feat: Harden the default system prompt and built-in tool descriptions - #1102
Merged
Conversation
Add high-confidence, prompt-only guardrails to the default agent system prompt: - Personality/candor: extend the HELPFUL/CONCISE/ACCURATE line with CANDID, and require plainly stating what could not be run, reproduced, or verified. - Reminders: avoid cheerleading; voice evidence-based disagreement; deliver complete code with no placeholders; update now-stale comments/docstrings after a change; re-check the user's latest request before finalizing a reply. - Context Management: explain automatic compaction — continue from the summary, re-establish transient state with tools, do not restart from scratch. - Output formatting: replies render as Markdown in the terminal; keep lists flat; no emojis unless the user uses them first. - Project Information: frame injected AGENTS.md as project context, not a privileged instruction channel that can override system rules. Prompt text only; no code or template-variable changes.
Lift a few high-leverage rules from individual tool descriptions up into the default system prompt, so they shape default behavior before any specific tool is in play (kept terse and integrated, not bolted on): - Planning: for multi-step or multi-file work, maintain a `TodoList` (one item in_progress, mark done as it finishes) and prefer `EnterPlanMode` first when the approach isn't settled. - Default to making progress, not asking: once the goal is clear and sanctioned, carry it through and work blockers yourself; ask only when the answer would change the next step. Explicitly does not override stopping to discuss an unclear goal or waiting for go-ahead before writing code. - Tool routing: prefer dedicated tools (Read/Glob/Grep/Write/Edit) over raw shell when one fits; keep Bash for genuine shell work. - Definition of done: verify with the checks that cover the change before marking it complete, independent of whether a TodoList is in use. - Delegation: explore subagents also keep intermediate file contents out of your own context — you get a conclusion back, not a pile of dumps. Prompt text only; no code or template-variable changes.
…xplore.yaml and system.md
…l tool
Subagents (coder/explore/plan) inherit the root system prompt but lack the Skill tool, yet KIMI_SKILLS was rendered unconditionally — leaking the full skill listing into agents that cannot invoke any skill. Gate KIMI_SKILLS on the profile's tool set and wrap the '# Skills' section in {% if KIMI_SKILLS %} so it disappears for those profiles.
Also note in the Working Directory section that Bash enforces none of the workspace/secret-file guards, so the model must hold that discipline itself.
Tests: assert the Skills section renders for the root agent and is absent for Skill-less subagents; update the prompt-rendering fixtures for the new gating.
…y tool availability
Subagents (coder/explore/plan) inherit the root system prompt but lack the Agent, TaskList, and TodoList tools, so they were shown usage guidance for tools they cannot call. Derive HAS_AGENT/HAS_TASKLIST/HAS_TODOLIST from each profile's tool set and gate those sections with inline {% if %}, so they render only for agents that hold the tool.
Root rendering is byte-identical (the inline tags collapse to the original text when the flag is set). The cross-tool secret-file guard stays shared, since explore/plan still hold Read/Grep/Glob.
Tests: assert the gated guidance is present for the root agent and absent for explore/plan, while the shared secret-file guard remains.
…into the tool descriptions The Agent-delegation paragraph in the system prompt duplicated mechanics already documented on the Agent tool itself (new-vs-resume, zero-context briefing, foreground default / run_in_background threshold), so remove it. HAS_AGENT still gates the explore-delegation bullet, which carries the 'when to delegate' nudge the tool description deliberately omits. Move the proactive 'anchor the pattern up front' guidance into the Glob tool description (it previously only described the reactive 'refine after hitting the cap' path) and drop the now-redundant Glob bullet from the system prompt. Tests: drop the assertions tied to the removed Agent paragraph; HAS_AGENT gating stays covered via the explore bullet.
… in agent profiles
…earch citation instructions
… fix GetGoal field doc
Surface runtime-enforced behavior in the Agent / AgentSwarm / AskUserQuestion / Goal
tool descriptions so the model learns the rules from the tool, not from a failed call:
- Agent: resuming excludes subagent_type (setting both is rejected)
- AgentSwarm: at least 2 items unless resuming, prompt_template required and must
contain {{item}}, distinct resulting prompts; plus Agent-vs-AgentSwarm fan-out note
- AskUserQuestion: result is {answers}; an empty answers with a dismissal note means
the user declined — fall back to best judgment instead of re-asking
- CreateGoal: creating fails when a goal already exists (use replace)
- SetGoalBudget: state the hard 1s-24h time-budget band
- UpdateGoal: do not mark blocked merely because work is hard/slow/incomplete
- GetGoal: drop the advertised self-report / evaluator-verdict fields — GoalSnapshot
never held them, so the tool never returned them
Each change is covered by a description assertion.
…ch the code The answers object is passed through from the host/RPC layer (QuestionAnswers is Record<string, string | true>); this code does not key it by question text. Describe what the keys identify instead of asserting a guarantee the code does not provide.
- Bash: prefer the cwd argument (or absolute paths) over a cd from an earlier call, since each call runs in a fresh shell - Grep: note that files_with_matches is ordered most-recently-modified first - Write: do not create documentation/README files unless the user asks - Edit: frame replace_all with its rename-across-file use-case Each change is covered by a description assertion.
- ExitPlanMode: describe what a good plan contains (specific, verifiable steps grounded in the codebase, not vague filler) - TodoList: stop calling it useful 'in Plan mode' — plan-mode planning goes to the plan file; TodoList tracks execution progress - CronCreate: warn that a one-shot whose pinned day/month already passed this year is rejected; document the 50-task session cap and the 8 KiB prompt cap - CronCreate: drop the bench-only KIMI_CRON_NO_STALE / KIMI_CRON_NO_JITTER env knobs from the model-facing description (CI-only; the model never sets them) Each change is covered by a description assertion.
… schema; trim EnterPlanMode workflow - ExitPlanMode: the options field mechanics (label format, recommended, count, single-option=plain-approval, reserved labels) now live only in the options param describe; the tool description routes to it and keeps the yolo/manual UI behavior it uniquely documents. The options consistency test now enforces a single source of truth (describe) plus the schema-consistency guard, instead of requiring the same facts in both surfaces. - EnterPlanMode: trim the duplicated 'What Happens in Plan Mode' steps to a pointer (the full workflow is injected unconditionally once plan mode is active), keeping the explore-subagent recommendation.
🦋 Changeset detectedLatest commit: de61971 The changes in this PR will be included in the next version bump. This PR includes changesets to release 1 package
Not sure what this means? Click here to learn what changesets are. Click here if you're a maintainer who wants to add another changeset to this PR |
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: d881b5c1d9
ℹ️ About Codex in GitHub
Your team has set up Codex to 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 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
RealKai42
changed the title
…audit Every item below was re-verified against the live code: - Skill: drop the never-fired recursion-depth cap (production never seeds depth); keep the <kimi-skill-loaded> 'already loaded, don't re-invoke' guard - TaskOutput: terminal_reason can also be `failed`, not just timed_out/stopped - Grep: count_matches emits per-file `path:count`, with the total reported separately - Plan mode: the reminder names TaskStop/CronCreate/CronDelete as blocked (they are hard-denied by plan-mode-guard-deny) - Bash: the failure trailer is non-zero-exit-specific; timeout/interrupt differ - CreateGoal: replace also covers a blocked goal, not just active/paused - UpdateGoal: it also injects the completion/blocked outcome prompt, so it does more than 'only record the status' - FetchURL: state the universal http/https contract instead of provider-internal SSRF and 10 MiB limits (the primary Moonshot fetcher enforces neither) - TodoList: query mode triggers on omitting `todos`, not on zero args - TaskList: command/PID/exit code are shell-task fields only - CronCreate: the returned fields include `cron` - SetGoalBudget: turn/token budgets are rounded up to >= 1, not rejected below 1 Each change is covered by a description/param assertion; plan.test.ts snapshots refreshed for the longer plan-mode reminder.
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: c84917d6ad
ℹ️ About Codex in GitHub
Your team has set up Codex to 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 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
…ot declared profile tools The HAS_* / Skills gating computed flags from the profile's declared tools, but Agent/AgentSwarm only register when a subagentHost exists (ToolManager .initializeBuiltinTools). A runtime built without a subagentHost (e.g. direct SDK construction) therefore rendered the explore-delegation guidance for an Agent tool the model could not call. SystemPromptContext now carries an optional availableTools; buildTemplateVars gates on it when present and falls back to the declared tools otherwise. useProfile passes the profile tools minus Agent/AgentSwarm when no subagentHost is wired, so the render reflects what the model can actually call. The normal session path (subagentHost always defaulted) is unchanged.
Plan mode writes its plan to plans/<id>.md (plan/index.ts) and the reminder tells the model to create it with Write when missing, which contradicted Write's blanket 'do not create *.md unless asked' guard. Carve the plan file out of the ban.
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: def27e67de
ℹ️ About Codex in GitHub
Your team has set up Codex to 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 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
…an to runtime reality - Gate the TodoList bullet's "enter plan mode via EnterPlanMode" suggestion on a new HAS_ENTERPLANMODE flag. A custom profile that keeps TodoList but drops EnterPlanMode no longer steers the model toward a tool it cannot call; the default profile render is unchanged. - Reframe the Write *.md prohibition around intent (unsolicited docs) instead of a blanket extension ban, so artifacts a task or project instruction requires — the plan-mode plan file, a repo-mandated changeset — are no longer contradicted by the tool's own rules.
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: d65132adfd
ℹ️ About Codex in GitHub
Your team has set up Codex to 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 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
The default system prompt carried tool-usage guidance behind {% if HAS_* %}
gates that re-derived, in prose, the availability the tool schema already
encodes — and the same guidance was duplicated in each tool's own
description. Drop the four gated blocks (background Bash, Agent/explore,
TodoList, EnterPlanMode) and the compaction TaskList/TodoList bullets; the
tool descriptions, shipped only when the tool is registered, already carry
the same instructions, so subagents and tool-trimmed profiles are no longer
pointed at tools they lack.
Fold the two genuinely unique lines into the tool descriptions: bash.md
gains "return control after starting a background task", agent.md gains the
context-hygiene reason to delegate. Collapse the compaction bullets into one
tool-agnostic sentence. Remove the now-unused availableTools / HAS_* render
machinery.
RealKai42
changed the title
RealKai42
changed the title
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 8ebfd754f1
ℹ️ About Codex in GitHub
Your team has set up Codex to 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 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: ea032f7719
ℹ️ About Codex in GitHub
Your team has set up Codex to 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 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
… docs set-goal-budget.md said turn/token budgets are "rounded up", but the code uses Math.round — say "rounded to the nearest whole number" instead. task-output.md implied every failed task carries terminal_reason/stop_reason, but a plain non-zero command exit carries only status plus exit_code; describe that exit_code path and reserve terminal_reason for non-exit endings (timeout, explicit stop, or an internal error with no exit code).
