feat(agent): background server, management CLI, and mode-aware goal prompts - #134
Merged
Merged
Conversation
Phase 2 implementation: - Create plugins/lib/n00n/structured_output.lua with constants and helper functions for schema validation - Create plugins/lib/n00n/subagent.lua with unified subagent launch API - Migrate plugins/task/init.lua to use n00n.structured_output helper - Migrate plugins/workflow/init.lua to use n00n.structured_output helper - Add n00n agent run CLI command with stubs for future daemon commands - Add tests for structured_output helper in plugins/lib/tests/spec.lua - Keep team/roles.lua unchanged due to custom patterns (hardcoded audience, budget handling, preview support) All tests pass. Lua syntax validated with luac.
Implement `n00n agent run --background` Unix-socket server plus `message`, `stop`, and `list` client commands. Background agents keep state in `state_dir/agents/<id>/agent.json` and a `control.sock`. Server streams `TextDelta`/`ToolOutput`/`Done` events as NDJSON, and supports an initial `--prompt` before entering the command loop. The `--mode` flag selects the same workflow/team/etc modes used by the one-shot runner. - Add `AgentMode` CLI enum (General, Research, Task, Team, Workflow) - Add `async-lock` for per-agent message serialization - Add unit tests for CLI parsing, state serde, and state listing Phase 3 stubs (status, pause, resume, policy) remain unimplemented.
Contributor
📝 WalkthroughWalkthroughTask and workflow plugins now route structured-output execution through ChangesSubagent routing
Estimated code review effort: 4 (Complex) | ~45 minutes Possibly related PRs
Poem
🚥 Pre-merge checks | ✅ 4 | ❌ 1❌ Failed checks (1 warning)
✅ Passed checks (4 passed)
✨ Finishing Touches📝 Generate docstrings
🧪 Generate unit tests (beta)
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
Dependency Review✅ No vulnerabilities or license issues or OpenSSF Scorecard issues found.Scanned FilesNone |
Add `ClientCommand::Pause` and `Resume` to the background agent protocol. The server stores a shared `paused` flag and rejects new messages while paused, updating `agent.json` status accordingly. Implement `n00n agent status`, `pause`, and `resume` client commands. `status` reads the persisted agent state; `pause`/`resume` send a control command over the Unix socket. The `policy` subcommand remains a stub.
The text-mode client was streaming TextDelta chunks and then printing the full Done.text again when the run ended with an error. Only print the error in that case; the streamed text is already on the terminal.
![github-actions[bot]](https://avatars.githubusercontent.com/in/15368?s=60&v=4){kind=small}
There was a problem hiding this comment.
Criterion
Details
| Benchmark suite | Current: cf2092b | Previous: d464812 | Ratio |
|---|---|---|---|
fib/jit_mlua_hook |
7671510 ns/iter (± 229105) |
7284823 ns/iter (± 75675) |
1.05 |
fib/jit_watchdog |
1480787 ns/iter (± 36716) |
2603040 ns/iter (± 41118) |
0.57 |
fib/jit_none |
1477877 ns/iter (± 69862) |
2604751 ns/iter (± 36917) |
0.57 |
fib/interp_mlua_hook |
8105136 ns/iter (± 53510) |
7927475 ns/iter (± 19132) |
1.02 |
fib/interp_watchdog |
2898223 ns/iter (± 6957) |
3874557 ns/iter (± 81135) |
0.75 |
fib/interp_none |
2895160 ns/iter (± 10123) |
3816445 ns/iter (± 47350) |
0.76 |
buffer_rw/jit_mlua_hook |
735240 ns/iter (± 5791) |
572889 ns/iter (± 1758) |
1.28 |
buffer_rw/jit_watchdog |
89487 ns/iter (± 229) |
167764 ns/iter (± 597) |
0.53 |
buffer_rw/jit_none |
89519 ns/iter (± 1443) |
167883 ns/iter (± 194) |
0.53 |
buffer_rw/interp_mlua_hook |
1086904 ns/iter (± 6387) |
1062203 ns/iter (± 20350) |
1.02 |
buffer_rw/interp_watchdog |
456111 ns/iter (± 1465) |
642735 ns/iter (± 6550) |
0.71 |
buffer_rw/interp_none |
456154 ns/iter (± 6815) |
640283 ns/iter (± 7979) |
0.71 |
splash_render_120x40 |
51201 ns/iter (± 3296) |
65472 ns/iter (± 6958) |
0.78 |
splash_render_200x60 |
128923 ns/iter (± 4796) |
154383 ns/iter (± 12329) |
0.84 |
This comment was automatically generated by workflow using github-action-benchmark.
There was a problem hiding this comment.
⚠️ Performance Alert ⚠️
Possible performance regression was detected for benchmark 'Criterion'.
Benchmark result of this commit is worse than the previous benchmark result exceeding threshold 2.
| Benchmark suite | Current: 698a7b3 | Previous: 5883bbf | Ratio |
|---|---|---|---|
buffer_rw/jit_watchdog |
191486 ns/iter (± 313) |
75535 ns/iter (± 1409) |
2.54 |
buffer_rw/jit_none |
191954 ns/iter (± 473) |
76858 ns/iter (± 905) |
2.50 |
splash_render_120x40 |
80553 ns/iter (± 806) |
32557 ns/iter (± 1739) |
2.47 |
This comment was automatically generated by workflow using github-action-benchmark.
Codecov Report❌ Patch coverage is
📢 Thoughts on this report? Let us know! |
Extend `plugins/lib/n00n/subagent.lua` with options the team plugin needs:
- `system` override for role-specific prompts
- `preview` and `activity_label` for ActivityPreview integration
- `budget` with `:consume()` for agent-call budgets
- `fail_on_pricing_error` to keep the existing team semantics
- return the resolved `model_spec` as a fifth value
Migrate `plugins/team/roles.lua` to launch subagents through the shared
helper, preserving its `{ok, text, cost, model, usage, error}` return
shape and all existing tests.
Simplify `plugins/team/init.lua` `run_supervisor` by delegating the
structured-output plan generation to `n00n.subagent.launch` with
`output_schema = PLANNER_OUTPUT`.
Surface failed socket/directory removals in the background agent stop handler instead of silently ignoring them. The process still exits, but the warning gives operators a signal that stale state may remain.
Prevent path traversal from user-supplied agent ids by validating them before any filesystem operation, and reject ids that contain path separators, control characters, or other unsafe content. Limit ids to 64 ASCII alphanumeric/hyphen/underscore characters. Set the agent state directory to 0o700 and the Unix control socket to 0o600 so only the owning user can read or connect to background agent state.
Extract `prepare_agent_env` and `PreparedEnv` to remove the duplicated plugin/model/MCP initialization between `run` and `server`. Both paths now build their `HeadlessParams`/`InteractiveParams` from the same prepared environment, making the two entry points easier to keep in sync.
