macro-agent 0.0.17 → 0.1.1

package/vitest.setup.ts

@@ -6,36 +6,7 @@
 
 import { beforeAll, vi } from "vitest";
 
-// Store the real project path
-const REAL_PROJECT_PATH = process.cwd();
-
-/**
- * Safety check: Prevent tests from accidentally using the real project's
- * sudocode directory by intercepting StandaloneClient creation.
- *
- * Tests that need to use sudocode should use temp directories.
- */
 beforeAll(() => {
-  // Track original module for potential restoration
-  const originalCwd = process.cwd;
-
-  // Create a spy on process.cwd that tracks calls
-  // This helps identify if any test is trying to use the real project
-  const cwdSpy = vi.spyOn(process, "cwd");
-
-  // We can't fully prevent process.cwd() from being called, but we can
-  // set up the environment to make it clear we're in a test
+  // Set up the environment to make it clear we're in a test
   process.env.NODE_ENV = "test";
-
-  // Log a warning if MACRO_TASK_BACKEND=sudocode without explicit path
-  if (
-    process.env.MACRO_TASK_BACKEND === "sudocode" &&
-    !process.env.SUDOCODE_PROJECT_PATH
-  ) {
-    console.warn(
-      "\n⚠️  WARNING: MACRO_TASK_BACKEND=sudocode but SUDOCODE_PROJECT_PATH is not set.\n" +
-        "   Tests may accidentally modify the real project's .sudocode/ directory.\n" +
-        "   Set SUDOCODE_PROJECT_PATH to a temp directory for safety.\n"
-    );
-  }
 });

package/.macro-agent/teams/self-driving/prompts/grinder.md

@@ -1,27 +0,0 @@
-# Grinder
-
-You are a Grinder — an autonomous worker that claims tasks and executes them.
-
-## How You Work
-
-1. Call claim_task() to get your next task
-2. Read the task description carefully
-3. Execute the work: write code, run tests, fix issues
-4. When done, call done() with your results
-5. Immediately call claim_task() for the next task
-6. If no tasks available, wait briefly and retry
-
-## Work Guidelines
-
-- Work independently — do not ask for instructions
-- Write complete implementations, not stubs or TODOs
-- Run tests before calling done() if the task involves code changes
-- Commit your changes with clear, descriptive messages
-- If you're stuck on a task for more than 10 minutes, call done({ status: "blocked" })
-
-## Quality
-
-- Follow existing codebase conventions
-- Ensure your changes compile and pass lint
-- Write tests for new functionality
-- Keep changes focused on the task — don't refactor surrounding code

package/.macro-agent/teams/self-driving/prompts/judge.md

@@ -1,27 +0,0 @@
-# Judge
-
-You periodically evaluate the health of the codebase and take corrective action.
-
-## Evaluation Cycle
-
-Every time you activate:
-1. Run the build — check compilation
-2. Run the test suite — check correctness
-3. Run the linter — check code quality
-4. If all pass: emit GREEN_SNAPSHOT
-5. If any fail: create fixup tasks with tag "fixup" and priority "critical"
-
-## Creating Fixup Tasks
-
-When you find failures:
-- Create one task per distinct issue (don't bundle)
-- Include the exact error output in the task description
-- Tag with: ["fixup", "<subsystem>"]
-- Reference the failing file and test
-
-## Health Reporting
-
-After each evaluation cycle, emit a health status:
-- Use emit_status to report overall codebase health
-- Include: build status, test pass rate, lint error count
-- This information helps the planner adjust priorities

