Conversation
|
commit: |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 1506063502
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
| nextTurnId: state.turnIndex.nextTurnId, | ||
| branch: store.ref.branch, | ||
| head: deps.stores.tree.openBranch(store.ref.branch).head, | ||
| tokensBefore: estimateUsedContextTokens(state.history), |
There was a problem hiding this comment.
When usage metadata is unavailable and the system prompt or tool definitions are large, onBeforeStep calculates usage including those prefixes, while this snapshot records tokensBefore from history alone. The controller later assigns this smaller value to lastCompactedTokens, so if the fixed prefix still puts the compacted context above the threshold, the continuation immediately triggers another compaction; every round repeats the mismatch and can loop through paid summary requests without reaching the normal LLM call. Record and compare token counts using the same prefix-aware calculation.
Useful? React with 👍 / 👎.
| ({ context }) => { | ||
| if (context.cause !== 'user-abort') { | ||
| deps.actor.send({ type: 'input.continue' }); | ||
| } |
There was a problem hiding this comment.
When budget compaction is triggered before a prompt's first LLM step and summarization then fails, the before-step hook has already ended that turn without producing a response. This cancellation path sends only input.continue, but the idle agent resumes on that event only when history ends mid-tool-chain; a history ending with the triggering user message therefore stays idle, while the prompt has already settled as successfully completed with no answer. Auto-compaction failure needs to requeue a continuation or otherwise restart the blocked turn.
Useful? React with 👍 / 👎.
| return { | ||
| history: state.history, | ||
| queue: state.queue, | ||
| nextTurnId: state.turnIndex.nextTurnId, |
There was a problem hiding this comment.
An idle agent may legitimately have persisted input.remind entries waiting for its next turn, but the quiesce snapshot captures only history and queue. The fresh compaction branch therefore initializes reminders as empty, and replay starts after this snapshot's head, so any reminder already present before compaction is silently discarded and never delivered. Include the current reminder state in the seed, along with any other pending input state that must survive the reset.
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 619717435b
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
| function defaultContinuation(reason: CompactionReason): UserMessage | undefined { | ||
| if (reason === 'manual') return undefined; | ||
| return compactionContinuationMessage(); |
There was a problem hiding this comment.
When manual compaction begins while a turn is executing tools, the pause ends that turn at the next tool-result boundary, and branch switching replaces the mid-tool-chain history with user messages plus the summary. Returning no continuation for every manual run then leaves the reset agent idle, so the original prompt is reported as done without ever producing its final response. Enqueue a continuation when a manual compaction actually interrupted a running turn.
Useful? React with 👍 / 👎.
| deps.actor.send({ type: 'input.pause' }); | ||
| const waiting = waitFor(deps.actor, (s) => s.matches('idle'), { timeout: PAUSE_TIMEOUT_MS }); | ||
| void waiting.catch(() => undefined); | ||
| await Promise.race([waiting, aborted(signal)]); |
There was a problem hiding this comment.
When the agent has detached background tools, its state is idle.waiting, which still satisfies this broad matches('idle') predicate. Compaction therefore proceeds to store.reset; the agent reset aborts its scope and clears background, silently terminating those tools and discarding their eventual completion notifications. Quiescing should wait for an idle state with no background work, or explicitly preserve the detached tasks across the reset.
Useful? React with 👍 / 👎.
| goal_control: z.enum(['pause', 'resume', 'cancel']).optional(), | ||
| disabled_tools: z.array(z.string()).optional(), | ||
| prompt_id: z.string().min(1).optional(), | ||
| steer: z.boolean().optional(), |
There was a problem hiding this comment.
This adds a documented, user-visible prompt submission capability, but the reviewed diff contains no .changeset/ entry, so the CLI release changelog and versioning will omit it. Add the required user-facing changeset before merging.
AGENTS.md reference: AGENTS.md:L86-L87
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 1814aa78d5
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
| deps.actor.on('turn.aborting', () => { | ||
| active?.actor.send({ type: 'cancel', cause: 'user-abort' }); | ||
| }), |
There was a problem hiding this comment.
Once budget compaction has quiesced the agent, the agent is paused in idle, where input.abort has no handler and therefore emits no turn.aborting; this subscription is consequently unable to cancel the active summarization. An abort issued during that potentially long or paid phase does nothing, and a successful automatic compaction subsequently submits its continuation and restarts work the user tried to stop. Ensure aborts reach the active compaction even when no turn is running.
Useful? React with 👍 / 👎.
| const seedStore = await createEventStore({ journal, slices: agentSlices }); | ||
| try { | ||
| await seedStore.dispatch([...opts.seed]); |
There was a problem hiding this comment.
When the original agent store has a registered stateful slice such as todoSlice, this fresh branch is populated only with the fixed agentSlices and the compaction seed's history/queue events. Resetting the original store then refolds its registered todo slice over that branch without any state.updated event, so a nonempty structured TodoList becomes empty; appending rendered todo text to the LLM summary does not restore what the TodoList tool or UI reads. Seed the current registered state, including the todo snapshot, before resetting.
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 5f639d6e62
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
| compactionStarted(): boolean { | ||
| if (this.compacting) return false; | ||
| this.compacting = true; | ||
| if (this.status !== 'running' || this.turn === undefined) return false; |
There was a problem hiding this comment.
When POST /sessions/{id}:compact starts manual compaction, AgentFullCompactionService.begin requires the loop to be idle, so this tracker normally has status === 'idle' and no turn. This guard therefore suppresses the agent.state update, and because this change also removes compacting from session.state, clients receive no indication that the asynchronous manual compaction is running. Transition idle agents to a running { status: 'compacting' } turn here as well.
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: be69ace220
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
| index: number, | ||
| member: SwarmMemberResult | undefined, | ||
| ): void => { | ||
| if (tasks.has(agentId)) return; |
There was a problem hiding this comment.
