feat: experimental dynamic workflows — runtime, TUI, SDK, mode, autocomplete, docs by lucasfelipe24 · Pull Request #2280 · MoonshotAI/kimi-code

lucasfelipe24 · 2026-07-27T21:13:18Z

Problem

Kimi Code lacks a way to orchestrate multiple subagents with structured phases, parallel fan-out, pipelines, and validated outputs. Users who need multi-step research or complex automation must manually manage each subagent turn.

What changed

This PR adds Dynamic Workflows: an experimental feature that runs user-approved JavaScript scripts orchestrating subagents inside a restricted sandbox.

Engine v1 (packages/agent-core)

Workflow runtime: sandboxed node:vm with agent/parallel/pipeline API, JSON Schema validation, limits, cancellation, explicit failure/refusal propagation
Discovery in project/user/extra/builtin scopes with first-match precedence
WorkflowRunManager + WorkflowBackgroundTask with live events
Tool Workflow (proposal via model, approval via workflow_run display, background execution)
Permission policy workflow-run-review-ask (manual/yolo/auto)
WorkflowMode (orthogonal mode via /workflow on|off)

Engine v2 (packages/agent-core-v2)

IWorkflowCatalogService (App), IWorkflowRunService (Session), WorkflowModeService (Agent)
Tool/approval/config/flag ported

Protocol

WorkflowTaskInfo, WorkflowPhaseInfo, workflow.run.* events, display kind workflow_run

SDK (packages/node-sdk)

8 workflow methods + setWorkflowMode + events

TUI (apps/kimi-code)

/workflow command with 10 subcommands
Dynamic autocomplete (workflow names after /workflow run)
Full-screen runs browser with phase progress, logs, cancel/save/view
Footer badge for active workflows
Subagent suppression in transcript
Workflow mode badge

Builtin

deep-research.js (5 phases, adversarial verification, <=50 agent calls)

Docs

New pages EN+ZH + config/env/slash-commands updates

Checklist

Tests added (engine v1: 4113, v2: 4229, TUI: 2434, SDK: 238, protocol: 523)
Changeset generated
Docs updated

Register AI/ML, architecture, backend, database, debugging, design, DevOps, embedded, frontend, game dev, languages, project management, research writing, security, and testing skills as built-in bundles in both agent-core and agent-core-v2.

…ve bundle check Code examples in domain skill .md files contained require('uuid'), require('commander'), import('vue'), import ... from 'vue', and import('./types/user.mjs') patterns that the native bundle checker treated as real runtime dependencies.

# Conflicts: # apps/kimi-code/src/cli/commands.ts # docs/en/reference/kimi-command.md # docs/zh/reference/kimi-command.md

…tool usage in system prompt

…oading in system prompt

Restore the references/, rules/, and AGENTS.md files of the original domain skills into the built-in skill trees of both engines, and make them loadable at runtime: - SkillDefinition gains a resources map (skill-relative path to content), populated from ?raw imports in each bundle registration. - The Skill tool gains a resource parameter that returns a bundled resource file without invoking the skill (no activation, no depth). - The loaded skill block lists available resources under <bundled-resources>, and the system prompt documents the mechanism. - Escape require()/import() patterns in resource content so the native bundle check passes.

- Compare entry guards against pathToFileURL(argv[1]) so native build steps are not silently skipped when the checkout path contains spaces or non-ASCII characters (01-bundle.mjs, minidb server). - Use fileURLToPath instead of URL.pathname for MCP stdio fixtures and vis session fixtures, which broke under percent-encoded paths.

# Conflicts: # packages/agent-core-v2/src/app/agentProfileCatalog/system.md # packages/agent-core-v2/src/app/auth/webSearch/webSearchService.ts # packages/agent-core-v2/test/agent/loop/loop.test.ts # packages/agent-core-v2/test/tool/tool.test.ts

# Conflicts: # apps/kimi-code/src/tui/components/dialogs/api-key-input-dialog.ts

…al plugins

Drop the 16 domain expertise skill bundles (ai-ml, architecture, backend-frameworks, database, debugging, design, devops-infra, embedded-systems, find-skills, frontend-frameworks, game-developer, languages, project-management, research-writing, security, testing) from both engines, along with their SKILL.md files, sub-skills and bundled reference material. Revert the shared skill infrastructure to its prior state: the builtin registries, the SkillDefinition resources map, the Skill tool resource parameter, the loaded-block bundled-resources rendering, and the skill sections of both system prompts.