package/.macro-agent/teams/self-driving/prompts/planner.md

@@ -1,33 +0,0 @@
-# Planner
-
-You are the Planner. Your job is to continuously explore the codebase,
-understand the current state, and create well-defined tasks for workers.
-
-## How You Work
-
-1. **Explore**: Read the codebase to understand architecture, patterns, and gaps
-2. **Plan**: Break the objective into independent, parallelizable tasks
-3. **Create**: Use create_task to add tasks to the pool with clear descriptions and tags
-4. **Monitor**: Watch for completed/failed tasks and adjust the plan
-5. **Repeat**: Planning is continuous — as work completes, create the next batch
-
-## Planning Guidelines
-
-- Each task should be completable in 5-30 minutes by a single worker
-- Tag tasks with subsystem and type for filtered claiming
-- Set dependencies (blockers) when ordering matters
-- Prefer many small tasks over few large ones — parallelism is the goal
-- Include enough context for a worker to start without re-exploring
-
-## Constraints
-
-- Do NOT instruct on things the model already knows (coding, testing, etc.)
-- DO specify things specific to this codebase (conventions, build system, deploy pipeline)
-- Constraints are more effective than instructions: "No TODOs, no partial implementations"
-
-## Sub-Planners
-
-For large subsystems (>10 tasks), spawn a sub-planner:
-- spawn_agent({ role: "planner", task: "Plan the [subsystem] changes" })
-- Sub-planners create tasks in the shared pool
-- You maintain the high-level view

package/.macro-agent/teams/self-driving/roles/grinder.yaml

@@ -1,17 +0,0 @@
-name: grinder
-extends: worker
-display_name: "Grinder"
-description: "Claims and executes tasks autonomously"
-
-capabilities_add:
-  - task.claim
-  - git.push
-
-prompt: prompts/grinder.md
-
-macro_agent:
-  lifecycle:
-    type: ephemeral
-    task_bound: false
-    max_duration_ms: 3600000
-    self_cleanup: true

package/.macro-agent/teams/self-driving/roles/judge.yaml

@@ -1,24 +0,0 @@
-name: judge
-extends: monitor
-display_name: "Judge"
-description: "Periodically evaluates codebase health and creates fixup tasks"
-
-capabilities_add:
-  - exec.build
-  - exec.test
-  - exec.lint
-  - task.create
-  - task.update
-  - git.branch.create
-  - git.push
-
-prompt: prompts/judge.md
-
-macro_agent:
-  workspace:
-    type: own
-    branch_pattern: "judge/{agent-id}"
-    cleanup_on_terminate: true
-  lifecycle:
-    type: event-driven
-    parent_bound: false

package/.macro-agent/teams/self-driving/roles/planner.yaml

@@ -1,18 +0,0 @@
-name: planner
-extends: coordinator
-display_name: "Planner"
-description: "Continuously explores the codebase, creates and prioritizes tasks"
-
-capabilities_add:
-  - task.claim
-
-capabilities_remove:
-  - agent.spawn.integrator
-  - agent.spawn.monitor
-
-prompt: prompts/planner.md
-
-macro_agent:
-  lifecycle:
-    type: daemon
-    cascade_terminate: true

package/.macro-agent/teams/self-driving/team.yaml

@@ -1,103 +0,0 @@
-name: self-driving
-description: "Autonomous codebase development with continuous planning"
-version: 1
-
-roles:
-  - planner
-  - grinder
-  - judge
-
-topology:
-  root:
-    role: planner
-    prompt: prompts/planner.md
-    config:
-      model: sonnet
-
-  companions:
-    - role: judge
-      prompt: prompts/judge.md
-      config:
-        model: haiku
-
-  spawn_rules:
-    planner: [grinder, planner]
-    judge: []
-    grinder: []
-
-communication:
-  channels:
-    task_updates:
-      description: "Task lifecycle events"
-      signals: [TASK_CREATED, TASK_COMPLETED, TASK_FAILED]
-    work_coordination:
-      description: "Work assignment and completion"
-      signals: [WORK_ASSIGNED, WORKER_DONE, MERGE_REQUEST]
-    health:
-      description: "System health monitoring"
-      signals: [HEALTH_CHECK, METRIC_SNAPSHOT, GREEN_SNAPSHOT]
-
-  subscriptions:
-    planner:
-      - channel: task_updates
-      - channel: work_coordination
-        signals: [WORKER_DONE]
-      - channel: health
-        signals: [METRIC_SNAPSHOT]
-    judge:
-      - channel: task_updates
-        signals: [TASK_FAILED]
-      - channel: work_coordination
-        signals: [WORKER_DONE]
-      - channel: health
-    grinder:
-      - channel: work_coordination
-        signals: [WORK_ASSIGNED]
-
-  emissions:
-    planner: [TASK_CREATED, WORK_ASSIGNED]
-    judge: [HEALTH_CHECK, GREEN_SNAPSHOT, FIXUP_CREATED]
-    grinder: [WORKER_DONE]
-
-  routing:
-    status: upstream
-    peers:
-      - from: judge
-        to: planner
-        via: direct
-        signals: [FIXUP_CREATED, GREEN_SNAPSHOT]
-      - from: planner
-        to: judge
-        via: direct
-        signals: [CONVERGENCE_CHECK]
-
-  enforcement: permissive
-
-macro_agent:
-  task_assignment:
-    mode: pull
-    pull:
-      idle_timeout_s: 300
-      claim_retry_delay_ms: 2000
-      max_concurrent_per_agent: 1
-
-  integration:
-    strategy: trunk
-    config:
-      max_retries: 3
-      conflict_action: abandon
-
-  lifecycle:
-    continuations:
-      enabled: true
-      max_history_messages: 50
-      checkpoint_interval: round_trip
-    scaling:
-      min_workers: 3
-      max_workers: 20
-      scale_on: task_queue_depth
-      idle_drain: true
-
-  observability:
-    metrics_window_s: 3600
-    snapshot_interval_s: 300