Record the user-facing additions, refactor, and security hardening from PR #134 in changelog.d.
The background agent Stop command now waits for the agent task to finish after sending the cancel signal, instead of exiting immediately. This prevents dropping active tool calls and unfinished MCP sessions and ensures state is persisted before cleanup.
The `goal` flag on `n00n agent run` was parsed but never used, and the `policy` subcommands were stubs. Remove both until they are wired to actual behavior.
Thread the `thinking` option through `n00n.agent.session` in both the task and workflow plugins, and include it in the workflow journal key so cached results honor the thinking setting.
…tive `headless::spawn_interactive` was calling `permissions.toggle_yolo()` when `params.yolo` was true, which flipped an already-true yolo state back to false. This broke `--yolo` in background agent mode (and TUI/ACP when always_yolo was set). Add `PermissionManager::set_yolo` and use it to reliably enable yolo when requested, leaving the config-derived default otherwise.
Re-add --goal to n00n agent run and wire it into mode-specific prompts. Team mode now emits a team tool call with goal, mode, max_agents, waves, and auto-tier flags. Workflow mode emits a workflow tool call with the prompt as the script and goal merged into inputs. Task mode emits a task tool call with description/prompt. General/Research modes prepend the goal to the prompt. Also introduce AgentRunOptions to keep run/server signatures tidy and avoid excessive boolean parameters.
w0wl0lxd
changed the title
Remove the hard-coded 24/32 agent-call ceilings in team and workflow. - team `max_agents` is now configurable with no hard maximum and an optional `timeout_secs`. - workflow `max_agents_per_run` is configurable with no hard maximum. - Add `plugins/lib/n00n/guard.lua` to enforce call budgets, wall-clock timeouts, repeated-prompt loops, and consecutive subagent errors. - Wire the guard through `n00n.subagent.launch`, `team`, and `workflow`.
…tive errors - Move the used counter increment from record() into check() so the budget is reserved before any async yield; this prevents concurrent subagent calls from overshooting max_calls. - Check consecutive_errors in check() and use > in record() so the guard blocks the next call after the threshold instead of wasting a token on the failing call. - Add lib spec coverage for budget, repeated-prompt, consecutive-error, and legacy consume() behavior.
w0wl0lxd
marked this pull request as ready for review
July 27, 2026 06:48
![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: 22d4ff09f0
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".
4 of 7 tasks
w0wl0lxd
added a commit
that referenced
this pull request
Jul 27, 2026
Unify background worker socks under n00n agent run --background with daemon-first list/status/message/pause/resume/stop, plus agent daemon subcommand for headless worker planes.
Owner
Author
Owner
Author
|
cursor review |
Bugbot couldn't run - usage limit reachedBugbot is counted against Cursor usage for this user or team, and this run hit a usage or spend limit. A user or team admin can review and increase usage limits in the Cursor dashboard. (requestId: serverGenReqId_f8e60ee7-46ac-4537-9ea3-0f5f814aa589) |
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
Contributor
There was a problem hiding this comment.
Actionable comments posted: 5
Caution
Some comments are outside the diff and can’t be posted inline due to platform limitations.
⚠️ Outside diff range comments (1)
plugins/task/init.lua (1)
157-177: 📐 Maintainability & Code Quality | 🟠 Major | ⚡ Quick winDead
local_toolstable — remove it.
local_toolsis assigned here but never read; the legacy path builds its owndone_toolat Line 252, and the structured path delegates tosubagent.launch. The handler at Line 172-174 also dropsvalueentirely, making this block a misleading duplicate. Luacheck flags it (W231).♻️ Remove the unused block
local preview = make_preview(ctx, input.description or "task") - - -- Build local tools: either structured_output (with schema) or done tool - local local_tools - if not input.output_schema then - local_tools = { - [DONE_NAME] = { - description = DONE_DESCRIPTION, - input_schema = { - type = "object", - properties = { - answer = { type = "string", description = "Final answer to return to the parent agent." }, - }, - required = { "answer" }, - }, - handler = function(value) - return "Done." - end, - }, - } - end🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@plugins/task/init.lua` around lines 157 - 177, Remove the unused local_tools declaration and its conditional DONE_NAME handler block from the task initialization flow after make_preview; retain the existing legacy done_tool construction and structured subagent.launch path unchanged.Source: Linters/SAST tools
🤖 Prompt for all review comments with AI agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
Inline comments:
In `@plugins/task/init.lua`:
- Around line 307-318: Update the local assignment in do_poll to discard the
unused second return value from sess:get_progress, using a single progress
variable or an explicit underscore while preserving the existing polling and
completion logic.
- Around line 217-222: Update the return path that builds the `llm_output`
result to capture and validate the error returned by
`n00n.json.encode(captured)`, matching the checked encoding patterns elsewhere
in the handler. Handle encoding failure before returning so `ctx:finish` never
receives a nil `llm_output`.
In `@plugins/workflow/init.lua`:
- Around line 413-421: Update the subagent.launch result handling in the
workflow initialization flow to avoid silently capturing unused cost and
usage_val values: either remove them from the destructuring or propagate them
through existing progress/telemetry, preferably the agent_done event, so
sub-agent token spend remains observable.
- Around line 430-439: Update the completion handling around launch_err so
progress.agent_done and the logger.log("agent_done", ...) telemetry execute only
when launch_err is absent; preserve the existing error("sub-agent error: " ..
launch_err, 0) path for failed launches.
- Around line 441-450: Update the output handling around subagent.launch so
captured string results are stored unchanged, while table results are
JSON-encoded with the existing encode error handling. Use the type of captured
to select the path, preserving the empty fallback for nil values and assigning
the encoded value only for structured table output.
---
Outside diff comments:
In `@plugins/task/init.lua`:
- Around line 157-177: Remove the unused local_tools declaration and its
conditional DONE_NAME handler block from the task initialization flow after
make_preview; retain the existing legacy done_tool construction and structured
subagent.launch path unchanged.
🪄 Autofix (Beta)
Fix all unresolved CodeRabbit comments on this PR:
- Push a commit to this branch (recommended)
- Create a new PR with the fixes
ℹ️ Review info
⚙️ Run configuration
Configuration used: Organization UI
Review profile: ASSERTIVE
Plan: Pro Plus
Run ID: 0dc01059-d475-4f57-82a8-e59938efc332
📒 Files selected for processing (3)
changelog.d/134.changed.mdplugins/task/init.luaplugins/workflow/init.lua
📜 Review details
⏰ Context from checks skipped due to timeout. (12)
- GitHub Check: Test (macOS)
- GitHub Check: Test (Windows)
- GitHub Check: Build (Windows)
- GitHub Check: MSRV (1.97)
- GitHub Check: Coverage
- GitHub Check: Lint (macOS)
- GitHub Check: Lint
- GitHub Check: Lint (Windows)
- GitHub Check: Test
- GitHub Check: Build
- GitHub Check: Docs
- GitHub Check: Analyze (rust)
🧰 Additional context used
📓 Path-based instructions (2)
**/*
📄 CodeRabbit inference engine (AGENTS.md)
**/*: Non-trivial or multi-file changes should use a dedicated git worktree and new branch; do not modify unrelated user changes, force-push, or push to main.