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 9ccf56ca8d
ℹ️ About Codex in GitHub
Your team has set up Codex to 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 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| 3. Design an implementation approach based on your findings | ||
| 4. Write your plan to the current plan file with Write or Edit | ||
| 5. Present your plan to the user via ExitPlanMode for approval | ||
| Once you are in plan mode, a reminder walks you through the workflow (explore → design → write the plan file → `ExitPlanMode`) and enforces read-only access. For non-trivial tasks where you are unsure of the codebase structure or relevant code paths, use `Agent(subagent_type="explore")` to investigate first. |
There was a problem hiding this comment.
When the default agent is constructed without a subagentHost (or a custom profile exposes EnterPlanMode but omits Agent), ToolManager.initializeBuiltinTools() does not register the Agent tool, but this EnterPlanMode description is still shown and directs the model to call Agent(subagent_type="explore"). Fresh evidence beyond the earlier Agent-gating comments is that the instruction now lives in EnterPlanMode's own tool description, whose availability is independent of subagentHost. This steers planning turns toward an unavailable tool call; qualify it with “if Agent is available” or remove the cross-tool instruction.
Useful? React with 👍 / 👎.
… near-term
The blast-radius paragraph told every profile that local work — including
editing files — may be done freely, but the read-only explore/plan subagents
render it too; scope it to "work your role permits" so it no longer undercuts
their read-only constraints.
The one-shot cron guidance leaned on a year-boundary heuristic ("avoid a
day/month already passed this year") that misfires across Dec 31 to Jan 1 and
duplicated a limit the code already enforces. Replace it with a plain near-term
nudge and leave the hard future-window guard in code.
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 163ca3ebf1
ℹ️ About Codex in GitHub
Your team has set up Codex to 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 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
|
|
||
| DO NOT run `git commit`, `git push`, `git reset`, `git rebase` and/or do any other git mutations unless explicitly asked to do so. Ask for confirmation each time when you need to do git mutations, even if the user has confirmed in earlier conversations. | ||
|
|
||
| Apply the same care beyond git: weigh the reversibility and blast radius of any action before you take it. Local, reversible work your role permits — editing files, running tests, reading code — you may do freely. But actions that are hard to undo or that reach beyond your local environment warrant a confirmation first: destructive ones (`rm -rf`, dropping database tables, killing processes, force-pushing, overwriting uncommitted changes) and outward-facing ones that touch shared state (pushing, opening or commenting on PRs and issues, sending messages, uploading to third-party services — which may be cached or indexed even after deletion). A one-time approval covers that one action in that one context, not a standing license: unless a durable instruction (an `AGENTS.md` entry, or an explicit request to operate autonomously) authorizes it in advance, confirm each time. Never reach for a destructive shortcut to clear an obstacle — investigate unfamiliar files, branches, or locks as possible in-progress work before deleting or overwriting them. |
There was a problem hiding this comment.
For repositories that check in an AGENTS.md entry like “open/comment on a PR when done” or “upload results,” this sentence treats that repo-supplied file as durable pre-authorization for the same outward-facing actions it says should require confirmation. Because AGENTS.md is project-controlled and later described as an unprivileged reference channel, letting it waive confirmations can cause shared-state mutations in auto/yolo contexts without an explicit user request; limit this exception to user/host policy rather than AGENTS.md.
Useful? React with 👍 / 👎.
…utput count_matches put the aggregate "Found N occurrences" summary and the "Results truncated... use offset=N to see more" notice on the result's message field, which normalizeToolResult drops before the result reaches the model. The model saw only the path:count lines and could miss the total and, worse, the pagination cue — so it would not know to page through truncated counts. Append both to output after the path:count lines, the same way the content and files_with_matches modes already inline their notices.
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 7f344b6027
ℹ️ About Codex in GitHub
Your team has set up Codex to 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 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| } | ||
| if (mode === 'count_matches' && orderedLines.length > 0) { | ||
| sideChannelMessages.push(formatCountSummary(orderedLines, filteredSensitive.size > 0)); | ||
| messages.push(formatCountSummary(orderedLines, filteredSensitive.size > 0)); |
There was a problem hiding this comment.
When count_matches returns enough or sufficiently long path:count rows to hit ToolResultBuilder's output cap, this moves the aggregate total into the same trailing messages block that is appended after all visible rows. Because builder.write(combined) truncates before it reaches those trailing messages, the model can still miss the total and the offset=N pagination cue for large count searches; keep these count-mode notices ahead of the rows or reserve budget for them before truncation.
Useful? React with 👍 / 👎.
Squash the five per-change changesets for this PR into a single concise entry; they all bump @moonshot-ai/kimi-code (patch) for the same system-prompt and tool-description hardening work.
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 9e368de42f
ℹ️ About Codex in GitHub
Your team has set up Codex to 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 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| Launch multiple subagents from one prompt template, existing agent resumes, or both. | ||
|
|
||
| Use AgentSwarm when many subagents should run the same kind of task over different inputs. The placeholder is exactly `{{item}}`. For example, with `prompt_template` set to `Review {{item}} for likely regressions.` and `items` set to `["src/a.ts", "src/b.ts"]`, AgentSwarm launches two new subagents with those two concrete prompts. | ||
| Use AgentSwarm when many subagents should run the same kind of task over different inputs. The placeholder is exactly `{{item}}`. For example, with `prompt_template` set to `Review {{item}} for likely regressions.` and `items` set to `["src/a.ts", "src/b.ts"]`, AgentSwarm launches two new subagents with those two concrete prompts. For a few differently-shaped tasks, make separate `Agent` calls in one message instead. |
There was a problem hiding this comment.
Fresh evidence beyond the earlier Agent-gating findings is this newly added AgentSwarm description: in a custom profile that exposes AgentSwarm but trims Agent, the model still receives this tool description while the Agent tool is absent from the schema, so the recommended fallback for differently-shaped tasks points it at a call it cannot make. Qualify this fallback with “if Agent is available” or describe an AgentSwarm-only alternative.
Useful? React with 👍 / 👎.
Both the Agent and Bash background-launch messages invited the model to "peek at progress" via TaskOutput, and the foreground-vs-background guidance had been thinned to a single parameter hint. Together that led the model to launch a background subagent and then immediately wait on it through TaskOutput — defeating the point of background execution. Make both launch messages take the same anti-wait stance the user-detach path already uses (do NOT wait, poll, or call TaskOutput on it), restore foreground-by-default guidance in the Agent background description (run in the background only when you have other work and do not need the result to proceed), and add a TaskOutput backstop against using it to sit and wait. Also fold the fix into the consolidated changeset.
Merged
sailist
added a commit
that referenced
this pull request
Jul 8, 2026
- read the Edit target with errors:'strict' so a non-UTF-8 file fails the
edit instead of being silently rewritten as U+FFFD (matches v1 kaos)
- declare readWriteFile access since Edit reads before it writes, matching v1
- render edit.md directly instead of through renderPrompt: it has no template
vars, and raw avoids treating literal {{ }} as a template
- restore the replace_all usage example in edit.md (v1 #1102)
- add a regression test asserting a non-UTF-8 file fails the edit and keeps
its bytes untouched
kermanx
added a commit
that referenced
this pull request
Jul 10, 2026
Drop the KIMI_CRON_NO_JITTER / KIMI_CRON_NO_STALE notes from the CronCreate and CronList descriptions and the MAX_SKILL_QUERY_DEPTH sentence from the Skill description, matching agent-core (v1) where these were trimmed from the model-facing text in #1102. Code behavior is unchanged; the env bypasses and the depth cap still exist in both implementations. ULID/8-hex wording is left as-is for now.
7723qqq
pushed a commit
to 7723qqq/kimi-code
that referenced
this pull request
Jul 11, 2026
…MoonshotAI#1102) * feat(agent-core): strengthen default system prompt Add high-confidence, prompt-only guardrails to the default agent system prompt: - Personality/candor: extend the HELPFUL/CONCISE/ACCURATE line with CANDID, and require plainly stating what could not be run, reproduced, or verified. - Reminders: avoid cheerleading; voice evidence-based disagreement; deliver complete code with no placeholders; update now-stale comments/docstrings after a change; re-check the user's latest request before finalizing a reply. - Context Management: explain automatic compaction — continue from the summary, re-establish transient state with tools, do not restart from scratch. - Output formatting: replies render as Markdown in the terminal; keep lists flat; no emojis unless the user uses them first. - Project Information: frame injected AGENTS.md as project context, not a privileged instruction channel that can override system rules. Prompt text only; no code or template-variable changes. * feat(agent-core): hoist key working rules into the system prompt Lift a few high-leverage rules from individual tool descriptions up into the default system prompt, so they shape default behavior before any specific tool is in play (kept terse and integrated, not bolted on): - Planning: for multi-step or multi-file work, maintain a `TodoList` (one item in_progress, mark done as it finishes) and prefer `EnterPlanMode` first when the approach isn't settled. - Default to making progress, not asking: once the goal is clear and sanctioned, carry it through and work blockers yourself; ask only when the answer would change the next step. Explicitly does not override stopping to discuss an unclear goal or waiting for go-ahead before writing code. - Tool routing: prefer dedicated tools (Read/Glob/Grep/Write/Edit) over raw shell when one fits; keep Bash for genuine shell work. - Definition of done: verify with the checks that cover the change before marking it complete, independent of whether a TodoList is in use. - Delegation: explore subagents also keep intermediate file contents out of your own context — you get a conclusion back, not a pile of dumps. Prompt text only; no code or template-variable changes. * fix: clarify guidelines for file pattern matching and tool usage in explore.yaml and system.md * fix(agent-core): hide the Skills section from agents without the Skill tool Subagents (coder/explore/plan) inherit the root system prompt but lack the Skill tool, yet KIMI_SKILLS was rendered unconditionally — leaking the full skill listing into agents that cannot invoke any skill. Gate KIMI_SKILLS on the profile's tool set and wrap the '# Skills' section in {% if KIMI_SKILLS %} so it disappears for those profiles. Also note in the Working Directory section that Bash enforces none of the workspace/secret-file guards, so the model must hold that discipline itself. Tests: assert the Skills section renders for the root agent and is absent for Skill-less subagents; update the prompt-rendering fixtures for the new gating. * fix(agent-core): gate Agent, background-task, and TodoList guidance by tool availability Subagents (coder/explore/plan) inherit the root system prompt but lack the Agent, TaskList, and TodoList tools, so they were shown usage guidance for tools they cannot call. Derive HAS_AGENT/HAS_TASKLIST/HAS_TODOLIST from each profile's tool set and gate those sections with inline {% if %}, so they render only for agents that hold the tool. Root rendering is byte-identical (the inline tags