v0.6.2 — Scoped research and product discovery

Supervisor mode now supports bounded discovery work

@supervisor can now handle more than concrete execution tasks. In v0.6.2, it supports bounded research, comparison, recommendation, and MVP-scoping prompts.

Examples:

@supervisor Compare Sentry, Honeycomb, and Datadog for frontend error monitoring and recommend one.

@supervisor Turn this customer problem into an MVP product scope with user stories and acceptance criteria.

What changed

• broadened supervisor planning heuristics for discovery-style prompts
• replaced naive comma-fragment planning with synthesis-oriented work-unit generation
• improved mixed execution + discovery handling so concrete steps are preserved
• preserved secondary asks in MVP-scoping prompts
• kept open-ended brainstorms and long-horizon strategy requests unsupported

Still intentionally unsupported

Examples that should still fail closed with rewrite guidance:

• @supervisor brainstorm startup ideas
• @supervisor research the entire CRM market
• @supervisor define our 12-month strategy

Validation

• 50 test files / 476 tests passing
• typecheck clean
• local refresh + verify: 61/61 OK
• runtime smoke tests previously validated on the branch before release

Full changelog

See CHANGELOG.md

v0.6.1 — npm CLI packaging fix

Patch release

v0.6.1 fixes npm CLI packaging compatibility for the published package.

What's fixed

• normalize the published bin path for npm v11 compatibility
• ensure the published CLI package exposes the correct executable metadata
• keep the runtime/plugin behavior identical to v0.6.0

User impact

• no plugin/runtime behavior changes versus v0.6.0
• patch is only about the published CLI packaging surface

Install

npm install -g opencode-council@0.6.1
opencode-council --version

npm

opencode-council@0.6.1 is published on npm.

v0.6.0 — Supervisor Mode + Budget Profiles

Supervisor Mode

Type @supervisor [goal] and the plugin decomposes, plans, and launches real child sessions.

@supervisor Build an auth module, refactor the API layer, update docs

• Plugin detects @supervisor trigger
• Goal decomposed into lanes with roles and dependencies
• Plan preview shown in output
• LLM calls supervisor_launch tool to start child sessions
• Each child session runs as a separate OpenCode session

Budget Profiles

Choose how aggressively the plugin manages token budgets:

npx opencode-council config budget-profile extended

Profile	Description
conservative	Tight budgets, early compaction (CI/automation)
standard	Balanced budgets, moderate compaction (default)
extended	Higher budgets, late compaction (research/strategy)
unlimited	No proactive compaction, warnings only

Interactive selection during npx opencode-council init, or set via CLI / env var.

Performance

Supervisor modules are lazy-loaded — non-supervisor sessions load ~15 modules at startup instead of ~35.

Install

npx opencode-council init

Or without Node.js:

curl -fsSL https://raw.githubusercontent.com/marcel-tuinstra/opencode-council/v0.6.0/install.sh | bash

Validation

• 50 test files / 454 tests passing
• Typecheck clean
• OpenCode smoke tests: single role, multi-role, delegation, supervisor — all green

Full changelog

See CHANGELOG.md

v0.5.0 — Contract Freeze & Hardening

Contract Freeze

The stable package root is now intentionally small and frozen for 0.5.x:

import { AgentConversations, SUPPORTED_ROLES } from "opencode-council";

Supervisor helpers live behind an explicit experimental entry point:

import { ... } from "opencode-council/supervisor";

Compatibility Policy

A canonical compatibility and deprecation guide is now shipped with the package. Stable surfaces for 0.5.x:

• Root package exports
• CLI verbs: init, refresh, verify, uninstall
• Documented budget env vars
• Documented supervisor-policy.json keys

See docs/guides/compatibility-and-deprecations.md

Observability Hardening

• Structured diagnostics envelopes for runtime debug events
• Correlation metadata (sessionId, runId, laneId) where available
• Clearer reason-code and remediation surfacing for budget, MCP, and policy paths
• Budget runtime config provenance (default / policy / env)

Runtime Fix

The plugin entry file now exports only the plugin factory function. The npm package barrel is a separate index.ts. This fixes a runtime crash where OpenCode's loader tried to call an exported array as a plugin factory.

Install

npx opencode-council init
# or
curl -fsSL https://raw.githubusercontent.com/marcel-tuinstra/opencode-council/v0.5.0/install.sh | bash

Validation

• 43 test files / 213 tests passing
• Typecheck clean
• Local install refresh + verify: 54/54 OK
• OpenCode runtime smoke tests: plugin loads, roles detect, system injects, no errors

Full changelog

See CHANGELOG.md

v0.4.0 — OpenCode Council

Rename to OpenCode Council

This release renames the project from opencode-orchestration-workflows to opencode-council. The GitHub repo has been renamed (old URLs redirect automatically).

One-command install

Node.js:

npx opencode-council init

Without Node.js:

curl -fsSL https://raw.githubusercontent.com/marcel-tuinstra/opencode-council/v0.4.0/install.sh | bash

Install management

npx opencode-council refresh     # Update to latest version
npx opencode-council verify      # SHA-256 health check
npx opencode-council uninstall   # Clean removal

All commands support --dry-run, --backup, and --force flags.