Ship finished work with a clear Conventional Commit message, a pushed branch, and a draft pull request; never add AI-agent attribution to authored content.
Before investigating unfamiliar failures or third-party behavior, research documented behavior first and distinguish unrelated baseline failures from regressions in the touched surface.
Use structural tools before broad searches: prefercodegraphorarborfor cross-file relationships,indexbefore reading files, targeted reads, parallel calls, andcode_executionfor filtering large outputs.
Files:
changelog.d/134.changed.mdplugins/workflow/init.luaplugins/task/init.lua
plugins/**/*.lua
📄 CodeRabbit inference engine (AGENTS.md)
Built-in Lua plugins belong under
./pluginsand should use the repository's plugin tooling and conventions.
Files:
plugins/workflow/init.luaplugins/task/init.lua
🪛 Luacheck (1.2.0)
plugins/workflow/init.lua
[warning] 413-413: unused variable 'cost'
(W211)
[warning] 413-413: unused variable 'usage_val'
(W211)
plugins/task/init.lua
[warning] 160-160: variable 'local_tools' is never accessed
(W231)
[warning] 309-309: unused variable 'err'
(W211)
🪛 markdownlint-cli2 (0.23.0)
changelog.d/134.changed.md
[warning] 1-1: First line in a file should be a top-level heading
(MD041, first-line-heading, first-line-h1)
🔇 Additional comments (4)
plugins/task/init.lua (2)
12-12: LGTM!
223-306: LGTM!Also applies to: 319-327
plugins/workflow/init.lua (1)
24-24: LGTM!changelog.d/134.changed.md (1)
1-1: LGTM!
- Fix background detach: fork and setsid to daemonize server process - Fix stop command: close input_tx and await outer task for proper cleanup - Fix error handling: break event loop on Error to prevent state hang - Fix task plugin: remove unused error from get_progress, fix encode error handling - Fix workflow plugin: emit agent_done only on success, remove unused cost/usage returns
Bugbot couldn't run - usage limit reachedBugbot is counted against Cursor usage for this user or team, and this run hit a usage or spend limit. A user or team admin can review and increase usage limits in the Cursor dashboard. (requestId: serverGenReqId_78dfef73-cbf8-494b-ae25-b3342b839a63) |
Owner
Author
|
cursor review |
Bugbot couldn't run - usage limit reachedBugbot is counted against Cursor usage for this user or team, and this run hit a usage or spend limit. A user or team admin can review and increase usage limits in the Cursor dashboard. (requestId: serverGenReqId_c081034c-8f4f-45de-bfbb-f8217c1523a9) |
- Return error on AgentEvent::Error in one-shot run for proper exit code - Preserve agent state on read failures (only delete on not found) - Avoid overwriting paused status when active run completes - UTF-8 truncation already uses character-aware .chars().take()
Bugbot couldn't run - usage limit reachedBugbot is counted against Cursor usage for this user or team, and this run hit a usage or spend limit. A user or team admin can review and increase usage limits in the Cursor dashboard. (requestId: serverGenReqId_41e21cfe-671c-42ee-839c-e8ae3d69b1bc) |
Owner
Author
|
cursor review |
Bugbot couldn't run - usage limit reachedBugbot is counted against Cursor usage for this user or team, and this run hit a usage or spend limit. A user or team admin can review and increase usage limits in the Cursor dashboard. (requestId: serverGenReqId_c2b66b2f-d547-4fab-8a69-8226271b2d75) |
Bugbot couldn't run - usage limit reachedBugbot is counted against Cursor usage for this user or team, and this run hit a usage or spend limit. A user or team admin can review and increase usage limits in the Cursor dashboard. (requestId: serverGenReqId_6f4f499a-0e4f-4e75-8caf-f4f725a5a07f) |
Owner
Author
|
cursor review |
Bugbot couldn't run - usage limit reachedBugbot is counted against Cursor usage for this user or team, and this run hit a usage or spend limit. A user or team admin can review and increase usage limits in the Cursor dashboard. (requestId: serverGenReqId_4e14ddea-1325-4b7e-8e38-036d7bc38f57) |
…ctions - Add AgentMode::Research and is_readonly() so build_system_prompt picks the research prompt and filter_tools_for_mode strips execute-kind tools. - Plumb runtime AgentMode through HeadlessParams/InteractiveParams and persist it in background AgentState so follow-up messages keep research mode. - Exclude write/edit/multiedit/edit_lines/insert_lines for research runs. - Fix task plugin preview to use ActivityPreview (required by n00n.subagent) and return raw structured-output validation errors for tests.
Bugbot couldn't run - usage limit reachedBugbot is counted against Cursor usage for this user or team, and this run hit a usage or spend limit. A user or team admin can review and increase usage limits in the Cursor dashboard. (requestId: serverGenReqId_81fc75de-b42a-4ec6-844b-03c9cb07562d) |
…tables Subagent.launch may return either a plain string or a structured table. Store strings unchanged, JSON-encode tables with error handling, and use an empty string for nil output instead of encoding nil.