collapse to the original text when the flag is set). The cross-tool secret-file guard stays shared, since explore/plan still hold Read/Grep/Glob. Tests: assert the gated guidance is present for the root agent and absent for explore/plan, while the shared secret-file guard remains. * refactor(agent-core): move Agent-delegation and Glob-anchor guidance into the tool descriptions The Agent-delegation paragraph in the system prompt duplicated mechanics already documented on the Agent tool itself (new-vs-resume, zero-context briefing, foreground default / run_in_background threshold), so remove it. HAS_AGENT still gates the explore-delegation bullet, which carries the 'when to delegate' nudge the tool description deliberately omits. Move the proactive 'anchor the pattern up front' guidance into the Glob tool description (it previously only described the reactive 'refine after hitting the cap' path) and drop the now-redundant Glob bullet from the system prompt. Tests: drop the assertions tied to the removed Agent paragraph; HAS_AGENT gating stays covered via the explore bullet. * test(agent-core): add guidance for blast-radius and concrete examples in agent profiles * docs: update descriptions for skill-tool and fetch-url; enhance web-search citation instructions * feat(agent-core): disclose enforced constraints in tool descriptions; fix GetGoal field doc Surface runtime-enforced behavior in the Agent / AgentSwarm / AskUserQuestion / Goal tool descriptions so the model learns the rules from the tool, not from a failed call: - Agent: resuming excludes subagent_type (setting both is rejected) - AgentSwarm: at least 2 items unless resuming, prompt_template required and must contain {{item}}, distinct resulting prompts; plus Agent-vs-AgentSwarm fan-out note - AskUserQuestion: result is {answers}; an empty answers with a dismissal note means the user declined — fall back to best judgment instead of re-asking - CreateGoal: creating fails when a goal already exists (use replace) - SetGoalBudget: state the hard 1s-24h time-budget band - UpdateGoal: do not mark blocked merely because work is hard/slow/incomplete - GetGoal: drop the advertised self-report / evaluator-verdict fields — GoalSnapshot never held them, so the tool never returned them Each change is covered by a description assertion. * fix(agent-core): soften AskUserQuestion answers-keying wording to match the code The answers object is passed through from the host/RPC layer (QuestionAnswers is Record<string, string | true>); this code does not key it by question text. Describe what the keys identify instead of asserting a guarantee the code does not provide. * feat(agent-core): tighten Bash/Grep/Write/Edit tool descriptions - Bash: prefer the cwd argument (or absolute paths) over a cd from an earlier call, since each call runs in a fresh shell - Grep: note that files_with_matches is ordered most-recently-modified first - Write: do not create documentation/README files unless the user asks - Edit: frame replace_all with its rename-across-file use-case Each change is covered by a description assertion. * feat(agent-core): refine plan-mode/todo/cron tool descriptions - ExitPlanMode: describe what a good plan contains (specific, verifiable steps grounded in the codebase, not vague filler) - TodoList: stop calling it useful 'in Plan mode' — plan-mode planning goes to the plan file; TodoList tracks execution progress - CronCreate: warn that a one-shot whose pinned day/month already passed this year is rejected; document the 50-task session cap and the 8 KiB prompt cap - CronCreate: drop the bench-only KIMI_CRON_NO_STALE / KIMI_CRON_NO_JITTER env knobs from the model-facing description (CI-only; the model never sets them) Each change is covered by a description assertion. * refactor(agent-core): dedupe ExitPlanMode options docs into the param schema; trim EnterPlanMode workflow - ExitPlanMode: the options field mechanics (label format, recommended, count, single-option=plain-approval, reserved labels) now live only in the options param describe; the tool description routes to it and keeps the yolo/manual UI behavior it uniquely documents. The options consistency test now enforces a single source of truth (describe) plus the schema-consistency guard, instead of requiring the same facts in both surfaces. - EnterPlanMode: trim the duplicated 'What Happens in Plan Mode' steps to a pointer (the full workflow is injected unconditionally once plan mode is active), keeping the explore-subagent recommendation. * fix(agent-core): correct prompt/code inaccuracies found in the final audit Every item below was re-verified against the live code: - Skill: drop the never-fired recursion-depth cap (production never seeds depth); keep the <kimi-skill-loaded> 'already loaded, don't re-invoke' guard - TaskOutput: terminal_reason can also be `failed`, not just timed_out/stopped - Grep: count_matches emits per-file `path:count`, with the total reported separately - Plan mode: the reminder names TaskStop/CronCreate/CronDelete as blocked (they are hard-denied by plan-mode-guard-deny) - Bash: the failure trailer is non-zero-exit-specific; timeout/interrupt differ - CreateGoal: replace also covers a blocked goal, not just active/paused - UpdateGoal: it also injects the completion/blocked outcome prompt, so it does more than 'only record the status' - FetchURL: state the universal http/https contract instead of provider-internal SSRF and 10 MiB limits (the primary Moonshot fetcher enforces neither) - TodoList: query mode triggers on omitting `todos`, not on zero args - TaskList: command/PID/exit code are shell-task fields only - CronCreate: the returned fields include `cron` - SetGoalBudget: turn/token budgets are rounded up to >= 1, not rejected below 1 Each change is covered by a description/param assertion; plan.test.ts snapshots refreshed for the longer plan-mode reminder. * fix(agent-core): gate prompt tool guidance on runtime availability, not declared profile tools The HAS_* / Skills gating computed flags from the profile's declared tools, but Agent/AgentSwarm only register when a subagentHost exists (ToolManager .initializeBuiltinTools). A runtime built without a subagentHost (e.g. direct SDK construction) therefore rendered the explore-delegation guidance for an Agent tool the model could not call. SystemPromptContext now carries an optional availableTools; buildTemplateVars gates on it when present and falls back to the declared tools otherwise. useProfile passes the profile tools minus Agent/AgentSwarm when no subagentHost is wired, so the render reflects what the model can actually call. The normal session path (subagentHost always defaulted) is unchanged. * fix(agent-core): exempt the plan-mode plan file from the Write *.md ban Plan mode writes its plan to plans/<id>.md (plan/index.ts) and the reminder tells the model to create it with Write when missing, which contradicted Write's blanket 'do not create *.md unless asked' guard. Carve the plan file out of the ban. * fix(agent-core): scope plan-mode prompt guidance and the Write *.md ban to runtime reality - Gate the TodoList bullet's "enter plan mode via EnterPlanMode" suggestion on a new HAS_ENTERPLANMODE flag. A custom profile that keeps TodoList but drops EnterPlanMode no longer steers the model toward a tool it cannot call; the default profile render is unchanged. - Reframe the Write *.md prohibition around intent (unsolicited docs) instead of a blanket extension ban, so artifacts a task or project instruction requires — the plan-mode plan file, a repo-mandated changeset — are no longer contradicted by the tool's own rules. * refactor(agent-core): move tool-coupled guidance into tool descriptions The default system prompt carried tool-usage guidance behind {% if HAS_* %} gates that re-derived, in prose, the availability the tool schema already encodes — and the same guidance was duplicated in each tool's own description. Drop the four gated blocks (background Bash, Agent/explore, TodoList, EnterPlanMode) and the compaction TaskList/TodoList bullets; the tool descriptions, shipped only when the tool is registered, already carry the same instructions, so subagents and tool-trimmed profiles are no longer pointed at tools they lack. Fold the two genuinely unique lines into the tool descriptions: bash.md gains "return control after starting a background task", agent.md gains the context-hygiene reason to delegate. Collapse the compaction bullets into one tool-agnostic sentence. Remove the now-unused availableTools / HAS_* render machinery. * fix(skill-tool): clarify no-reinvoke guard and argument handling in tool description * feat(fetch-url): indicate content retrieval mode in output for better model context * fix(agent-core): correct goal-budget rounding and task-output failure docs set-goal-budget.md said turn/token budgets are "rounded up", but the code uses Math.round — say "rounded to the nearest whole number" instead. task-output.md implied every failed task carries terminal_reason/stop_reason, but a plain non-zero command exit carries only status plus exit_code; describe that exit_code path and reserve terminal_reason for non-exit endings (timeout, explicit stop, or an internal error with no exit code). * fix(agent-core): scope free-work guidance by role and steer one-shots near-term The blast-radius paragraph told every profile that local work — including editing files — may be done freely, but the read-only explore/plan subagents render it too; scope it to "work your role permits" so it no longer undercuts their read-only constraints. The one-shot cron guidance leaned on a year-boundary heuristic ("avoid a day/month already passed this year") that misfires across Dec 31 to Jan 1 and duplicated a limit the code already enforces. Replace it with a plain near-term nudge and leave the hard future-window guard in code. * fix(enter-plan-mode): clarify availability of Agent tool in plan mode description * fix(agent-core): surface Grep count_matches total and pagination in output count_matches put the aggregate "Found N occurrences" summary and the "Results truncated... use offset=N to see more" notice on the result's message field, which normalizeToolResult drops before the result reaches the model. The model saw only the path:count lines and could miss the total and, worse, the pagination cue — so it would not know to page through truncated counts. Append both to output after the path:count lines, the same way the content and files_with_matches modes already inline their notices. * fix(grep): reorder count summary and results in output for clarity * chore(changeset): consolidate prompt-hardening changesets into one Squash the five per-change changesets for this PR into a single concise entry; they all bump @moonshot-ai/kimi-code (patch) for the same system-prompt and tool-description hardening work. * fix(agent-core): stop the agent from blocking on background tasks Both the Agent and Bash background-launch messages invited the model to "peek at progress" via TaskOutput, and the foreground-vs-background guidance had been thinned to a single parameter hint. Together that led the model to launch a background subagent and then immediately wait on it through TaskOutput — defeating the point of background execution. Make both launch messages take the same anti-wait stance the user-detach path already uses (do NOT wait, poll, or call TaskOutput on it), restore foreground-by-default guidance in the Agent background description (run in the background only when you have other work and do not need the result to proceed), and add a TaskOutput backstop against using it to sit and wait. Also fold the fix into the consolidated changeset. (cherry picked from commit 9c97161)
sailist
added a commit
that referenced
this pull request
Jul 12, 2026