What's new

• CLI installer (bin/cli.mjs): zero-dependency Node.js CLI with init, refresh, verify, uninstall commands
• Shell installer (install.sh): POSIX shell installer for non-Node users, pinned to release tag
• Config detection: existing install detection, supervisor-policy.json preservation
• Stale file pruning: refresh/init cleans up removed files from previous versions
• engines.node >= 22.0.0: matches CI test matrix

Full changelog

See CHANGELOG.md

v0.3.0-beta

v0.3.0-beta

DelegationPlan: lead-first orchestration

• Introduce DelegationPlan type for modeling explicit agent-led delegation as a first-class runtime object
• Wire delegation plan through all plugin hooks: message transform, system transform, and text complete
• Inject delegation-lead system instructions when delegation mode is active (single-role prompt tells the agent it may delegate via <<DELEGATE:...>> markers)
• Parse max parallel agents N from user prompts

Visible delegated launches

• Rewrite generic orchestrator narration into structured [Supervisor] delegation.launch lines
• Render wave-based delegation provenance (delegated wave 1 by CTO: FE, BE, UX)
• Render max-parallel metadata (max parallel agents: 6)

CLI control leakage stripping

• Strip interactive CLI hints (ctrl+x down view subagents) from delegated transcripts

Known limitation

• The DelegationPlan is built and stored in session policy, but the LLM does not yet reliably emit <<DELEGATE:...>> markers to trigger actual subagent launches. The plan is passive metadata until the execution bridge is built.

Test coverage

• 3 new acceptance tests for delegated launch transcript rendering
• 197/197 tests green, no regressions

v0.2.3-beta

v0.2.3-beta

Delegation provenance tracking

• Detect explicit agent-led delegation requests (let @ceo delegate, have @ceo decide, use @cto first then let it pull in others)
• Persist delegation metadata in session policy (DelegationRequest with mode, primaryRole, requestedByUser)
• Carry delegation provenance through delegated thread expansion so transcripts show user-requested and agent-delegated roles separately
• Pass requestedByUser provenance to both delegated-thread and multi-role-thread paths for consistency
• Omit empty addedByOrchestrator output instead of rendering none

Test coverage

• New acceptance test suite for delegation intent detection, provenance tracking, and transcript annotation
• Additional output tests for empty orchestrator provenance and multi-role-thread requested-by-user rendering
• Strengthened domain intent assertions

Refs #57

v0.2.2-beta

fix: harden manager role delegation

Changes

• Broaden manager-role delegation guards beyond CTO to cover CEO, PM, and other non-execution roles
• Reduce false positives by distinguishing real implementation ownership from review, planning, and product responsibilities
• Expand regression coverage so manager-role plans fail only when they actually absorb implementation work

Before

• v0.2.1-beta closed the immediate CTO takeover bug, but responsibility detection was still heuristic and broad enough to over-flag review or planning language
• Broader manager-role coverage for CEO and PM was not yet enforced explicitly

After

• Manager-role delegation hardening now applies across the non-execution leadership roles
• Review, planning, requirements, and roadmap responsibilities no longer trip the implementation-owner guard by accident
• Regression tests now cover both rejection and valid non-implementation manager assignments

Example

• A plan where PM owns Build the release checklist flow now fails, while a plan where PM owns Deliver roadmap and Review test plan remains valid

v0.2.1-beta: Manager Delegation Boundary Fix

Highlights

This patch fixes a delegation-boundary gap discovered after the Supervisor Beta promotion.

The issue: a manager role such as CTO could still absorb implementation responsibilities in a delegated plan without handing execution to DEV, FE, BE, or UX.

This release reached main via PR #55 (fix/manager-delegation-boundaries -> main, merge commit ad9a2fc).

What's Included

• Regression coverage for manager roles absorbing implementation work
• Validation that rejects implementation-scoped delegation plans owned directly by manager roles
• Validation that requires at least one execution-role assignment (DEV, FE, BE, or UX) for implementation-scoped runs
• Positive coverage proving that CTO remains valid when architecture/review stays with CTO and execution is delegated

Validation

• npm run typecheck passed
• npm test -- tests/supervisor-delegation.test.ts tests/supervisor-config.test.ts tests/roles.test.ts passed
• GitHub CI passed on PR #55

Known Limits

• Responsibility detection is still heuristic and text-based
• Broader manager-boundary hardening for CEO, PM, and related non-execution roles continues in the next patch line

What's Next

• Tighten manager-role responsibility detection beyond simple keyword heuristics
• Add broader regression coverage for CEO, PM, and other non-execution roles
• Continue hardening delegation-boundary policy for the 0.2.2-beta patch

v0.2.0-beta: Supervisor Beta

Highlights

Supervisor Beta promotes the full supervisor workflow to main.

This release moves beyond Alpha foundations and ships a coherent end-to-end supervisor stack: goal intake, lane decomposition, dispatch and scheduler flow, governance and review routing, protected-path enforcement, data lifecycle reporting, execution workflow, golden traces, and specialist FE/BE/UX role routing.

It reached main via PR #54 (beta -> main, merge commit f7eadd5).

What's Included

• Goal planning and lane decomposition for delegated work
• Lane contract, handoff validation, and checkpoint escalation routing
• Supervisor dispatch planning, scheduler flow, and execution workflow
• Delegate-only supervisor guardrails with explicit integration-agent support
• Governance review routing and protected-path policy enforcement
• Guardrail threshold evidence and observability-aligned review support
• Supervisor bootstrap preview and dry-run data lifecycle reporting
• Golden trace scenario coverage for release-readiness validation
• Specialist role routing for DEV, FE, BE, and UX
• Operator docs for local sync, role sanity testing, and worktree discipline

Validation

• npm run typecheck passed
• npm test passed (38 files, 171 tests)
• GitHub CI passed on the promotion PR for Node 22 and 24
• Local plugin and agent sync completed from main

Known Limits

• Supervisor runtime remains conservative and manual-merge-first
• Specialist role sanity is documented as an operator check, not yet a CI gate
• Local installs still require syncing both plugin files and agent profiles after updates

What's Next

• Keep tightening specialist role boundary checks and regression coverage
• Harden local install and smoke-test flows further
• Continue evolving the supervisor from safe Beta baseline toward stronger post-Beta orchestration