Bugbot couldn't run - usage limit reachedBugbot is counted against Cursor usage for this user or team, and this run hit a usage or spend limit. A user or team admin can review and increase usage limits in the Cursor dashboard. (requestId: serverGenReqId_726566c2-a243-4022-b30d-fdbecb0c86b0) |
w0wl0lxd
added a commit
that referenced
this pull request
Jul 29, 2026
…rompts (#134) * feat(agent): add structured output helper and agent CLI command Phase 2 implementation: - Create plugins/lib/n00n/structured_output.lua with constants and helper functions for schema validation - Create plugins/lib/n00n/subagent.lua with unified subagent launch API - Migrate plugins/task/init.lua to use n00n.structured_output helper - Migrate plugins/workflow/init.lua to use n00n.structured_output helper - Add n00n agent run CLI command with stubs for future daemon commands - Add tests for structured_output helper in plugins/lib/tests/spec.lua - Keep team/roles.lua unchanged due to custom patterns (hardcoded audience, budget handling, preview support) All tests pass. Lua syntax validated with luac. * fix(agent-orchestration): repair route_tier and structured_output helpers * feat(agent): add background agent server and management CLI Implement `n00n agent run --background` Unix-socket server plus `message`, `stop`, and `list` client commands. Background agents keep state in `state_dir/agents/<id>/agent.json` and a `control.sock`. Server streams `TextDelta`/`ToolOutput`/`Done` events as NDJSON, and supports an initial `--prompt` before entering the command loop. The `--mode` flag selects the same workflow/team/etc modes used by the one-shot runner. - Add `AgentMode` CLI enum (General, Research, Task, Team, Workflow) - Add `async-lock` for per-agent message serialization - Add unit tests for CLI parsing, state serde, and state listing Phase 3 stubs (status, pause, resume, policy) remain unimplemented. * feat(agent): implement status, pause, and resume commands Add `ClientCommand::Pause` and `Resume` to the background agent protocol. The server stores a shared `paused` flag and rejects new messages while paused, updating `agent.json` status accordingly. Implement `n00n agent status`, `pause`, and `resume` client commands. `status` reads the persisted agent state; `pause`/`resume` send a control command over the Unix socket. The `policy` subcommand remains a stub. * fix(agent): avoid duplicate text when message run ends with error The text-mode client was streaming TextDelta chunks and then printing the full Done.text again when the run ended with an error. Only print the error in that case; the streamed text is already on the terminal. * refactor(team): migrate roles and supervisor to n00n.subagent Extend `plugins/lib/n00n/subagent.lua` with options the team plugin needs: - `system` override for role-specific prompts - `preview` and `activity_label` for ActivityPreview integration - `budget` with `:consume()` for agent-call budgets - `fail_on_pricing_error` to keep the existing team semantics - return the resolved `model_spec` as a fifth value Migrate `plugins/team/roles.lua` to launch subagents through the shared helper, preserving its `{ok, text, cost, model, usage, error}` return shape and all existing tests. Simplify `plugins/team/init.lua` `run_supervisor` by delegating the structured-output plan generation to `n00n.subagent.launch` with `output_schema = PLANNER_OUTPUT`. * fix(agent): log cleanup warnings in stop handler Surface failed socket/directory removals in the background agent stop handler instead of silently ignoring them. The process still exits, but the warning gives operators a signal that stale state may remain. * fix(agent): validate agent ids and lock down socket permissions Prevent path traversal from user-supplied agent ids by validating them before any filesystem operation, and reject ids that contain path separators, control characters, or other unsafe content. Limit ids to 64 ASCII alphanumeric/hyphen/underscore characters. Set the agent state directory to 0o700 and the Unix control socket to 0o600 so only the owning user can read or connect to background agent state. * refactor(agent): share one-shot and background agent setup Extract `prepare_agent_env` and `PreparedEnv` to remove the duplicated plugin/model/MCP initialization between `run` and `server`. Both paths now build their `HeadlessParams`/`InteractiveParams` from the same prepared environment, making the two entry points easier to keep in sync. * docs(changelog): add fragments for agent CLI and team refactor Record the user-facing additions, refactor, and security hardening from PR #134 in changelog.d. * fix(agent): await task completion before stop exits The background agent Stop command now waits for the agent task to finish after sending the cancel signal, instead of exiting immediately. This prevents dropping active tool calls and unfinished MCP sessions and ensures state is persisted before cleanup. * refactor(cli): remove unused goal arg and stub policy commands The `goal` flag on `n00n agent run` was parsed but never used, and the `policy` subcommands were stubs. Remove both until they are wired to actual behavior. * fix(task,workflow): pass thinking config to subagent sessions Thread the `thinking` option through `n00n.agent.session` in both the task and workflow plugins, and include it in the workflow journal key so cached results honor the thinking setting. * fix(n00n-agent): make yolo enable rather than toggle in spawn_interactive `headless::spawn_interactive` was calling `permissions.toggle_yolo()` when `params.yolo` was true, which flipped an already-true yolo state back to false. This broke `--yolo` in background agent mode (and TUI/ACP when always_yolo was set). Add `PermissionManager::set_yolo` and use it to reliably enable yolo when requested, leaving the config-derived default otherwise. * feat(agent): add --goal and mode-aware team/workflow/task prompts Re-add --goal to n00n agent run and wire it into mode-specific prompts. Team mode now emits a team tool call with goal, mode, max_agents, waves, and auto-tier flags. Workflow mode emits a workflow tool call with the prompt as the script and goal merged into inputs. Task mode emits a task tool call with description/prompt. General/Research modes prepend the goal to the prompt. Also introduce AgentRunOptions to keep run/server signatures tidy and avoid excessive boolean parameters. * feat(agent): configurable agent-call limits and runaway guard Remove the hard-coded 24/32 agent-call ceilings in team and workflow. - team `max_agents` is now configurable with no hard maximum and an optional `timeout_secs`. - workflow `max_agents_per_run` is configurable with no hard maximum. - Add `plugins/lib/n00n/guard.lua` to enforce call budgets, wall-clock timeouts, repeated-prompt loops, and consecutive subagent errors. - Wire the guard through `n00n.subagent.launch`, `team`, and `workflow`. * fix(guard): reserve call budget atomically in check and guard consecutive errors - Move the used counter increment from record() into check() so the budget is reserved before any async yield; this prevents concurrent subagent calls from overshooting max_calls. - Check consecutive_errors in check() and use > in record() so the guard blocks the next call after the threshold instead of wasting a token on the failing call. - Add lib spec coverage for budget, repeated-prompt, consecutive-error, and legacy consume() behavior. * fix(agent): gate background agent IPC behind Unix-only cfg Background agents use Unix domain sockets and permission modes that are not available on Windows. Return a clear error on unsupported platforms so the crate builds on all CI targets. * fix(nix): use platform-correct loader env var in wrappers Use DYLD_LIBRARY_PATH on Darwin and LD_LIBRARY_PATH elsewhere, and wrap the computed package binary path to avoid runtime loader failures on non-Linux builds. * refactor(plugins): migrate task and workflow to n00n.subagent.launch for structured output * fix(agent): resolve P1 issues in background server and error handling - Fix background detach: fork and setsid to daemonize server process - Fix stop command: close input_tx and await outer task for proper cleanup - Fix error handling: break event loop on Error to prevent state hang - Fix task plugin: remove unused error from get_progress, fix encode error handling - Fix workflow plugin: emit agent_done only on success, remove unused cost/usage returns * fix(agent): address P2 review feedback in agent.rs - Return error on AgentEvent::Error in one-shot run for proper exit code - Preserve agent state on read failures (only delete on not found) - Avoid overwriting paused status when active run completes - UTF-8 truncation already uses character-aware .chars().take() * fix(team): enforce wall-clock timeout during subagent calls * feat(agent): wire research mode through system prompt and tool restrictions - Add AgentMode::Research and is_readonly() so build_system_prompt picks the research prompt and filter_tools_for_mode strips execute-kind tools. - Plumb runtime AgentMode through HeadlessParams/InteractiveParams and persist it in background AgentState so follow-up messages keep research mode. - Exclude write/edit/multiedit/edit_lines/insert_lines for research runs. - Fix task plugin preview to use ActivityPreview (required by n00n.subagent) and return raw structured-output validation errors for tests. * fix(workflow): preserve string subagent output and encode structured tables Subagent.launch may return either a plain string or a structured table. Store strings unchanged, JSON-encode tables with error handling, and use an empty string for nil output instead of encoding nil.