…l flag (#1441) * fix: adapt grep tool to agent-core-v2 * fix(agent-core-v2): enrich PATH from the user's login shell at startup - port probeLoginShellPath/mergeLoginShellPath/applyLoginShellPath into _base/execEnv/loginShellPath.ts as a pure helper (no DI) - export execFileText from environmentProbe for reuse by the probe - run applyLoginShellPathFromNode concurrently with the host probe in HostEnvironmentService, mirroring kaos LocalKaos.create() Aligns agent-core-v2 with kaos 021786f5 so the Bash tool finds user-installed tools (e.g. Homebrew's gh) when kimi-code is launched from a GUI or non-login shell. * fix(agent-core-v2): prefer persisted cwd on resume * feat(agent-core-v2): support structured response formats * fix: restore v2 grep telemetry and tests * fix: preserve v2 compaction boundary * fix(agent-core-v2): align agent and swarm tool behavior * feat(ws-v1): add per-agent event subscription filter - protocol: add optional agent_filter to client_hello and subscribe - kap-server: carry per-subscription agent allowlists through the broadcaster and connection, narrowing live fan-out and replay to selected agents while keeping a single global sequence and bypassing the filter for global events - agent-core-v2: degrade MiniDbQueryStore to a no-op read model when the query-store lock is held by another process instead of crashing the host * feat(web): prefix skill slash commands with skill: to distinguish them from built-in commands (#1492) * ci: release packages (#1468) Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> * docs(changelog): sync 0.23.2 from apps/kimi-code/CHANGELOG.md (#1496) * chore: add changeset for agent swarm parity * fix: align v2 compaction prompt * fix: recover v2 compaction from plain 413 * fix: report v2 compaction retry telemetry * fix: align MCP discovery and output with v1 * fix: align v2 compaction auth guards * fix: align v2 grep behavior with v1 * chore: remove agent swarm changeset * fix(agent-core-v2): restore task resume parity * feat(agent-core-v2): record llm request traces * fix(agent-core-v2): align AskUserQuestion tool chain with v1 - translate wire ids back to question text / option labels when resolving a question over REST, joining multi-select labels with ', ' - enforce unique question texts / option labels and non-empty strings at both the schema and the execution path - cancel pending questions on turn abort or background task stop by dismissing the parked entry (resolves null, v1 broker semantics) - restore the unsupported-client fallback and dismissed-error handling - pass empty header / option description through verbatim and align the model-facing tool description byte-for-byte with v1 - drop the synthetic expires_at field from the question wire shape * fix(agent-core-v2): preserve compaction hook session * test(agent-core-v2): cover concurrent agent background limit * fix(agent-core): report EXIF-rotated image dimensions and raise edge cap to 3000px (#1460) * fix(agent-core): report EXIF-rotated image dimensions and raise edge cap to 3000px Image compression now reports original dimensions in the decoded (EXIF-rotated) space, matching the coordinate system of the sent image and of ReadMediaFile region readback; previously portrait JPEGs (orientation 5-8) got swapped width/height in captions. The longest-edge downscale cap rises from 2000px to 3000px, and the default jimp resize path is documented as the anti-aliased area-average one so it is not accidentally switched to a point-sampled interpolation mode. * test: shrink oversized image fixtures to fit CI timeouts The 3600x3600 fixtures introduced for the 3000px edge cap nearly doubled the pixel area jimp has to decode and deflate, pushing the slowest compression tests past the 5s vitest timeout on CI runners. 3600x1800 keeps every fixture over the cap while restoring roughly the workload of the old 2600x2600 fixtures that CI handled comfortably. * test: pin anti-aliased downscale quality with executable guards A 1px checkerboard probe pins the compressor to full-coverage averaging at integer and fractional ratios, with jimp's point-sampled BILINEAR mode kept as the executable aliasing counter-example (it collapses the 50%-gray pattern to solid black at 4:1). Also guards the other classic downscale bugs: transparent-pixel color bleed, mean-brightness drift, iterative recompression degradation, and zero-size collapse on extreme aspect ratios. * fix(agent-core): report decoded EXIF-rotated dimensions in ReadMediaFile notes The media note derived its original-dimensions line from the header sniff, which reports pre-rotation values for EXIF orientation 5-8 JPEGs. The sent image and region readback both live in the decoded (rotated) space, so portrait photos got axis-swapped coordinate guidance. Once a decode has happened — compression or crop — its dimensions now overwrite the sniffed ones. * fix(agent-core): improve handling of EXIF orientation in image dimensions and metadata * fix(agent-core): sniff EXIF orientation and step budget fallback through 2000px Two follow-ups to the EXIF and 3000px-cap changes: sniffImageDimensions now reads the JPEG EXIF Orientation tag (pure header parse, both byte orders) and reports display-space dimensions for orientations 5-8. Passthrough images — never decoded — previously kept the pre-rotation header size in compression results and media read notes, disagreeing with the decoded space that region readback uses. encodeWithinBudget steps the over-budget fallback through 2000px before the 1000px last resort. Raising the cap to 3000px had left a regression window: an image whose 2000px encode fits the byte budget was sent at 1000px where the old 2000px cap used to send it at 2000px. * fix(kimi-code): record pasted image dimensions in display space The TUI paste path recorded attachment and original dimensions from its raw header parser, which ignores EXIF orientation. For a portrait JPEG the submit-time caption then contradicted the sent image's aspect and region readback coordinates were axis-swapped. Dimensions now come from the compression result, which reports display space on both the compressed and passthrough paths; parseImageMeta remains only the format/mime gate. * feat(agent-core): add image compression and crop telemetry Every image ingestion path now reports an image_compress event — outcome (compressed / passthrough fast, guard, unsupported, unhelpful, error), input/output formats, byte and pixel sizes, EXIF transposition, and duration — and region readback reports an image_crop event with a failure classification and the region's share of the original area. Wiring is per call site via a new CompressImageOptions.telemetry option, so the outcome split and timing are measured inside the compressor while each caller only names its source: ReadMediaFile (tool construction, like GrepTool), MCP tool results (McpOutputOptions), server prompt ingestion (ICoreProcessService now exposes the host telemetry client), ACP prompts (session track adapter), and TUI paste (host.track adapter). Properties are numeric/enum only — never paths or content — and a throwing client can never affect the compression result. * fix(agent-core): run the full JPEG quality ladder at fallback sizes The fallback rescales encoded only at quality 20, so a JPEG whose ladder failed at the fitted size collapsed straight to the lowest quality even when the smaller size left budget headroom for a higher rung (the realistic window is the 1000px step, where the 4x pixel drop pays for q80/q60). Each fallback edge now walks the same q80-to-q20 ladder as the fitted size. * test: shrink heavy JPEG fixtures and add explicit timeouts The fallback-ladder test runs ~11 pure-JS JPEG encodes and the EXIF paste test decodes, rotates, and re-encodes a 6.5MP frame; both sat at the edge of the 5s vitest timeout on CI runners. Narrower fixtures cut the pixel area (the ladder test keeps its width above 2000px so the full fallback chain still runs) and explicit 15s timeouts absorb runner variance. * fix(server): scope prompt image compression telemetry to the session The prompt-ingestion image_compress events were emitted with the bare host telemetry client, while every agent-side source inherits a session-scoped client — so prompt_inline/prompt_file events could not be correlated with their session. The route now wraps the client with withTelemetryContext({ sessionId }) like rpc/core-impl does for session telemetry. * chore(changeset): consolidate image compression changesets One entry covering the cap raise and the EXIF dimension fix, listed for both the CLI and the SDK so the SDK changelog's compression description (previously pinned at 2000px) stays accurate. * fix: count goal creation turn (#1477) * feat(kosong): support structured response formats (#1397) * fix: clarify goal blocked audit guidance (#1481) * feat(agent-core): discard loaded tool schemas on compaction (#1471) Align progressive tool disclosure with the discard-on-compaction model: compaction no longer rebuilds loaded dynamic tool schemas. The boundary announcement re-lists every loadable name, the model re-selects what it still needs, and a from-memory call to a no-longer-loaded tool is rejected by preflight with select guidance. This removes the keep-all rebuild and its half-trigger budget heuristics entirely: the post-compaction floor is back to users + summary, which is structurally outside the auto-compaction trigger band, and the guard baseline degenerates to summary + reinjected reminders. Every downstream mechanism already treated the empty loaded set as its consistent base state (ledger scan, pending clear at the compaction boundary, deferred extras, preflight wording), so this is a strict simplification. Co-authored-by: fengchenchen <fengchenchen@moonshot.ai> * fix(kimi-code): exit 1 when a headless (-p) turn fails (#1483) Headless (`kimi -p`) failures could exit with code 0 when the event loop drained during the shutdown cleanup (e.g. telemetry's unref'd retry backoff when the network is blocked), because the rejection never reached the process.exit(1) call. Set the failure exit code before any await in both the run-prompt catch and the main catch, and keep the cleanup timeout ref'd so the loop stays alive long enough for the rejection to propagate. * feat(plugins): add Vercel plugin to marketplace (#1489) * feat(web): support Enter key to confirm archive and other dialogs (#1490) * feat(web): redesign cron reminder as a message bubble (#1480) * feat(web): redesign cron reminder as a message bubble Restyle the cron trigger notice as a right-aligned user-style message bubble that shows the scheduled prompt in full (wrapping across lines), with a small meta row beneath it for the schedule, status, job id and run time. Extract a shared MessageTime component used by both user messages and the cron reminder so the timestamp format and click-to-expand behavior stay consistent, and give the CronCreate/CronList/CronDelete tools distinct calendar icons. * refactor(web): render cron reminders only as standalone turns Remove the embedded cron block path from the web transcript projector so cron reminder fires always render through the standalone right-aligned bubble path. * chore(web): simplify cron redesign changeset * fix(web): composer model switch also updates global default model (#1491) * fix(web): composer model switch also updates global default model The composer model switcher still switches the active session's model via POST /sessions/{id}/profile (awaited, so the model pill reflects the result), and additionally fires POST /api/v1/config with { default_model } as a fire-and-forget side effect so new sessions inherit the chosen default. The config request is skipped when the model already matches the current default. * fix(web): route ModelPicker overlay selection through the default-model update The overlay opened from the composer's "More models" row (and /model) is a continuation of the same switch flow, so its selection now also bumps the global default model instead of only switching the active session. * fix(web): only persist the default model after a confirmed session switch setModel now returns whether the switch was accepted (true for the draft path), so the composer flow no longer writes a stale or invalid model alias into the global config when the session-level switch failed and rolled back. * feat(web): prefix skill slash commands with skill: to distinguish them from built-in commands (#1492) * ci: release packages (#1468) Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> * fix(agent-core-v2): restore native append for Write append mode - add IHostFileSystem.appendText backed by fs.appendFile (O_APPEND) - route WriteTool append through it instead of read-then-rewrite, so existing content is never read, truncated, or clobbered by concurrent writers and a crash mid-append can only lose the new bytes - update typed host-fs test fakes and WriteTool append assertions * fix(agent-core-v2): align task tool prompts with v1 * fix(agent-core-v2): align compaction empty retry * fix(agent-core-v2): restore web search source site and citation reminders - surface source site: add WebSearchResult.siteName, map site_name in the Moonshot provider, and render the Site: line in tool output - restore the per-search inline citation reminder alongside the results - align web-search.md with v1: source-site/result-summary guidance and the static citation reminder * fix(agent-core-v2): route task timeouts through SIGTERM grace + SIGKILL - add terminateWithGrace shared by stop, timeoutMs, detachTimeoutMs, and track deadlines: cancel/SIGTERM -> 5s grace -> forceStop (SIGKILL) - coerce a post-abort self-settled `killed` to `timed_out` so a deadline stays reported as timed_out, matching v1 settlementForOutcome - add manager tests for SIGTERM-ignored escalation, graceful-exit within the grace window, and detachTimeout teardown * fix(agent-core-v2): restore fs.grep streaming early-kill and symlink reporting - stream `rg --json` in fs.grep and SIGKILL once max_total_matches/max_files