# Conflicts: # docs/en/configuration/config-files.md # docs/zh/configuration/config-files.md # packages/agent-core-v2/scripts/check-domain-layers.mjs # packages/agent-core-v2/src/app/auth/configSection.ts # packages/agent-core-v2/test/agent/mcp/stubs.ts # packages/agent-core/src/rpc/core-api.ts # packages/agent-core/src/rpc/core-impl.ts # packages/agent-core/test/harness/runtime.test.ts

# Conflicts: # packages/agent-core-v2/src/agent/tools/web-search/webSearchTool.ts # packages/agent-core-v2/src/app/auth/webSearch/webSearchService.ts # packages/agent-core-v2/src/index.ts

…omplete, docs Dynamic Workflows permitem orquestrar múltiplos subagentes a partir de scripts JS controlados (node:vm sandbox), com fases, fan-out paralelo, pipeline, saída estruturada validada por JSON Schema, logs/progresso e resultado final. Engine v1 (agent-core): - Workflow runtime sandbox (types, validate, script, runtime, discovery) - WorkflowRunManager + WorkflowBackgroundTask (kind workflow) - Descoberta em scopes: projeto (.kimi-code/workflows) → usuário → extra → builtin - Config section [workflows] (limites: concorrência, calls, duração, bytes) - Flag experimental 'dynamic-workflows' - Tool Workflow (proposal via model → approval → background run) - Permission policy workflow-run-review-ask (manual/yolo/auto) - WorkflowMode (modo ortogonal via system reminder) - argumentHint field na metadata Engine v2 (agent-core-v2): - IWorkflowCatalogService (App scope), IWorkflowRunService (Session scope) - Agent tool Workflow + approval + WorkflowModeService Protocol: - WorkflowTaskInfo, WorkflowPhaseInfo (+ zod schemas) - Eventos workflow.run.started/phase/log/agent_call/completed - Display kind workflow_run no ToolInputDisplay SDK (node-sdk): - listWorkflows, getWorkflow, reloadWorkflows, runWorkflow - listWorkflowRuns, getWorkflowRun, cancelWorkflowRun, saveWorkflow - setWorkflowMode TUI (kimi-code): - /workflow (alias /workflows) com subcomandos list/run/runs/show/cancel/save/reload/on/off - Autocomplete dinâmico: /workflow run <Tab> lista workflows disponíveis - Runs browser full-screen (status, fases, agend calls, logs, cancel, save, view script) - Workflow badge no footer ([N workflow(s) running]) - Painel de approval renderiza workflow_run display com script expansível - ProgressController + subagent suppression for cleaner transcript Builtin: - deep-research.js (5 fases: Scope/Search/Fetch/Verify/Synthesize, adversarial verification) Docs EN+ZH: - Páginas customization/workflows.md + seções em config-files, env-vars, slash-commands

changeset-bot · 2026-07-27T21:13:22Z

🦋 Changeset detected

Latest commit: c64223e

The changes in this PR will be included in the next version bump.

This PR includes changesets to release 1 package

Name	Type
@moonshot-ai/kimi-code	Minor

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

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: adc6853f73

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".

…workflow mode

lucasfelipe24 · 2026-07-27T21:24:28Z

Thanks for the thorough review! Here are my responses:

P1 — Sync loop blocking: Fixed in 3d57663. Added syncTimeoutMs: 100 to the compiled.run() call in both v1 (packages/agent-core/src/workflow/runtime.ts) and v2 (packages/agent-core-v2/src/app/workflow/runtime/runtime.ts). The script.ts already exposed the syncTimeoutMs option but the runtime was not passing it.

P1 — Builtin not bundled: Valid observation. The builtin deep-research.js is loaded via fs.readFile at runtime by the discovery module, so it needs to be present in the installed package. Currently both v1 and v2 only work in source-tree development. A follow-up fix should add the builtin directory to the package.json files field and ensure the build step copies *.js files from src/workflow/builtin/ to the output. This applies equally to v1 and v2.

P1 — Confirmation removed: This was an intentional design decision per user request. The /workflow run command is an explicit user action that doesn't need a second confirmation dialog. The model-invoked path (via the Workflow tool) still goes through the workflow-run-review-ask permission policy and shows the workflow_run approval panel. This matches how other explicit commands like /goal behave — they execute immediately without an extra confirmation popup.

P2 — WorkflowMode gate: Fixed in 3d57663. Added this.requireWorkflowsEnabled() guard to both enterWorkflowMode and exitWorkflowMode RPC handlers in packages/agent-core/src/session/rpc.ts, consistent with all other workflow RPC methods.