w0wl0lxd
added a commit
that referenced
this pull request
Jul 29, 2026
…rompts (#134) * feat(agent): add structured output helper and agent CLI command Phase 2 implementation: - Create plugins/lib/n00n/structured_output.lua with constants and helper functions for schema validation - Create plugins/lib/n00n/subagent.lua with unified subagent launch API - Migrate plugins/task/init.lua to use n00n.structured_output helper - Migrate plugins/workflow/init.lua to use n00n.structured_output helper - Add n00n agent run CLI command with stubs for future daemon commands - Add tests for structured_output helper in plugins/lib/tests/spec.lua - Keep team/roles.lua unchanged due to custom patterns (hardcoded audience, budget handling, preview support) All tests pass. Lua syntax validated with luac. * fix(agent-orchestration): repair route_tier and structured_output helpers * feat(agent): add background agent server and management CLI Implement `n00n agent run --background` Unix-socket server plus `message`, `stop`, and `list` client commands. Background agents keep state in `state_dir/agents/<id>/agent.json` and a `control.sock`. Server streams `TextDelta`/`ToolOutput`/`Done` events as NDJSON, and supports an initial `--prompt` before entering the command loop. The `--mode` flag selects the same workflow/team/etc modes used by the one-shot runner. - Add `AgentMode` CLI enum (General, Research, Task, Team, Workflow) - Add `async-lock` for per-agent message serialization - Add unit tests for CLI parsing, state serde, and state listing Phase 3 stubs (status, pause, resume, policy) remain unimplemented. * feat(agent): implement status, pause, and resume commands Add `ClientCommand::Pause` and `Resume` to the background agent protocol. The server stores a shared `paused` flag and rejects new messages while paused, updating `agent.json` status accordingly. Implement `n00n agent status`, `pause`, and `resume` client commands. `status` reads the persisted agent state; `pause`/`resume` send a control command over the Unix socket. The `policy` subcommand remains a stub. * fix(agent): avoid duplicate text when message run ends with error The text-mode client was streaming TextDelta chunks and then printing the full Done.text again when the run ended with an error. Only print the error in that case; the streamed text is already on the terminal. * refactor(team): migrate roles and supervisor to n00n.subagent Extend `plugins/lib/n00n/subagent.lua` with options the team plugin needs: - `system` override for role-specific prompts - `preview` and `activity_label` for ActivityPreview integration - `budget` with `:consume()` for agent-call budgets - `fail_on_pricing_error` to keep the existing team semantics - return the resolved `model_spec` as a fifth value Migrate `plugins/team/roles.lua` to launch subagents through the shared helper, preserving its `{ok, text, cost, model, usage, error}` return shape and all existing tests. Simplify `plugins/team/init.lua` `run_supervisor` by delegating the structured-output plan generation to `n00n.subagent.launch` with `output_schema = PLANNER_OUTPUT`. * fix(agent): log cleanup warnings in stop handler Surface failed socket/directory removals in the background agent stop handler instead of silently ignoring them. The process still exits, but the warning gives operators a signal that stale state may remain. * fix(agent): validate agent ids and lock down socket permissions Prevent path traversal from user-supplied agent ids by validating them before any filesystem operation, and reject ids that contain path separators, control characters, or other unsafe content. Limit ids to 64 ASCII alphanumeric/hyphen/underscore characters. Set the agent state directory to 0o700 and the Unix control socket to 0o600 so only the owning user can read or connect to background agent state. * refactor(agent): share one-shot and background agent setup Extract `prepare_agent_env` and `PreparedEnv` to remove the duplicated plugin/model/MCP initialization between `run` and `server`. Both paths now build their `HeadlessParams`/`InteractiveParams` from the same prepared environment, making the two entry points easier to keep in sync. * docs(changelog): add fragments for agent CLI and team refactor Record the user-facing additions, refactor, and security hardening from PR #134 in changelog.d. * fix(agent): await task completion before stop exits The background agent Stop command now waits for the agent task to finish after sending the cancel signal, instead of exiting immediately. This prevents dropping active tool calls and unfinished MCP sessions and ensures state is persisted before cleanup. * refactor(cli): remove unused goal arg and stub policy commands The `goal` flag on `n00n agent run` was parsed but never used, and the `policy` subcommands were stubs. Remove both until they are wired to actual behavior. * fix(task,workflow): pass thinking config to subagent sessions Thread the `thinking` option through `n00n.agent.session` in both the task and workflow plugins, and include it in the workflow journal key so cached results honor the thinking setting. * fix(n00n-agent): make yolo enable rather than toggle in spawn_interactive `headless::spawn_interactive` was calling `permissions.toggle_yolo()` when `params.yolo` was true, which flipped an already-true yolo state back to false. This broke `--yolo` in background agent mode (and TUI/ACP when always_yolo was set). Add `PermissionManager::set_yolo` and use it to reliably enable yolo when requested, leaving the config-derived default otherwise. * feat(agent): add --goal and mode-aware team/workflow/task prompts Re-add --goal to n00n agent run and wire it into mode-specific prompts. Team mode now emits a team tool call with goal, mode, max_agents, waves, and auto-tier flags. Workflow mode emits a workflow tool call with the prompt as the script and goal merged into inputs. Task mode emits a task tool call with description/prompt. General/Research modes prepend the goal to the prompt. Also introduce AgentRunOptions to keep run/server signatures tidy and avoid excessive boolean parameters. * feat(agent): configurable agent-call limits and runaway guard Remove the hard-coded 24/32 agent-call ceilings in team and workflow. - team `max_agents` is now configurable with no hard maximum and an optional `timeout_secs`. - workflow `max_agents_per_run` is configurable with no hard maximum. - Add `plugins/lib/n00n/guard.lua` to enforce call budgets, wall-clock timeouts, repeated-prompt loops, and consecutive subagent errors. - Wire the guard through `n00n.subagent.launch`, `team`, and `workflow`. * fix(guard): reserve call budget atomically in check and guard consecutive errors - Move the used counter increment from record() into check() so the budget is reserved before any async yield; this prevents concurrent subagent calls from overshooting max_calls. - Check consecutive_errors in check() and use > in record() so the guard blocks the next call after the threshold instead of wasting a token on the failing call. - Add lib spec coverage for budget, repeated-prompt, consecutive-error, and legacy consume() behavior. * fix(agent): gate background agent IPC behind Unix-only cfg Background agents use Unix domain sockets and permission modes that are not available on Windows. Return a clear error on unsupported platforms so the crate builds on all CI targets. * fix(nix): use platform-correct loader env var in wrappers Use DYLD_LIBRARY_PATH on Darwin and LD_LIBRARY_PATH elsewhere, and wrap the computed package binary path to avoid runtime loader failures on non-Linux builds. * refactor(plugins): migrate task and workflow to n00n.subagent.launch for structured output * fix(agent): resolve P1 issues in background server and error handling - Fix background detach: fork and setsid to daemonize server process - Fix stop command: close input_tx and await outer task for proper cleanup - Fix error handling: break event loop on Error to prevent state hang - Fix task plugin: remove unused error from get_progress, fix encode error handling - Fix workflow plugin: emit agent_done only on success, remove unused cost/usage returns * fix(agent): address P2 review feedback in agent.rs - Return error on AgentEvent::Error in one-shot run for proper exit code - Preserve agent state on read failures (only delete on not found) - Avoid overwriting paused status when active run completes - UTF-8 truncation already uses character-aware .chars().take() * fix(team): enforce wall-clock timeout during subagent calls * feat(agent): wire research mode through system prompt and tool restrictions - Add AgentMode::Research and is_readonly() so build_system_prompt picks the research prompt and filter_tools_for_mode strips execute-kind tools. - Plumb runtime AgentMode through HeadlessParams/InteractiveParams and persist it in background AgentState so follow-up messages keep research mode. - Exclude write/edit/multiedit/edit_lines/insert_lines for research runs. - Fix task plugin preview to use ActivityPreview (required by n00n.subagent) and return raw structured-output validation errors for tests. * fix(workflow): preserve string subagent output and encode structured tables Subagent.launch may return either a plain string or a structured table. Store strings unchanged, JSON-encode tables with error handling, and use an empty string for nil output instead of encoding nil.