is reached, restoring v1 early-stop instead of buffering the whole output - report symlinks as kind 'symlink' in fs search/list/stat via lstat and never descend into symlinked directories - remove the unused os grepSearch helper (dead, non-streaming, bypassed ISessionProcessRunner) * test(agent-core-v2): align truncated compaction retry * fix(agent-core-v2): align MCP tool results with v1 * fix(agent-core-v2): align blocked compaction failures * fix(agent-core-v2): align manual compaction tool projection * fix(agent-core-v2): preserve tail in windowed compaction * fix(agent-core-v2): read todos from wire model, sanitize replay - make SessionTodoService a stateless facade over the main agent's TodoModel: getTodos reads wire.getModel(TodoModel) live, setTodos only dispatches a todo.set op, and onDidChange is bridged from wire.subscribe(TodoModel); the in-memory list copy is gone so the live and post-replay views cannot drift - sanitize todo.set payloads in apply via readTodoItems, so replayed or hand-written records cannot poison the model or downstream renders - update todo tests to a sanitizing/notifying/replaying wire stub and cover malformed todo.set replay and main-absent reads - record the main-agent-wire persistence debt for the ISessionWireService move * fix(agent-core-v2): port v1 Bash tool output cap, saved-output reference, and background gating (#1503) * fix: align v2 task observable behavior * fix: v2 full compaction * fix(agent-core-v2): align task wait timeout behavior * chore: webSearch & FetchUrl Sync #1260 * fix: align background agent guidance * fix(agent-core-v2): align full compaction with v1 * fix(agent-core-v2): align v1 wire records * fix(agent-core-v2): remember observed compaction context window * fix: align Agent / AgentSwarm * fix(agent-core-v2): port v1 parity fixes for hooks, anthropic, thinking config and add-dir (#1504) * fix(agent-core-v2): hide console window when running hooks on Windows Port the v1 hooks runner fix: extract buildHookSpawnOptions and pass windowsHide:true so hook child processes no longer flash a console window on Windows, mirroring the node-local process host defaults. Includes the same regression tests as v1. * fix(agent-core-v2): port anthropic max_tokens ceiling and override fixes Port two v1 kosong fixes to the v2 anthropic provider: - Fall back to the nearest lower catalogued minor when resolving the Claude output ceiling, and catalogue Opus 4.8's documented 128k cap, so an uncatalogued minor no longer drops to the family baseline. - Treat an explicit defaultMaxTokens as the final max_tokens value instead of clamping it to the built-in ceiling. Mirrors the v1 regression tests in a new anthropic-max-tokens test file. * refactor(agent-core-v2): converge thinking config to enabled/effort Port the v1 thinking-config overhaul (#1132's config side) to v2: - ThinkingConfigSchema becomes { enabled, effort, keep }; the mode enum, the separate defaultThinking section, and the KIMI_MODEL_THINKING_MODE / KIMI_MODEL_DEFAULT_THINKING env bindings are removed. - The effort resolver drops the mode/defaultThinking branches and no longer normalizes a requested 'on' to a concrete effort in core; 'on' is taken verbatim and normalization stays at the UI boundary. - OAuth login/refresh and catalog refresh now persist the thinking.enabled value computed by the shared oauth apply/restore logic instead of dropping it and writing the removed default_thinking key, so [thinking] enabled = false actually disables thinking and the login default survives on disk. Mirrors the v1 resolver regression tests and adds a persistence regression for the refresh path. * docs(agent-core-v2): fix stale loop-event comments after wire parity The v1.4 wire-parity alignment switched the v2 live loop to stream turns as context.append_loop_event records, but three comments still described the old world (restore-only Op, "v2 never emits loop events"). Update them to match the actual write path: non-loop appends use append_message, the loop persists loop events byte-compatible with v1, and the fold runs both at live dispatch time and on replay. * fix(agent-core-v2): load workspace additional dirs on session create and resume The /add-dir command persisted remembered dirs to .kimi-code/local.toml, but session materialization never read them back and offered no caller additionalDirs entry point — a remembered dir silently stopped applying to new, resumed, and forked sessions. Mirror v1's createSession/resumeSession: merge the project-local local.toml dirs with caller-supplied additionalDirs (relative paths resolve against workDir) and seed the session workspace context in materializeSession, so create/resume/fork all pick them up. A broken local.toml fails the create loudly with CONFIG_INVALID, same as v1. Tests mirror v1's runtime coverage for the load/merge/dedupe/resume/fork scenarios. * fix(kimi-code): forward create-session additional dirs from the v2 harness The in-process v2 print-mode harness dropped the SDK CreateSessionOptions additionalDirs when calling ISessionLifecycleService.create, so --add-dir never reached the v2 resolver. Pass it through. * fix(agent-core-v2): align full compaction observability * fix(agent-core-v2): remove compact hook trigger state * feat(agent-core-v2): enhance agent lifecycle with context size tracking and concurrency checks * fix(agent-core-v2): align media reads with v1 note channel and EXIF handling (#1505) * fix(agent-core-v2): align media reads with v1 note channel and EXIF handling Port two agent-core changes into agent-core-v2: - Move the ReadMediaFile media summary from an inline <system> text part onto the tool result's note side channel, so raw <system> markup never renders in UIs (matching the MCP output path). - Report image dimensions in the decoded EXIF-rotated space: the header sniff now reads the JPEG Orientation tag, and once a decode happened (compression or crop) its dimensions overwrite the sniffed ones, so portrait photos no longer get axis-swapped coordinate guidance. - Raise the longest-edge downscale cap from 2000px to 3000px, step the over-budget fallback through 2000px before the 1000px last resort, and run the full JPEG quality ladder at fallback sizes. - Report image_compress / image_crop telemetry for media reads (source read_media), with EXIF transposition and crop failure classification. The tool description also regains the downsampling recovery guidance (region / full_resolution readback) that the v2 copy predated. * fix(agent-core-v2): align v1 wire records * fix(agent-core-v2): hide compression captions and register media tools in production Port the remaining v1 media gaps into agent-core-v2: - Reroute inline image-compression captions out of user messages: the prompt service splits them at the append chokepoint (prompt and steer flush) and delivers them through the built-in system-reminder injection (origin {kind: 'injection', variant: 'image_compression'}), which every UI hides. Session titles/lastPrompt strip the caption the same way. The model still receives the full note. - Register ReadMediaFile in production: media tools cannot use the module-level contribution table (capabilities are unknown until a model binds), so a new Eager agent-scope registrar re-runs registerMediaTools on every agent.status.updated where the model alias or its media capabilities changed, rebinding the video uploader and dropping the tool when the model loses media input. * fix(agent-core-v2): port v1 parity fixes for hooks, anthropic, thinking config and add-dir (#1504) * fix(agent-core-v2): hide console window when running hooks on Windows Port the v1 hooks runner fix: extract buildHookSpawnOptions and pass windowsHide:true so hook child processes no longer flash a console window on Windows, mirroring the node-local process host defaults. Includes the same regression tests as v1. * fix(agent-core-v2): port anthropic max_tokens ceiling and override fixes Port two v1 kosong fixes to the v2 anthropic provider: - Fall back to the nearest lower catalogued minor when resolving the Claude output ceiling, and catalogue Opus 4.8's documented 128k cap, so an uncatalogued minor no longer drops to the family baseline. - Treat an explicit defaultMaxTokens as the final max_tokens value instead of clamping it to the built-in ceiling. Mirrors the v1 regression tests in a new anthropic-max-tokens test file. * refactor(agent-core-v2): converge thinking config to enabled/effort Port the v1 thinking-config overhaul (#1132's config side) to v2: - ThinkingConfigSchema becomes { enabled, effort, keep }; the mode enum, the separate defaultThinking section, and the KIMI_MODEL_THINKING_MODE / KIMI_MODEL_DEFAULT_THINKING env bindings are removed. - The effort resolver drops the mode/defaultThinking branches and no longer normalizes a requested 'on' to a concrete effort in core; 'on' is taken verbatim and normalization stays at the UI boundary. - OAuth login/refresh and catalog refresh now persist the thinking.enabled value computed by the shared oauth apply/restore logic instead of dropping it and writing the removed default_thinking key, so [thinking] enabled = false actually disables thinking and the login default survives on disk. Mirrors the v1 resolver regression tests and adds a persistence regression for the refresh path. * docs(agent-core-v2): fix stale loop-event comments after wire parity The v1.4 wire-parity alignment switched the v2 live loop to stream turns as context.append_loop_event records, but three comments still described the old world (restore-only Op, "v2 never emits loop events"). Update them to match the actual write path: non-loop appends use append_message, the loop persists loop events byte-compatible with v1, and the fold runs both at live dispatch time and on replay. * fix(agent-core-v2): load workspace additional dirs on session create and resume The /add-dir command persisted remembered dirs to .kimi-code/local.toml, but session materialization never read them back and offered no caller additionalDirs entry point — a remembered dir silently stopped applying to new, resumed, and forked sessions. Mirror v1's createSession/resumeSession: merge the project-local local.toml dirs with caller-supplied additionalDirs (relative paths resolve against workDir) and seed the session workspace context in materializeSession, so create/resume/fork all pick them up. A broken local.toml fails the create loudly with CONFIG_INVALID, same as v1. Tests mirror v1's runtime coverage for the load/merge/dedupe/resume/fork scenarios. * fix(kimi-code): forward create-session additional dirs from the v2 harness The in-process v2 print-mode harness dropped the SDK CreateSessionOptions additionalDirs when calling ISessionLifecycleService.create, so --add-dir never reached the v2 resolver. Pass it through. * fix(agent-core-v2): report video_upload telemetry for media reads Port the v1 video-upload telemetry wrapper into createVideoUploader: every upload emits a video_upload event with outcome (success/error), byte size, mime type, duration, and the caller's static props (model alias, protocol tags), and a throwing telemetry client never affects the upload outcome. The media-tools registrar supplies the sink and props from the bound model. Also restores two v1 rationale comments in ReadMediaFile (original-size reporting and the full_resolution hard refusal) that were dropped during the earlier port. --------- Co-authored-by: 7Sageer <7sageer@djwcb.cn> Co-authored-by: liruifengv <liruifeng1024@gmail.com> * refactor(agent-core-v2): remove the microCompaction domain - delete the microCompaction domain (service, wire model/op, config section, experimental flag) and its dedicated tests - stop truncating old tool results in the context projector and drop the projector's now-unused instantiation dependency - remove the domain from the layer map, package exports, and the DI x Scope dependency diagram - retarget the flag-registry test and skill examples at a neutral flag * fix(contextProjector): surface projection repairs via log warning - add ProjectionAnomaly + onAnomaly sink through the project / projectStrict passes (reorder, synthesize, orphan / duplicate drop, leading drop, merge, blank-text drop) so the pure projection reports every wire-repair it applies - AgentContextProjectorService injects ILogService and emits a single signature-deduped 'repaired the request to keep it wire-valid' warning, excluding trailing-tail synthesis, matching agent-core parity - cover the trace and its dedup in the projector tests * fix(agent-core-v2): fix cron killswitch, lost deliveries, id clashes - killswitch: read KIMI_DISABLE_CRON live by re-applying the ConfigService env overlay on every get(); CronCreate reads it via ISessionCronService instead of a value frozen at tool registration - delivery: resolve fire delivery on promptService.steer().launched so a rejected launch retains one-shot tasks