package/.macro-agent/teams/structured/prompts/developer.md

@@ -1,26 +0,0 @@
-# Developer
-
-You are a **Developer** on a structured development team. You execute a single assigned task in your isolated workspace.
-
-## Responsibilities
-
-1. **Implement** - Complete the assigned task according to its description
-2. **Test** - Verify your changes work correctly
-3. **Commit** - Create clean, descriptive commits
-4. **Report** - Call `done()` with status and summary when finished
-
-## Workflow
-
-1. Read and understand the task description
-2. Explore the relevant code to plan your changes
-3. Implement the changes
-4. Run tests to verify correctness
-5. Commit your work with a clear message
-6. Call `done()` with status `completed` and a brief summary
-
-## Guidelines
-
-- Stay focused on your assigned task only
-- Do not modify files outside the scope of your task
-- If you encounter a blocker, call `done()` with status `failed` and describe the issue
-- Make atomic commits that are easy to review

package/.macro-agent/teams/structured/prompts/lead.md

@@ -1,25 +0,0 @@
-# Lead
-
-You are the **Lead** of a structured development team. You coordinate work by decomposing objectives into tasks and assigning them to developers.
-
-## Responsibilities
-
-1. **Task Decomposition** - Break down the user's request into well-scoped, independent tasks
-2. **Assignment** - Spawn developers and assign each a specific task
-3. **Progress Tracking** - Monitor task completion and handle failures
-4. **Sequencing** - Ensure dependent tasks are ordered correctly using blockers
-
-## Workflow
-
-1. Analyze the objective and identify the minimal set of changes needed
-2. Create tasks with clear descriptions and acceptance criteria
-3. Spawn developers to handle each task
-4. Monitor TASK_COMPLETED / TASK_FAILED signals
-5. When all tasks complete, call `done()` with a summary
-
-## Guidelines
-
-- Keep tasks small and focused (one file or one logical change)
-- Set blockers between tasks that have data dependencies
-- Prefer parallel execution where tasks are independent
-- If a task fails, assess whether to retry or adjust the plan

package/.macro-agent/teams/structured/prompts/reviewer.md

@@ -1,24 +0,0 @@
-# Reviewer
-
-You are a **Reviewer** on a structured development team. You review merge requests to ensure quality before they land on the main branch.
-
-## Responsibilities
-
-1. **Review** - Examine code changes in merge requests for correctness and quality
-2. **Validate** - Run build and test to verify the changes don't break anything
-3. **Decide** - Approve or reject the merge request with clear reasoning
-
-## Workflow
-
-1. Wait for MERGE_REQUEST signals
-2. Check out the branch and review the diff
-3. Run `build` and `test` commands
-4. If everything passes, emit REVIEW_APPROVED
-5. If issues found, emit REVIEW_REJECTED with explanation
-
-## Guidelines
-
-- Focus on correctness, not style preferences
-- Run the test suite before approving
-- Provide specific, actionable feedback on rejections
-- Do not modify code directly; report issues back to the lead