w0wl0lxd
added a commit
that referenced
this pull request
Jul 29, 2026
…rompts (#134) * feat(agent): add structured output helper and agent CLI command Phase 2 implementation: - Create plugins/lib/n00n/structured_output.lua with constants and helper functions for schema validation - Create plugins/lib/n00n/subagent.lua with unified subagent launch API - Migrate plugins/task/init.lua to use n00n.structured_output helper - Migrate plugins/workflow/init.lua to use n00n.structured_output helper - Add n00n agent run CLI command with stubs for future daemon commands - Add tests for structured_output helper in plugins/lib/tests/spec.lua - Keep team/roles.lua unchanged due to custom patterns (hardcoded audience, budget handling, preview support) All tests pass. Lua syntax validated with luac. * fix(agent-orchestration): repair route_tier and structured_output helpers * feat(agent): add background agent server and management CLI Implement `n00n agent run --background` Unix-socket server plus `message`, `stop`, and `list` client commands. Background agents keep state in `state_dir/agents/<id>/agent.json` and a `control.sock`. Server streams `TextDelta`/`ToolOutput`/`Done` events as NDJSON, and supports an initial `--prompt` before entering the command loop. The `--mode` flag selects the same workflow/team/etc modes used by the one-shot runner. - Add `AgentMode` CLI enum (General, Research, Task, Team, Workflow) - Add `async-lock` for per-agent message serialization - Add unit tests for CLI parsing, state serde, and state listing Phase 3 stubs (status, pause, resume, policy) remain unimplemented. * feat(agent): implement status, pause, and resume commands Add `ClientCommand::Pause` and `Resume` to the background agent protocol. The server stores a shared `paused` flag and rejects new messages while paused, updating `agent.json` status accordingly. Implement `n00n agent status`, `pause`, and `resume` client commands. `status` reads the persisted agent state; `pause`/`resume` send a control command over the Unix socket. The `policy` subcommand remains a stub. * fix(agent): avoid duplicate text when message run ends with error The text-mode client was streaming TextDelta chunks and then printing the full Done.text again when the run ended with an error. Only print the error in that case; the streamed text is already on the terminal. * refactor(team): migrate roles and supervisor to n00n.subagent Extend `plugins/lib/n00n/subagent.lua` with options the team plugin needs: - `system` override for role-specific prompts - `preview` and `activity_label` for ActivityPreview integration - `budget` with `:consume()` for agent-call budgets - `fail_on_pricing_error` to keep the existing team semantics - return the resolved `model_spec` as a fifth value Migrate `plugins/team/roles.lua` to launch subagents through the shared helper, preserving its `{ok, text, cost, model, usage, error}` return shape and all existing tests. Simplify `plugins/team/init.lua` `run_supervisor` by delegating the structured-output plan generation to `n00n.subagent.launch` with `output_schema = PLANNER_OUTPUT`. * fix(agent): log cleanup warnings in stop handler Surface failed socket/directory removals in the background agent stop handler instead of silently ignoring them. The process still exits, but the warning gives operators a signal that stale state may remain. * fix(agent): validate agent ids and lock down socket permissions Prevent path traversal from user-supplied agent ids by validating them before any filesystem operation, and reject ids that contain path separators, control characters, or other unsafe content. Limit ids to 64 ASCII alphanumeric/hyphen/underscore characters. Set the agent state directory to 0o700 and the Unix control socket to 0o600 so only the owning user can read or connect to background agent state. * refactor(agent): share one-shot and background agent setup Extract `prepare_agent_env` and `PreparedEnv` to remove the duplicated plugin/model/MCP initialization between `run` and `server`. Both paths now build their `HeadlessParams`/`InteractiveParams` from the same prepared environment, making the two entry points easier to keep in sync. * docs(changelog): add fragments for agent CLI and team refactor Record the user-facing additions, refactor, and security hardening from PR #134 in changelog.d. * fix(agent): await task completion before stop exits The background agent Stop command now waits for the agent task to finish after sending the cancel signal, instead of exiting immediately. This prevents dropping active tool calls and unfinished MCP sessions and ensures state is persisted before cleanup. * refactor(cli): remove unused goal arg and stub policy commands The `goal` flag on `n00n agent run` was parsed but never used, and the `policy` subcommands were stubs. Remove both until they are wired to actual behavior. * fix(task,workflow): pass thinking config to subagent sessions Thread the `thinking` option through `n00n.agent.session` in both the task and workflow plugins, and include it in the workflow journal key so cached results honor the thinking setting. * fix(n00n-agent): make yolo enable rather than toggle in spawn_interactive `headless::spawn_interactive` was calling `permissions.toggle_yolo()` when `params.yolo` was true, which flipped an already-true yolo state back to false. This broke `--yolo` in background agent mode (and TUI/ACP when always_yolo was set). Add `PermissionManager::set_yolo` and use it to reliably enable yolo when requested, leaving the config-derived default otherwise. * feat(agent): add --goal and mode-aware team/workflow/task prompts Re-add --goal to n00n agent run and wire it into mode-specific prompts. Team mode now emits a team tool call with goal, mode, max_agents, waves, and auto-tier flags. Workflow mode emits a workflow tool call with the prompt as the script and goal merged into inputs. Task mode emits a task tool call with description/prompt. General/Research modes prepend the goal to the prompt. Also introduce AgentRunOptions to keep run/server signatures tidy and avoid excessive boolean parameters. * feat(agent): configurable agent-call limits and runaway guard Remove the hard-coded 24/32 agent-call ceilings in team and workflow. - team `max_agents` is now configurable with no hard maximum and an optional `timeout_secs`. - workflow `max_agents_per_run` is configurable with no hard maximum. - Add `plugins/lib/n00n/guard.lua` to enforce call budgets, wall-clock timeouts, repeated-prompt loops, and consecutive subagent errors. - Wire the guard through `n00n.subagent.launch`, `team`, and `workflow`. * fix(guard): reserve call budget atomically in check and guard consecutive errors - Move the used counter increment from record() into check() so the budget is reserved before any async yield; this prevents concurrent subagent calls from overshooting max_calls. - Check consecutive_errors in check() and use > in record() so the guard blocks the next call after the threshold instead of wasting a token on the failing call. - Add lib spec coverage for