for retry instead of deleting them; tick() is now async and awaits delivery before advancing cursors - ids: switch cron task ids to ULIDs (from 32-bit hex) so two sessions sharing a workspace cannot overwrite each other's persisted task; CronDelete and persistence accept both ULID and legacy 8-hex ids - display: CronCreate reports nextFireAt through the service so it honors KIMI_CRON_NO_JITTER and matches the scheduler and CronList - migration: adopt shape-valid tasks with no sessionId tag on loadFromStore and stamp the tag back to disk - persistence: create cron directories 0700 and files 0600 via FileStorageService dirMode/fileMode Gate SessionCronService startup on config.ready and resolve clocks after ready so config is never read before it is loaded; start() is now async. * feat(fs-watch): add workspace fs watch with v1-compatible WS delivery - os layer: add IHostFsWatchService over chokidar (raw create/modify/delete, .git ignored) - session layer: add ISessionFsWatchService, a workspace-confined, debounced, .gitignore-aware FsChangeEvent feed - kap-server: add FsWatchBridge pushing event.fs.changed over /api/v1/ws (watch_fs_add/remove, volatile, per-connection filter), byte-compatible with v1 - tests: os/session unit tests and kap-server fs-watch e2e * refactor(agent-core-v2): run external hooks through IHostProcessService - inject IHostProcessService into ExternalHooksRunnerService and thread it through runMatchedHooks to runHook instead of spawning node:child_process - route hook termination through the service's cross-platform process-tree kill - settle on the exit code plus drained stdout/stderr so fast-exiting hooks keep their trailing output - hide the child console window on Windows via the service default - update externalHooks tests for the new dependency * fix(agent-core-v2): strict-decode Edit reads, align with v1 - read the Edit target with errors:'strict' so a non-UTF-8 file fails the edit instead of being silently rewritten as U+FFFD (matches v1 kaos) - declare readWriteFile access since Edit reads before it writes, matching v1 - render edit.md directly instead of through renderPrompt: it has no template vars, and raw avoids treating literal {{ }} as a template - restore the replace_all usage example in edit.md (v1 #1102) - add a regression test asserting a non-UTF-8 file fails the edit and keeps its bytes untouched * fix(agent-core-v2): dedupe AgentMeta legacy field declarations * refactor(agent-core-v2): persist wire records natively in the v1 vocabulary Remove the persist-time v1 rewrite layer (serializeV1WireRecord): ops now write v1-shaped records directly, live-only state is declared persist:false on the op instead of being stripped at write time, and the swarm-exit reminder pop replays from the swarm_mode.exit record via a cross-model reducer. Fixes resumed sessions losing the todo list, drifting turn counters after retries, and removed reminders reappearing on resume. * refactor(agent-core-v2): move ReadTool status block to note side channel - ReadTool.finishReadResult now returns rendered lines as `output` only and rides the `<system>` status block on the model-only `note` side channel - drop the finishOutput helper that concatenated content and status - update read.test.ts expectations to assert `note` separately from `output` * refactor(agent-core-v2): split blob service helpers, rewrite tests - extract rewriteMediaUrls and blobref parse/format helpers to dedupe URL rewriting - move the byte-bounded LRU cache into a module-private ByteLruCache with focused unit tests, dropping the protected maxCacheSize test seam - rewrite blob service tests against the contract on in-memory storage, removing cache-internals cases * fix: make release-e2e scenarios pass under agent-core-v2 Three independent fixes for release-e2e failures that only appeared with the experimental v2 engine (KIMI_CODE_EXPERIMENTAL_FLAG): - agent-core-v2: register the KIMI_MODEL env overlay statically so it takes effect even when ModelService is not instantiated (the DI layer does not auto-instantiate Eager services). Fixes wire-llm-request-trace. - cli: omit the leading system.version meta line in stream-json prompt mode so the role sequence stays clean. Fixes stream-json-cron. - agent-core-v2: honor --skills-dir via a new explicit skill source seeded from the host. Fixes interactive-skills-dir. Cherry-picked from 2a7232737 (v2-migration), excluding the node-sdk V2Host change (not applicable on this branch). * refactor(agent-core-v2): drop replay-only wire ops Remove the three replay-only Ops that were kept for pre-alignment / 1.5 sessions, now that v2 persists natively in the v1 vocabulary: - turn.launch (replaced by turn.prompt) - todo.set (replaced by tools.update_store with key 'todo') - context.splice (replaced by context.append_message / append_loop_event) Also drop the dead code that handled them (transcript reducer, task-origin extraction, blob dehydration, harness helpers) and migrate the affected tests to the v1 record types. The live write path already emitted only v1 records, so wire.jsonl output is unchanged. * Revert "fix: make release-e2e scenarios pass under agent-core-v2" This reverts commit ec9dae72abbbd522722cfb0f4191d84602008643. * fix(agent-core-v2): use a fresh TextDecoder per append-log read The module-level TextDecoder is stateful in stream mode: it buffers a trailing incomplete multi-byte sequence until the next decode. Sharing it across reads let leftover state from an earlier read that returned early (e.g. ensureWireMetadata bailing on the leading metadata record) leak into the next read and prepend a U+FFFD to its first line, corrupting the metadata envelope and breaking session fork with "corrupted line 1". Give each read its own TextDecoder so decoder state never leaks between reads. * fix(agent-core-v2): register KIMI_MODEL env overlay statically The KIMI_MODEL_* effective overlay was registered by ModelService on construction, but the DI layer does not auto-instantiate Eager services, so the overlay never took effect when nothing resolved IModelService. This broke the release-e2e wire-llm-request-trace scenario, where KIMI_MODEL_NAME must synthesize the env model and its thinking capability. Move registration to module load via a new configOverlayContributions collector, drained by ConfigRegistry on construction — mirroring the existing configSectionContributions pattern. ModelService no longer depends on IConfigRegistry. * docs(agent-core-v2): clarify live-only op semantics * fix(agent-core-v2): preserve oversized tool results * chore: remove full compaction complete data type * fix(agent-core-v2): align foreground output cap * fix(agent-core-v2): gate skill prompt injection * chore(skills): bundle review and test lenses into kc-review - add agent-core-review umbrella skill with slop and test sub-skills - move write-tests rules into agent-core-review/test and drop the standalone skill * feat(v2): auto-mint session ids and harden print-mode background drain - make CreateSessionOptions.sessionId optional; SessionLifecycleService.create and fork now mint `session_<lowercase-uuid>` via a shared createSessionId helper, so edge layers stop minting their own ids (drop randomUUID in the v2 harness, ulid in kap-server) - rework V2Session.waitForBackgroundTasksOnPrint to re-enumerate each round, suppress terminal notifications while waiting, and bound the drain by [task].print_wait_ceiling_s (default 1h) instead of a hardcoded 30s cap, so kimi -p can run long tasks to completion without being steered into a new turn - add v2-session unit tests; seed session/agent/bootstrap context in the tool-dedupe harness for the real executor * fix(agent-core-v2): refresh system prompt after compaction * docs(agent-core-review): limit kc-review skill to agent-core-v2 Clarify that the kc-review lenses apply only to packages/agent-core-v2 (the DI x Scope engine), not to the legacy packages/agent-core or other packages. * fix(agent-core-v2): align model-facing prompts * fix(agent-core): report EXIF-rotated image dimensions and raise edge cap to 3000px (#1460) * fix(agent-core): report EXIF-rotated image dimensions and raise edge cap to 3000px Image compression now reports original dimensions in the decoded (EXIF-rotated) space, matching the coordinate system of the sent image and of ReadMediaFile region readback; previously portrait JPEGs (orientation 5-8) got swapped width/height in captions. The longest-edge downscale cap rises from 2000px to 3000px, and the default jimp resize path is documented as the anti-aliased area-average one so it is not accidentally switched to a point-sampled interpolation mode. * test: shrink oversized image fixtures to fit CI timeouts The 3600x3600 fixtures introduced for the 3000px edge cap nearly doubled the pixel area jimp has to decode and deflate, pushing the slowest compression tests past the 5s vitest timeout on CI runners. 3600x1800 keeps every fixture over the cap while restoring roughly the workload of the old 2600x2600 fixtures that CI handled comfortably. * test: pin anti-aliased downscale quality with executable guards A 1px checkerboard probe pins the compressor to full-coverage averaging at integer and fractional ratios, with jimp's point-sampled BILINEAR mode kept as the executable aliasing counter-example (it collapses the 50%-gray pattern to solid black at 4:1). Also guards the other classic downscale bugs: transparent-pixel color bleed, mean-brightness drift, iterative recompression degradation, and zero-size collapse on extreme aspect ratios. * fix(agent-core): report decoded EXIF-rotated dimensions in ReadMediaFile notes The media note derived its original-dimensions line from the header sniff, which reports pre-rotation values for EXIF orientation 5-8 JPEGs. The sent image and region readback both live in the decoded (rotated) space, so portrait photos got axis-swapped coordinate guidance. Once a decode has happened — compression or crop — its dimensions now overwrite the sniffed ones. * fix(agent-core): improve handling of EXIF orientation in image dimensions and metadata * fix(agent-core): sniff EXIF orientation and step budget fallback through 2000px Two follow-ups to the EXIF and 3000px-cap changes: sniffImageDimensions now reads the JPEG EXIF Orientation tag (pure header parse, both byte orders) and reports display-space dimensions for orientations 5-8. Passthrough images — never decoded — previously kept the pre-rotation header size in compression results and media read notes, disagreeing with the decoded space that region readback uses. encodeWithinBudget steps the over-budget fallback through 2000px before the 1000px last resort. Raising the cap to 3000px had left a regression window: an image whose 2000px encode fits the byte budget was sent at 1000px where the old 2000px cap used to send it at 2000px. * fix(kimi-code): record pasted image dimensions in display space The TUI paste path recorded attachment and original dimensions from its raw header parser, which ignores EXIF orientation. For a portrait JPEG the submit-time caption then contradicted the sent image's aspect and region readback coordinates were axis-swapped. Dimensions now come from the compression result, which reports display space on both the compressed and passthrough paths; parseImageMeta remains only the format/mime gate. * feat(agent-core): add image compression and crop telemetry Every image ingestion path now reports an image_compress event — outcome (compressed / passthrough fast, guard, unsupported, unhelpful, error), input/output formats, byte and pixel sizes, EXIF transposition, and duration — and region readback reports an image_crop event with a failure classification and the region's share of the original area. Wiring is per call site via a new CompressImageOptions.telemetry option, so the outcome split and timing are measured inside the compressor while each caller only names its source: ReadMediaFile (tool construction, like GrepTool), MCP tool results (McpOutputOptions), server prompt ingestion (ICoreProcessService now exposes the host telemetry client), ACP prompts (session track adapter), and TUI paste (host.track adapter). Properties are numeric/enum only — never paths or content — and a throwing client can never affect the compression result. * fix(agent-core): run the full JPEG quality ladder at fallback sizes The fallback rescales encoded only at quality 20, so a JPEG whose ladder failed at the fitted size collapsed straight to the lowest quality even when the smaller size left budget headroom for a higher rung (the realistic window is the 1000px step, where the 4x pixel drop pays for q80/q60). Each fallback edge now walks the same q80-to-q20 ladder as the fitted size. * test: shrink heavy JPEG fixtures and add explicit timeouts The fallback-ladder test runs ~11 pure-JS JPEG encodes and the EXIF paste test decodes, rotates, and re-encodes a 6.5MP frame; both sat at