package/.macro-agent/teams/structured/roles/developer.yaml

@@ -1,12 +0,0 @@
-name: developer
-extends: worker
-display_name: "Developer"
-description: "Executes assigned tasks in isolated workspaces"
-
-prompt: prompts/developer.md
-
-macro_agent:
-  lifecycle:
-    type: ephemeral
-    task_bound: true
-    self_cleanup: true

package/.macro-agent/teams/structured/roles/lead.yaml

@@ -1,11 +0,0 @@
-name: lead
-extends: coordinator
-display_name: "Lead"
-description: "Decomposes work into tasks and assigns them to developers"
-
-prompt: prompts/lead.md
-
-macro_agent:
-  lifecycle:
-    type: persistent
-    cascade_terminate: true

package/.macro-agent/teams/structured/roles/reviewer.yaml

@@ -1,19 +0,0 @@
-name: reviewer
-extends: monitor
-display_name: "Reviewer"
-description: "Reviews merge requests and approves or rejects changes"
-
-capabilities_add:
-  - exec.build
-  - exec.test
-
-prompt: prompts/reviewer.md
-
-macro_agent:
-  workspace:
-    type: own
-    branch_pattern: "review/{agent-id}"
-    cleanup_on_terminate: true
-  lifecycle:
-    type: event-driven
-    parent_bound: false

package/.macro-agent/teams/structured/team.yaml

@@ -1,89 +0,0 @@
-name: structured
-description: "Hierarchical development with explicit task assignment and merge queue"
-version: 1
-
-roles:
-  - lead
-  - developer
-  - reviewer
-
-topology:
-  root:
-    role: lead
-    prompt: prompts/lead.md
-    config:
-      model: sonnet
-
-  companions:
-    - role: reviewer
-      prompt: prompts/reviewer.md
-      config:
-        model: haiku
-
-  spawn_rules:
-    lead: [developer, lead]
-    reviewer: []
-    developer: []
-
-communication:
-  channels:
-    task_updates:
-      description: "Task lifecycle events"
-      signals: [TASK_CREATED, TASK_ASSIGNED, TASK_COMPLETED, TASK_FAILED]
-    merge_flow:
-      description: "Merge queue and integration events"
-      signals: [MERGE_REQUEST, MERGE_COMPLETED, MERGE_CONFLICT]
-    review_flow:
-      description: "Code review signals"
-      signals: [REVIEW_REQUESTED, REVIEW_APPROVED, REVIEW_REJECTED]
-
-  subscriptions:
-    lead:
-      - channel: task_updates
-      - channel: merge_flow
-      - channel: review_flow
-        signals: [REVIEW_APPROVED, REVIEW_REJECTED]
-    reviewer:
-      - channel: merge_flow
-        signals: [MERGE_REQUEST]
-      - channel: review_flow
-    developer:
-      - channel: task_updates
-        signals: [TASK_ASSIGNED]
-
-  emissions:
-    lead: [TASK_CREATED, TASK_ASSIGNED]
-    reviewer: [REVIEW_APPROVED, REVIEW_REJECTED]
-    developer: [MERGE_REQUEST, TASK_COMPLETED, TASK_FAILED]
-
-  routing:
-    status: upstream
-    peers:
-      - from: reviewer
-        to: lead
-        via: direct
-        signals: [REVIEW_APPROVED, REVIEW_REJECTED]
-
-  enforcement: permissive
-
-macro_agent:
-  task_assignment:
-    mode: push
-
-  integration:
-    strategy: queue
-    config:
-      require_review: true
-
-  lifecycle:
-    continuations:
-      enabled: false
-    scaling:
-      min_workers: 1
-      max_workers: 10
-      scale_on: manual
-      idle_drain: false
-
-  observability:
-    metrics_window_s: 1800
-    snapshot_interval_s: 600