budget, repeated-prompt, consecutive-error, and legacy consume() behavior. * fix(agent): gate background agent IPC behind Unix-only cfg Background agents use Unix domain sockets and permission modes that are not available on Windows. Return a clear error on unsupported platforms so the crate builds on all CI targets. * fix(nix): use platform-correct loader env var in wrappers Use DYLD_LIBRARY_PATH on Darwin and LD_LIBRARY_PATH elsewhere, and wrap the computed package binary path to avoid runtime loader failures on non-Linux builds. * refactor(plugins): migrate task and workflow to n00n.subagent.launch for structured output * fix(agent): resolve P1 issues in background server and error handling - Fix background detach: fork and setsid to daemonize server process - Fix stop command: close input_tx and await outer task for proper cleanup - Fix error handling: break event loop on Error to prevent state hang - Fix task plugin: remove unused error from get_progress, fix encode error handling - Fix workflow plugin: emit agent_done only on success, remove unused cost/usage returns * fix(agent): address P2 review feedback in agent.rs - Return error on AgentEvent::Error in one-shot run for proper exit code - Preserve agent state on read failures (only delete on not found) - Avoid overwriting paused status when active run completes - UTF-8 truncation already uses character-aware .chars().take() * fix(team): enforce wall-clock timeout during subagent calls * feat(agent): wire research mode through system prompt and tool restrictions - Add AgentMode::Research and is_readonly() so build_system_prompt picks the research prompt and filter_tools_for_mode strips execute-kind tools. - Plumb runtime AgentMode through HeadlessParams/InteractiveParams and persist it in background AgentState so follow-up messages keep research mode. - Exclude write/edit/multiedit/edit_lines/insert_lines for research runs. - Fix task plugin preview to use ActivityPreview (required by n00n.subagent) and return raw structured-output validation errors for tests. * fix(workflow): preserve string subagent output and encode structured tables Subagent.launch may return either a plain string or a structured table. Store strings unchanged, JSON-encode tables with error handling, and use an empty string for nil output instead of encoding nil.
w0wl0lxd
added a commit
that referenced
this pull request
Jul 29, 2026
…rompts (#134) * feat(agent): add structured output helper and agent CLI command Phase 2 implementation: - Create plugins/lib/n00n/structured_output.lua with constants and helper functions for schema validation - Create plugins/lib/n00n/subagent.lua with unified subagent launch API - Migrate plugins/task/init.lua to use n00n.structured_output helper - Migrate plugins/workflow/init.lua to use n00n.structured_output helper - Add n00n agent run CLI command with stubs for future daemon commands - Add tests for structured_output helper in plugins/lib/tests/spec.lua - Keep team/roles.lua unchanged due to custom patterns (hardcoded audience, budget handling, preview support) All tests pass. Lua syntax validated with luac. * fix(agent-orchestration): repair route_tier and structured_output helpers * feat(agent): add background agent server and management CLI Implement `n00n agent run --background` Unix-socket server plus `message`, `stop`, and `list` client commands. Background agents keep state in `state_dir/agents/<id>/agent.json` and a `control.sock`. Server streams `TextDelta`/`ToolOutput`/`Done` events as NDJSON, and supports an initial `--prompt` before entering the command loop. The `--mode` flag selects the same workflow/team/etc modes used by the one-shot runner. - Add `AgentMode` CLI enum (General, Research, Task, Team, Workflow) - Add `async-lock` for per-agent message serialization - Add unit tests for CLI parsing, state serde, and state listing Phase 3 stubs (status, pause, resume, policy) remain unimplemented. * feat(agent): implement status, pause, and resume commands Add `ClientCommand::Pause` and `Resume` to the background agent protocol. The server stores a shared `paused` flag and rejects new messages while paused, updating `agent.json` status accordingly. Implement `n00n agent status`, `pause`, and `resume` client commands. `status` reads the persisted agent state; `pause`/`resume` send a control command over the Unix socket. The `policy` subcommand remains a stub. * fix(agent): avoid duplicate text when message run ends with error The text-mode client was streaming TextDelta chunks and then printing the full Done.text again when the run ended with an error. Only print the error in that case; the streamed text is already on the terminal. * refactor(team): migrate roles and supervisor to n00n.subagent Extend `plugins/lib/n00n/subagent.lua` with options the team plugin needs: - `system` override for role-specific prompts - `preview` and `activity_label` for ActivityPreview integration - `budget` with `:consume()` for agent-call budgets - `fail_on_pricing_error` to keep the existing team semantics - return the resolved `model_spec` as a fifth value Migrate `plugins/team/roles.lua` to launch subagents through the shared helper, preserving its `{ok, text, cost, model, usage, error}` return shape and all existing tests. Simplify `plugins/team/init.lua` `run_supervisor` by delegating the structured-output plan generation to `n00n.subagent.launch` with `output_schema = PLANNER_OUTPUT`. * fix(agent): log cleanup warnings in stop handler Surface failed socket/directory removals in the background agent stop handler instead of silently ignoring them. The process still exits, but the warning gives operators a signal that stale state may remain. * fix(agent): validate agent ids and lock down socket permissions Prevent path traversal from user-supplied agent ids by validating them before any filesystem operation, and reject ids that contain path separators, control characters, or other unsafe content. Limit ids to 64 ASCII alphanumeric/hyphen/underscore characters. Set the agent state directory to 0o700 and the Unix control socket to 0o600 so only the owning user can read or connect to background agent state. * refactor(agent): share one-shot and background agent setup Extract `prepare_agent_env` and `PreparedEnv` to remove the duplicated plugin/model/MCP initialization between `run` and `server`. Both paths now build their `HeadlessParams`/`InteractiveParams` from the same prepared environment, making the two entry points easier to keep in sync. * docs(changelog): add fragments for agent CLI and team refactor Record the user-facing additions, refactor, and security hardening from PR #134 in changelog.d. * fix(agent): await task completion before stop exits The background agent Stop command now waits for the agent task to finish after sending the cancel signal, instead of exiting immediately. This prevents dropping active tool calls and unfinished MCP sessions and ensures state is persisted before cleanup. * refactor(cli): remove unused goal arg and stub policy commands The `goal` flag on `n00n agent