… engine

lucasfelipe24 · 2026-07-27T21:33:56Z

Engine v2 migration progress

As part of the ongoing migration to the v2 engine architecture (agent-core-v2), this PR now includes:

kap-server REST routes (8 new endpoints)

GET    /api/v1/sessions/{sid}/workflows              — list catalog
GET    /api/v1/sessions/{sid}/workflows/{name}        — get detail
POST   /api/v1/sessions/{sid}/workflows/run           — start run
GET    /api/v1/sessions/{sid}/workflows/runs          — list runs
GET    /api/v1/sessions/{sid}/workflows/runs/{runId}  — get run detail
POST   /api/v1/sessions/{sid}/workflows/runs/{runId}/cancel — cancel
POST   /api/v1/sessions/{sid}/workflows/save          — save to disk
POST   /api/v1/sessions/{sid}/workflows/reload        — reload catalog

Built on top of IWorkflowCatalogService (App scope) and IWorkflowRunService (Session scope), following the same patterns as tasks.ts.

klient contracts

workflowCatalogContract and workflowRunContract with Zod schemas
Registered in globalContract with facade methods on SessionFacade

TUI WorkflowV2Client

WorkflowV2Client adapter class that delegates to v2 (kap-server) when available, falling back to v1 Session
All workflow TUI commands now go through this adapter, enabling gradual engine migration without breaking existing functionality

Tests

kap-server: 6 route tests (kap-server/test/workflows.test.ts)
klient: 14 contract schema tests (klient/test/contract/session/workflow.test.ts)
TUI: 11 command tests passing

lucasfelipe24 added 21 commits July 18, 2026 17:10

feat(web-search): add LangSearch provider and reranking

1265c49

fix(web-search): preserve provider fallback credentials

7ed2c11

feat: add personal local build script and documentation

87956bc

a

2ee52b0

Merge remote-tracking branch 'upstream/main' into personal

f12b0b0

# Conflicts: # apps/kimi-code/src/cli/commands.ts # docs/en/reference/kimi-command.md # docs/zh/reference/kimi-command.md

fix: unlink native binary before copy to avoid ETXTBSY during SEA build

c53ac03

feat: make built-in skill bundles model-invocable and document Skill …

02ecf28

…tool usage in system prompt

Merge remote-tracking branch 'upstream/main' into personal

17f7771

feat: add when-to-use triggers to built-in skills and enforce skill l…

b447ad1

…oading in system prompt

Merge remote-tracking branch 'upstream/main' into personal

f0b5b89

# Conflicts: # apps/kimi-code/src/tui/components/dialogs/api-key-input-dialog.ts

feat(plugins): vendor image_generation and audio_generation as offici…

9fffd5b

…al plugins

Merge branch 'main' into personal

4327acf

# Conflicts: # packages/agent-core-v2/src/agent/tools/web-search/webSearchTool.ts # packages/agent-core-v2/src/app/auth/webSearch/webSearchService.ts # packages/agent-core-v2/src/index.ts

.

427e12d

chatgpt-codex-connector Bot reviewed Jul 27, 2026

View reviewed changes

Comment thread packages/agent-core/src/workflow/runtime.ts Outdated

Comment thread packages/agent-core/src/workflow/discovery.ts

Comment thread apps/kimi-code/src/tui/commands/workflow.ts Outdated

Comment thread packages/agent-core/src/session/rpc.ts

fix: address code review — sync timeout in vm runtime, flag gate for …

3d57663

…workflow mode

feat: add kap-server REST routes and klient contracts for workflow v2…

b40dc59

… engine

feat(web): add workflow system UI and server-side workflow mode wiring

c64223e

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

feat: experimental dynamic workflows — runtime, TUI, SDK, mode, autocomplete, docs - #2280

feat: experimental dynamic workflows — runtime, TUI, SDK, mode, autocomplete, docs#2280
lucasfelipe24 wants to merge 24 commits into
MoonshotAI:mainfrom
lucasfelipe24:feat/dynamic-workflows

lucasfelipe24 commented Jul 27, 2026

Uh oh!

changeset-bot Bot commented Jul 27, 2026 •

edited

Loading

Uh oh!

chatgpt-codex-connector Bot left a comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

lucasfelipe24 commented Jul 27, 2026

Uh oh!

lucasfelipe24 commented Jul 27, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Uh oh!

Conversation

lucasfelipe24 commented Jul 27, 2026

Problem

What changed

Engine v1 (packages/agent-core)

Engine v2 (packages/agent-core-v2)

Protocol

SDK (packages/node-sdk)

TUI (apps/kimi-code)

Builtin

Docs

Checklist

Uh oh!

changeset-bot Bot commented Jul 27, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

🦋 Changeset detected

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

lucasfelipe24 commented Jul 27, 2026

Uh oh!

lucasfelipe24 commented Jul 27, 2026

Engine v2 migration progress

kap-server REST routes (8 new endpoints)

klient contracts

TUI WorkflowV2Client

Tests

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

changeset-bot Bot commented Jul 27, 2026 •

edited

Loading