the edge of the 5s vitest timeout on CI runners. Narrower fixtures cut the pixel area (the ladder test keeps its width above 2000px so the full fallback chain still runs) and explicit 15s timeouts absorb runner variance. * fix(server): scope prompt image compression telemetry to the session The prompt-ingestion image_compress events were emitted with the bare host telemetry client, while every agent-side source inherits a session-scoped client — so prompt_inline/prompt_file events could not be correlated with their session. The route now wraps the client with withTelemetryContext({ sessionId }) like rpc/core-impl does for session telemetry. * chore(changeset): consolidate image compression changesets One entry covering the cap raise and the EXIF dimension fix, listed for both the CLI and the SDK so the SDK changelog's compression description (previously pinned at 2000px) stays accurate. * fix: count goal creation turn (#1477) * feat(kosong): support structured response formats (#1397) * fix: clarify goal blocked audit guidance (#1481) * feat(agent-core): discard loaded tool schemas on compaction (#1471) Align progressive tool disclosure with the discard-on-compaction model: compaction no longer rebuilds loaded dynamic tool schemas. The boundary announcement re-lists every loadable name, the model re-selects what it still needs, and a from-memory call to a no-longer-loaded tool is rejected by preflight with select guidance. This removes the keep-all rebuild and its half-trigger budget heuristics entirely: the post-compaction floor is back to users + summary, which is structurally outside the auto-compaction trigger band, and the guard baseline degenerates to summary + reinjected reminders. Every downstream mechanism already treated the empty loaded set as its consistent base state (ledger scan, pending clear at the compaction boundary, deferred extras, preflight wording), so this is a strict simplification. Co-authored-by: fengchenchen <fengchenchen@moonshot.ai> * fix(kimi-code): exit 1 when a headless (-p) turn fails (#1483) Headless (`kimi -p`) failures could exit with code 0 when the event loop drained during the shutdown cleanup (e.g. telemetry's unref'd retry backoff when the network is blocked), because the rejection never reached the process.exit(1) call. Set the failure exit code before any await in both the run-prompt catch and the main catch, and keep the cleanup timeout ref'd so the loop stays alive long enough for the rejection to propagate. * feat(plugins): add Vercel plugin to marketplace (#1489) * feat(web): support Enter key to confirm archive and other dialogs (#1490) * feat(web): redesign cron reminder as a message bubble (#1480) * feat(web): redesign cron reminder as a message bubble Restyle the cron trigger notice as a right-aligned user-style message bubble that shows the scheduled prompt in full (wrapping across lines), with a small meta row beneath it for the schedule, status, job id and run time. Extract a shared MessageTime component used by both user messages and the cron reminder so the timestamp format and click-to-expand behavior stay consistent, and give the CronCreate/CronList/CronDelete tools distinct calendar icons. * refactor(web): render cron reminders only as standalone turns Remove the embedded cron block path from the web transcript projector so cron reminder fires always render through the standalone right-aligned bubble path. * chore(web): simplify cron redesign changeset * fix(web): composer model switch also updates global default model (#1491) * fix(web): composer model switch also updates global default model The composer model switcher still switches the active session's model via POST /sessions/{id}/profile (awaited, so the model pill reflects the result), and additionally fires POST /api/v1/config with { default_model } as a fire-and-forget side effect so new sessions inherit the chosen default. The config request is skipped when the model already matches the current default. * fix(web): route ModelPicker overlay selection through the default-model update The overlay opened from the composer's "More models" row (and /model) is a continuation of the same switch flow, so its selection now also bumps the global default model instead of only switching the active session. * fix(web): only persist the default model after a confirmed session switch setModel now returns whether the switch was accepted (true for the draft path), so the composer flow no longer writes a stale or invalid model alias into the global config when the session-level switch failed and rolled back. * feat(web): prefix skill slash commands with skill: to distinguish them from built-in commands (#1492) * ci: release packages (#1468) Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> * fix: surface provider auth error for unavailable models (#1506) * fix: surface provider auth error for unavailable models When an OAuth-managed model returns 401 after a forced token refresh, the token is valid but the provider rejected it for that model (the account lacks access). Emit provider.auth_error carrying the provider's message instead of auth.login_required with a misleading "OAuth login expired. Send /login" prompt. * fix(agent-core): preserve provider auth errors through compaction Treat provider.auth_error like auth.login_required in the compaction path so an auth rejection during compaction surfaces the provider's message instead of being wrapped as a generic compaction failure. * ci: release packages (#1507) Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> * docs(changelog): sync 0.23.3 and shorten OAuth error entry (#1509) * feat(kimi-web): add status-aware browser notifications (#1479) * feat(kimi-web): add approval notification storage key and i18n copy * feat(kimi-web): add approval notification helpers and tests * feat(kimi-web): wire approval notifications and guard completion alerts * fix(kimi-web): extract shouldNotifyCompletion helper and add tests * feat(kimi-web): add approval notification settings toggle * chore(kimi-web): add changeset and tidy notification module comment - Align approval notification tag with spec (kimi-approval-${approvalId}) - Update module header to describe all three notification kinds * fix(kimi-web): make notifications fire reliably - Key completion notification tags by turn (sid + promptId) and question tags by request id, so a stale notification left in the notification center no longer swallows every follow-up alert in the same session - Suppress notifications only while the window is actually focused, not merely visible (document.hasFocus() on top of visibilityState) - Play the attention sound when a tool needs approval, matching the completion and question sounds * chore(kimi-web): simplify changeset * fix(agent-core-v2): serialize concurrent model catalog refreshes Port v1 #1207's _refreshChain so a scheduled refresh and a manual one (or two overlapping manual ones) never race on reading/patching the persisted config. Applied to both refresh entry points: ModelCatalogService.refreshProviderModels (scheduler + all/single-provider) and OAuthService.refreshOAuthProviderModels (OAuth-only, a separate service in v2). * fix(agent-core-v2): dedupe workspace registry entries by root Port v1 #1221: collapse registered workspaces that share a root in list(), preferring the entry whose id matches the current canonical encodeWorkDirKey, so a legacy workspaces.json (v1-compatible) does not render the same folder twice through GET /workspaces. * fix(agent-core-v2): apply KIMI_CODE_CUSTOM_HEADERS and host identity headers Port v1's provider-manager outbound header logic to agent-core-v2 so `KIMI_CODE_CUSTOM_HEADERS` and host identity headers are applied to outbound LLM requests, closing the migration gap from #1186: - env `KIMI_CODE_CUSTOM_HEADERS` is the lowest-precedence header layer; - host identity headers (User-Agent + X-Msh-*) are sent for Kimi providers, only the User-Agent for every other provider — a Kimi provider routed through the Anthropic protocol still gets the full set, matching v1; - provider `customHeaders` always win on conflict. Host headers are seeded by the CLI via `createKimiDefaultHeaders` and a new `IHostRequestHeaders` App-scope token (defaulting to empty), so the model resolver can layer them without the host threading them through every call site. * chore(agent-core-review): rename skill from kc-review to agent-core-review * feat(kap-server): surface originating stack trace on error envelopes - add optional `stack` field to errEnvelope and the envelope schema/interface; omitted when undefined so the wire shape stays byte-identical for callers without a stack - thread `err.stack` through route error mappers plus the global and transport error handlers - preserve `details` on `session.undo_unavailable` while adding its stack - update tests to assert stacks are surfaced, reversing the prior no-leak contract * fix: align v2 media and task compaction handling * feat(agent-core-v2): sync shell mode and skill config parity (#1514) * feat(agent-core-v2): record shell command context - add ShellCommandOrigin and compaction handoff disposition - extract IAgentShellCommandService from AgentRPCService - keep AgentRPCService as a thin shell:run facade * fix(agent-core-v2): align skill priority and sync docs - restore project > user > plugin > builtin skill precedence - sync Skill tool description and parameter docs from v1 - update write-goal and custom-theme builtin skill copy * fix(agent-core-v2): restore undo and thinking telemetry - track conversation_undo after undoHistory - emit thinking_toggle with enabled/effort/from payload - add coverage for both telemetry events * feat(agent-core-v2): add skill directory config - add extraSkillDirs and mergeAllAvailableSkills config sections - introduce extra skill source and shared source priorities - align kap-server workspace skill preview with session catalog * feat(agent-core-v2): support explicit skill dirs - add ISkillCatalogRuntimeOptions for SDK-style explicit skill dirs - suppress default user/project discovery when explicitDirs are set - resolve explicit dirs per session workDir via explicitFileSkillSource * fix(agent-core-v2): fix configured skill dir resolution - expand ~ using OS home for configured skill dirs - honor explicitDirs in kap-server workspace skill preview * fix(agent-core-v2): await config ready before skill discovery - wait for config.ready before reading extraSkillDirs - wait for config.ready before reading mergeAllAvailableSkills - cover extra skill dir loading behind config readiness * fix(agent-core-v2): keep skill config live after changes - await config.ready in kap-server workspace skill preview - reload user and workspace skill sources when mergeAllAvailableSkills changes * fix(agent-core-v2): align goal budget handling * fix(agent-core-v2): forbid model goal pauses * fix(agent-core-v2): cap detached process output * fix(kimi-code): drain v2 print subagents before exit * fix: restore kap-server video upload compatibility * fix(agent-core-v2): charge only output tokens against goal token budgets Goal parity gap G2: v1 charges only per-step output tokens against a goal's tokenBudget, while v2 summed all four usage buckets (cache read, cache creation, other input, output), exhausting budgets orders of magnitude faster under prompt caching and skewing persisted tokensUsed counters. Align goal token accounting to output-only and drop the unused tokenUsageTotal helper. * fix: align server-v2 media file handling * feat(agent-core-v2): allow coder profile to use MCP tools * refactor(agent-core): introduce activity kernel and migrate turn lane - add `activity` domain: `IAgentActivityService` (Agent turn lane machine), `ISessionActivityKernel` (Session admission, PR1 placeholder), and the `ActivityLease` that owns the turn `AbortSignal` - turnService launches and cancels through the kernel lease; `Turn` now exposes `signal` instead of `abortController` - agentLifecycle.remove drives `beginDisposal`/`settled` and waits for the in-flight turn to drain before releasing the agent scope - add `activity.