+
- The wrapper `finally` clears the lock without an ownership check. After a mid-stream retry resets the turn, a stale pass that unwinds late can clear the lock of a newer pass and break invariant 7. A per-pass owner token removes this gap in the model. Invariants 1 to 7 hold for that policy.
126
+
127
+
This model is not production-backed. The flags are fields on `Task`, not pure reducers. The real-presenter tests in `Task.spec.ts` and `presentAssistantMessage-custom-tool.spec.ts` tie it to production. A later change can extract the readiness check as a pure function so the checker and `Task` share it.
128
+
94
129
## Provider handoff and scheduler model
95
130
96
131
`scripts/check-provider-handoff-scheduler.ts` is a separate bounded adapter model for the runtime boundary that the persisted lifecycle graph does not represent. Its breadth-first explorer normalizes provider-keyed records and owner arrays before deduplicating canonical states, then exhaustively explores enabled action orderings through depth 15 with a 20,000-state budget. It imports `selectHandoffExecutionContext` and the existing `delegateTaskToChild` and `completeDelegatedChild` reducers. A direct saved, unsaved, and locked-profile matrix checks task-local configuration selection within those cases. Stale provider lookup is caught before this pure selector, so focused provider tests check the failed lookup, contextual log, and fallback. The protocol state then models two provider instances, their claims and parent snapshots, authoritative parent/child records, current task publication, commit/start ownership, the child scheduler permit, queued and resumed parent state, and one bounded redelegation generation.
0 commit comments