When a later AgentSwarm call resumes an agent that appeared in an earlier call, tasks already contains that agent ID, so this return discards the later run's status, result, and timeline placement. Cold history consequently keeps showing the first run even though the live projector has updated the member; update the existing task or key synthesized tasks by both tool call and agent.
Useful? React with 👍 / 👎.
| deps.actor.on('turn.done', () => { | ||
| overflowAttempts = 0; | ||
| }), | ||
| deps.actor.on('turn.failed', (event) => { | ||
| if (!isContextOverflowError(event.error) || overflowAttempts >= maxAutoAttempts) { |
There was a problem hiding this comment.
After an overflow triggers compaction, overflowAttempts is reset only by turn.done. If the resumed turn is aborted, the next unrelated prompt inherits the previous attempt count; at the configured cap its first context-overflow failure is no longer compacted and is surfaced immediately. Reset the counter on aborted turns and on terminal non-overflow failures while preserving it only across the same overflow recovery chain.
Useful? React with 👍 / 👎.
| const current = { actor, reason, startedAt: Date.now() }; | ||
| active = current; | ||
| const subscriptions = pipeEvents(actor); | ||
| await deps.stores.session(); |
There was a problem hiding this comment.
If opening the session store rejects because of an I/O or backend failure, this await runs before the try/finally, after active and the event subscriptions have already been installed. The caller receives the error, but every subsequent manual compaction reports busy and the subscriptions leak until the controller is recreated. Include setup and actor startup in the cleanup-protected region.
Useful? React with 👍 / 👎.
| function parseSwarmMembers(output: string): SwarmMemberResult[] { | ||
| if (!output.includes('<agent_swarm_result>')) return []; | ||
| const members: SwarmMemberResult[] = []; | ||
| for (const match of output.matchAll(/<subagent\b([^>]*)>([\s\S]*?)<\/subagent>/g)) { |
There was a problem hiding this comment.
When a subagent report contains the literal text </subagent>—for example while discussing this result format—the producer inserts the report body without escaping it, so this non-greedy regex closes at that text and stores a truncated result_summary or error in cold history. Recover member results from structured events or use an encoding that cannot conflict with arbitrary report content.
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 9329c09edb
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
| const isV3 = url === WS_PATH_V3 || url.startsWith(`${WS_PATH_V3}?`); | ||
| const isDebug = url === WS_DEBUG_PATH || url.startsWith(`${WS_DEBUG_PATH}?`); | ||
| const wss = isV1 ? wssV1 : isV3 ? wssV3 : isDebug ? wssDebug : undefined; | ||
| const wss = isV3 ? wssV3 : isDebug ? wssDebug : undefined; |
There was a problem hiding this comment.
At this commit, the committed apps/kimi-code/dist-web/assets/index-HU0LCM-X.js still constructs /api/v1/ws, contains no /api/v3/ws client, and calls the removed transcript/message REST endpoints, while this branch now accepts only v3/debug WebSocket upgrades and removes those REST registrations. Consequently, the web UI shipped by kimi web cannot establish its chat socket or load history; sync and commit the code-app bundle together with the protocol removal.
AGENTS.md reference: AGENTS.md:L18-L18
Useful? React with 👍 / 👎.
| if (getLiveSessionById(this.deps.core.accessor, sessionId) !== undefined) { | ||
| return existing.view; |
There was a problem hiding this comment.
When a live session is searched once and later creates a new subagent, this returns the existing view whose roster was populated only once by loadRoster. Session-wide search derives its agent IDs from view.agents(), so it never calls ensureAgentHistory for the new subagent and silently omits that agent's messages from all subsequent searches until the cached source is dropped; refresh the roster or subscribe to agent lifecycle changes.
Useful? React with 👍 / 👎.
| const cached = this.sessions.get(sessionId); | ||
| this.sessions.delete(sessionId); | ||
| if (cached === undefined) return []; |
There was a problem hiding this comment.
Deleting a cold session, or any session that has not emitted a created/meta/activity event since this translator started, always hits this empty-cache return. The session has already been removed from ISessionIndex, so it cannot be fetched afterward, and v3 clients receive no deletion message; for example, kimi-inspect retains its seeded session entry until a later poll or unrelated invalidation. Seed the cache from existing sessions or construct a deletion entity from lifecycle data.
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: fde8ccef73
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
| function replaceEntryText(entry: UserEntry, text: string): UserEntry { | ||
| return { ...entry, message: { ...entry.message, content: [{ type: 'text', text }] } }; |
There was a problem hiding this comment.
When a boundary user message exceeds the head/tail token budget and contains images, audio, or video alongside text, this replacement rebuilds its content from a single text part and silently discards every non-text part. For example, a recent prompt containing several screenshots can cross the 20k compaction budget and lose all screenshots from the compacted branch, leaving the continued turn unable to inspect the original inputs. Preserve the non-text parts on one retained fragment while accounting for their token cost.
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 0d48c459aa
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
| agentId: file.agentId, | ||
| role: e.role, | ||
| text: e.text.length > MAX_DOC_TEXT_CHARS ? e.text.slice(0, MAX_DOC_TEXT_CHARS) : e.text, | ||
| text: e.text, |
There was a problem hiding this comment.
When a user or assistant message exceeds 20,000 characters and the only match occurs in its suffix, the persistent route now indexes the entire e.text, while collectLiveDocs in searchService.ts still truncates the same message to MAX_DOC_TEXT_CHARS. Consequently, an identical session-scoped search can return no hit while the session is live and then return a hit after it becomes cold; retaining the existing cap here also prevents exceptionally large prompts from unnecessarily inflating the persisted text indexes.
Useful? React with 👍 / 👎.
1 participant
Changes