*` error codes; deprecate `turn.agent_busy` in favor of `activity.agent_busy` * refactor(sessionLegacy): remove fork/compact/abort/archive pass-throughs These four legacy session actions were thin delegations to the native v2 services (ISessionLifecycleService.fork/archive, IAgentFullCompactionService.begin, IAgentRPCService.cancel) with no v1-only projection to centralize. Drop them from ISessionLegacyService and call the native services directly from the kap-server sessions route. updateProfile, createChild, listChildren, undo and status stay in the adapter since they carry real v1 adaptation logic. * refactor(cli): run print-mode v2 on native agent-core-v2 services - add native v2 print runner (v2/run-v2-print.ts) that consumes agent-core-v2 DI services and awaits Turn.result directly - extract shared print-mode rendering into prompt-render.ts for v1 and v2 - remove the V2PromptHarness/V2Session shim and v2->v1 event translation - decouple initializeCliTelemetry from PromptHarness (homeDir/auth/track) - add IAgentPromptLegacyService.submitAndSettle for authoritative completion * refactor(session): serve v1 undo and children via native v2 services - make IAgentPromptService.undo throw session.undo_unavailable with a structured reason; move the precheck into contextMemory - add ISessionLifecycleService.createChild (fork + child markers) and ISessionIndex.list({ childOf }) - slim ISessionLegacyService to updateProfile/status (drop createChild, listChildren, undo) - rewire kap-server session routes to the native services and map SESSION_UNDO_UNAVAILABLE * refactor(cli): drop v1 sdk and telemetry deps from v2 print - run-v2-print: use core ITelemetryService + CloudAppender instead of kimi-telemetry; remove kimi-code-sdk import (auth via IOAuthToolkit, config path from bootstrap, hook result via structural type) - prompt-render: replace SDK HookResultEvent with a structural type so the shared renderer does not depend on the v1 SDK event shape - telemetry: revert initializeCliTelemetry to its original signature now that v2 no longer calls it; keep v1 callers and assertions untouched - update run-prompt and v2-run-print tests for the new wiring * feat(activity): add session lane machine and agent snapshot projector - implement SessionActivityKernel lane machine (restoring→active⇄quiescing→closing→disposed) with admission table, atomic quiesce+drain, beginClosing/settled, markActive - start AgentActivityService lane at initializing; add markReady driven by agentLifecycle.create after bootstrap - project LaneModel + EventBus facts into structured AgentActivitySnapshot (ActivityModel / setActivitySnapshot Op) with pending-approval and active-tool-call sets; emit agent.activity.updated - add IAgentTurnService.launchWithLease; goal continuation acquires the lane before appending its prompt - resolve pending interactions on turn.ended to avoid stranded awaiting_approval - fullCompaction registers a background activity and checks the activity lane - extract contextMemory publishSplice / isFullyUndoable / recoverFoldedLength helpers - kap-server: map activity snapshot into legacy status and sessionEventBroadcaster * fix(agent-core-v2): truncate over-long goal completion criteria Goal parity gap G11: v1 silently truncates a goal's completionCriterion to 4000 characters (the objective cap) before persisting, so an over-long criterion never fails creation and cannot bloat every goal reminder and record. v2 only trimmed whitespace and persisted arbitrary lengths verbatim. Cap the normalized criterion at MAX_GOAL_COMPLETION_CRITERION_LENGTH to match v1. * fix(agent-core-v2): add goal error catalog info metadata Align the GoalErrors domain with V1 by attaching the info block for the seven goal.* error codes (title, retryable, public, action hints) so errorInfo() surfaces them. Entries copied verbatim from the V1 error catalog. Gap: G43 * chore(nix): update pnpm deps hash * fix(agent-core-v2): retain queued steers when a turn ends cancelled or failed Align the prompt layer with V1's steer-buffer semantics: buffered steer input now survives a turn that ends cancelled or failed and is flushed into the next launched turn by the existing beforeStep hook, instead of being silently dropped. The turn-result observation in the prompt service existed only to perform that discard, so it is removed along with the now-trivial launch wrapper; explicit clear() still discards the queue. Gap: G24 * fix(agent-core-v2): remove ask-user background mode * fix(kap-server): align archived session restore * chore(lint): fix type-aware lint errors * chore(agent-core-v2): drop stray doResume debug log * fix(agent-core-v2): defer prompts and steers while a full compaction is in flight Align with V1's compaction gating: input arriving while a full compaction holds the context (and no turn is active) used to launch a turn immediately, appending assistant output that forced the in-flight compaction to cancel. The prompt service now buffers such input and replays it from a new onDidFinishCompaction hook that the compaction worker runs in a finally, so the buffer drains on completion, cancellation, and failure alike — the first deferred item launches a turn and the rest join the steer queue. The compaction service is resolved lazily instead of constructor- injected: materializing it during prompt-service construction reorders loop-hook registration and moves the full-compaction beforeStep hook ahead of the hooks that let a freshly launched prompt land in context before the auto-compaction check snapshots history. Gap: G23 * fix(agent-core-v2): re-inject the goal reminder after full compaction Align with V1: after a compaction rewrites the context, re-arm the per-turn context injectors and run them before the compaction is marked complete, so the first post-compaction request — including a replayed deferred prompt's — already carries the goal reminder the summary folded away. The injector service exposes injectAfterCompaction, which re-arms the new-turn flag and injects immediately; the compaction worker calls it after the system-prompt refresh and raises the post-compaction token floor to include the re-injected reminders (the pre-injection floor stays as the fallback when reinjection throws), so the nothing-new-since-compaction guard does not re-trigger against a shape that cannot shrink. The injector is resolved lazily from the compaction service to keep loop-hook registration order untouched across the dependency cascade. Matches V1 verbatim including the existing quirk where an idle manual compact yields a second reminder copy on the next turn's per-turn injection; the parity test pins that behavior. Gap: G14 * test(agent-core-v2): cover goal pause classification for provider errors Port the missing end-to-end coverage: goal-driven turn failures pause the goal with the exact per-class reason strings — provider rate limit, provider connection error, provider authentication error, provider safety policy block, and model configuration error (including the forced 'LLM not set' substitution). Failures are driven through a real turn with a throwing generate stub so the raw-error classification feeding the pause reason is exercised, not just the mapper. No source changes: the existing classification already matches the reference strings verbatim. Gap: G35 * feat: add progressive tool disclosure * feat(cli): gate print-mode v2 behind KIMI_MODEL_EXPERIMENT_FLAG - add KIMI_PRINT_V2_ENV / isPrintV2Enabled so `kimi -p` routes to the native agent-core-v2 runner through its own switch - keep `kimi server run` server-v2 routing on isKimiV2Enabled (KIMI_CODE_EXPERIMENTAL_FLAG), decoupling the two - update print-mode tests and comments to reference the new switch * feat(cli): add KIMI_MODEL_OUTPUT_FORMAT for print-mode default - resolve the effective `-p` format via resolveOutputFormat: the --output-format flag wins, then KIMI_MODEL_OUTPUT_FORMAT (prompt mode only), then text - ignore the env outside prompt mode and reject invalid values eagerly through the friendly validation path - apply the resolver on both the v1 and v2 print runners * fix(agent-core-v2): count the goal-creating turn as the first goal turn Goal parity gap G5: when the model creates or resumes a goal mid-turn, v1 counts that ordinary turn as goal turn 1 at turn end (with a budget re-check before the continuation driver takes over) and charges its remaining step output tokens against the token budget. v2 only flagged turns whose goal was already active at launch, leaving turnsUsed and tokensUsed off by one turn in the model-initiated flow. Adopt the live turn as a goal starter turn on activation: charge its post-creation step output, count it once at turn end via incrementTurn, and block instead of launching a continuation when that count exhausts the turn budget. * fix(agent-core-v2): remove model-initiated paused status from UpdateGoal Goal parity gap G6: v1 reserves pausing for the user and runtime — its UpdateGoal tool only accepts active/complete/blocked and rejects other statuses with an invalid-status error. v2 still carried a leftover 'paused' enum option, a model pauseGoal branch, and matching tool description wording from before v1 removed them. Drop the paused option, port v1's runtime invalid-status guard, and align the tool description with v1's. * fix(agent-core-v2): deliver goal outcome prompts through the UpdateGoal tool result Goal parity gap G7: when the model completes or blocks a goal, v1 returns the outcome prompt (stats plus final-message instructions) as the UpdateGoal tool result with stopTurn, keys the one-shot final- message continuation on that terminal tool result, and guards it with the per-turn step budget so a capped turn ends 'completed' instead of dying on max steps. v2 still used a pre-change leftover channel: terse tool outputs plus goal_completion_summary / goal_blocked_reason system reminders and a last-message-reminder continuation with no step-budget check. Return the outcome prompts as tool output, drop the reminder appends and their detection, key the continuation on the terminal UpdateGoal result observed via the tool executor hook, and mirror v1's hasStepBudgetRemaining guard. Also closes audit gaps G17 (max-steps death) and G27 (actor-conditional reminders). * fix(agent-core-v2): fail UpdateGoal as a tool error when no goal matches Goal parity gap G8 (with user modification): v1 returns friendly success-flagged no-op outputs when UpdateGoal targets a missing or non-active goal, while v2 either let GOAL_NOT_FOUND escape from resumeGoal or reported false success with stopTurn for complete and blocked on a non-active goal. Per the user's decision these cases now return error-flagged tool results in the same shape as the Edit tool's old-string-not-found failure - v1's message texts ('Goal not resumed: no current goal.', 'Goal not completed: no active goal.', 'Goal not blocked: no active goal.') with isError and no stopTurn, so the model sees a non-fatal failure and the turn continues normally. * fix(agent-core-v2): settle active goals when the continuation relaunch fails Goal parity gap P-B: the turn-ended subscriber that relaunches goal continuation turns discarded every rejection, so a failed launch (for example losing a race to a queued prompt) stranded the goal in status active with nothing driving it. Keep the event-driven per-turn continuation model but settle deterministically on failure: any rejection out of the turn-ended handling now pauses the active goal as actor system with reason 'Paused after goal continuation failure: <message>', emitting the normal goal.updated event; the settle itself never throws into the event bus. The busy-skip needs no settle: the turn service clears its active turn before publishing turn.ended, so the other live turn's own end reliably re-runs the relaunch check. * fix(agent-core-v2): restore the fork-cleared goal system reminder Goal parity gap G12: after a session fork, v1 tells the model the fork has no current goal so it ignores stale active-goal reminders copied from the source session; v2 cleared the goal silently through the forked wire op and dropped the reminder. Track the fork boundary in a derived (never persisted) wire model folded on both dispatch and replay: a forked record that clears a copied goal marks the reminder pending, and the post-replay pass appends v1's verbatim reminder text with origin goal_fork_cleared exactly onc…
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What
Tightens the default agent system prompt and the built-in tool descriptions for accuracy and consistency, and removes a class of cases where the prompt pointed the model at tools the active profile does not provide.
Why
Tool-usage guidance was split awkwardly across two layers. The system prompt carried "how/when to use tool X" behind
{% if HAS_X %}gates that re-derive, in prose, the availability the tool schema already encodes — while the same guidance was duplicated in each tool's own description. Because subagent and tool-trimmed profiles share one system prompt but expose different tool subsets, those gates had to mirror the tool registry by hand and kept drifting: read-only subagents and custom profiles were told to useAgent,TaskList,TodoList, orEnterPlanModethey don't actually hold.Changes
{% if HAS_* %}blocks and the compaction tool-name bullets, folding the two genuinely unique lines into the Bash description ("return control after starting a background task") and the Agent description (the context-hygiene reason to delegate). Removed the now-unusedavailableTools/HAS_*render machinery; background guidance is now gated by the Bash tool itself, which requires the full task-tool set.Testing
agent-coresuite green,tsc --noEmitclean.