run` was parsed but never used, and the `policy` subcommands were stubs. Remove both until they are wired to actual behavior. * fix(task,workflow): pass thinking config to subagent sessions Thread the `thinking` option through `n00n.agent.session` in both the task and workflow plugins, and include it in the workflow journal key so cached results honor the thinking setting. * fix(n00n-agent): make yolo enable rather than toggle in spawn_interactive `headless::spawn_interactive` was calling `permissions.toggle_yolo()` when `params.yolo` was true, which flipped an already-true yolo state back to false. This broke `--yolo` in background agent mode (and TUI/ACP when always_yolo was set). Add `PermissionManager::set_yolo` and use it to reliably enable yolo when requested, leaving the config-derived default otherwise. * feat(agent): add --goal and mode-aware team/workflow/task prompts Re-add --goal to n00n agent run and wire it into mode-specific prompts. Team mode now emits a team tool call with goal, mode, max_agents, waves, and auto-tier flags. Workflow mode emits a workflow tool call with the prompt as the script and goal merged into inputs. Task mode emits a task tool call with description/prompt. General/Research modes prepend the goal to the prompt. Also introduce AgentRunOptions to keep run/server signatures tidy and avoid excessive boolean parameters. * feat(agent): configurable agent-call limits and runaway guard Remove the hard-coded 24/32 agent-call ceilings in team and workflow. - team `max_agents` is now configurable with no hard maximum and an optional `timeout_secs`. - workflow `max_agents_per_run` is configurable with no hard maximum. - Add `plugins/lib/n00n/guard.lua` to enforce call budgets, wall-clock timeouts, repeated-prompt loops, and consecutive subagent errors. - Wire the guard through `n00n.subagent.launch`, `team`, and `workflow`. * fix(guard): reserve call budget atomically in check and guard consecutive errors - Move the used counter increment from record() into check() so the budget is reserved before any async yield; this prevents concurrent subagent calls from overshooting max_calls. - Check consecutive_errors in check() and use > in record() so the guard blocks the next call after the threshold instead of wasting a token on the failing call. - Add lib spec coverage for budget, repeated-prompt, consecutive-error, and legacy consume() behavior. * fix(agent): gate background agent IPC behind Unix-only cfg Background agents use Unix domain sockets and permission modes that are not available on Windows. Return a clear error on unsupported platforms so the crate builds on all CI targets. * fix(nix): use platform-correct loader env var in wrappers Use DYLD_LIBRARY_PATH on Darwin and LD_LIBRARY_PATH elsewhere, and wrap the computed package binary path to avoid runtime loader failures on non-Linux builds. * refactor(plugins): migrate task and workflow to n00n.subagent.launch for structured output * fix(agent): resolve P1 issues in background server and error handling - Fix background detach: fork and setsid to daemonize server process - Fix stop command: close input_tx and await outer task for proper cleanup - Fix error handling: break event loop on Error to prevent state hang - Fix task plugin: remove unused error from get_progress, fix encode error handling - Fix workflow plugin: emit agent_done only on success, remove unused cost/usage returns * fix(agent): address P2 review feedback in agent.rs - Return error on AgentEvent::Error in one-shot run for proper exit code - Preserve agent state on read failures (only delete on not found) - Avoid overwriting paused status when active run completes - UTF-8 truncation already uses character-aware .chars().take() * fix(team): enforce wall-clock timeout during subagent calls * feat(agent): wire research mode through system prompt and tool restrictions - Add AgentMode::Research and is_readonly() so build_system_prompt picks the research prompt and filter_tools_for_mode strips execute-kind tools. - Plumb runtime AgentMode through HeadlessParams/InteractiveParams and persist it in background AgentState so follow-up messages keep research mode. - Exclude write/edit/multiedit/edit_lines/insert_lines for research runs. - Fix task plugin preview to use ActivityPreview (required by n00n.subagent) and return raw structured-output validation errors for tests. * fix(workflow): preserve string subagent output and encode structured tables Subagent.launch may return either a plain string or a structured table. Store strings unchanged, JSON-encode tables with error handling, and use an empty string for nil output instead of encoding nil.
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.
Summary
n00n agent run --backgroundUnix-socket server for long-lived background agents.n00n agent message,stop,list,status,pause, andresumeclient commands.state_dir/agents/<id>/agent.jsonwith acontrol.sock.TextDelta,ToolOutput,Done,Error) as NDJSON and supports pause/resume.AgentModeenum supportsGeneral,Research,Task,Team, andWorkflow.--goal,--team-mode,--max-agents,--waves,--workflow-inputs, and--task-descriptionton00n agent runso team/workflow/task modes build the matching tool call automatically.plugins/team/init.luaandplugins/team/roles.luato launch subagents through the sharedn00n.subagenthelper, reducing duplication withplugins/taskandplugins/workflow.plugins/lib/n00n/structured_output.luaandplugins/lib/n00n/subagent.luamodules used by task, workflow, and team plugins.agent_idto prevent path traversal, sets state directory permissions to0o700, sets the Unix control socket to0o600, and waits for the agent task to finish before thestopcommand exits.thinkingoption throughtaskandworkflowsubagent sessions.Test plan
cargo fmt --allcargo check --all --testscargo clippy --all --tests -- -D warningscargo nextest run --workspacecargo test -p n00ncargo test -p n00n-lua --test specNote
Medium Risk
Research mode and read-only tool filtering change which tools run in production sessions; background agent fork/cancel/stop paths touch long-lived process control and persisted state.
Overview
Introduces
AgentMode::Researchend-to-end: read-only tool filtering (same path as plan mode viais_readonly()),PromptId::Researchfor system prompts, session/UI persistence asStoredMode::Research, and CLIresearchmode that excludes write/edit tools when spawning headless or interactive agents.plugins/taskandplugins/workflownow calln00n.subagent.launchwhen anoutput_schemais set; task still uses a directn00n.agent.sessionpath for the legacydonetool. Task previews move toActivityPreview. Team setsctx:set_deadlinefromtimeout_secs.Background
n00n agent runserver behavior is tightened: child processfork/setsid,modestored inagent.jsonand restored for client messages, pause sends cancel, stop drops input and cancels the interactive task, and error events end the message loop. Headless CLI runs fail withErrwhen the agent emits an error event.Reviewed by Cursor Bugbot for commit cf2092b. Bugbot is set up for automated code reviews on this repo. Configure here.