v0.9.0

Added

• New researcher agent for external knowledge research — fetches and synthesizes information from the web, official docs, APIs, and public sources during the comprehension phase (before planning)
• New brainstorm agent — helps developers discover and articulate what they want to build before planning starts. Run it before Orion to produce a structured product brief at docs/briefs/{project-name}.md.
• New harness agent — encodes emerging patterns as permanent mechanical enforcement artifacts (lint rules, CI workflows, AGENTS.md entries, guiding principles). Triggered by the user, Orion post-feature, or the Gardener on recurring drift.
• New planning agent — transforms complex or ambiguous requests into structured work contracts on disk (docs/exec-plans/). Returns inline plan simples for small tasks; full exec-plans for multi-session work.
• New gardener agent — periodic maintenance agent that fixes stale documentation and detects code drift against established rules. Opens targeted PRs; updates QUALITY_SCORE.md; escalates recurring patterns to harness.
• Orion now knows when to invoke planning (complex/ambiguous requests) and when to suggest harness post-delivery (recurring patterns).
• Five lifecycle tools now available directly to Orion — no delegation needed for project bookkeeping: project_state (full artifact inventory), check_artifacts (consistency scan), mark_block_done (check a block in an exec-plan), complete_plan (close a scope), and register_spec (create a new spec file). Orion calls these at mission start and after each delivery automatically.
• Exec-plans now support an optional brief: frontmatter field to trace the brainstorm → implementation link bidirectionally.

Changed

• The brainstorm agent now challenges your assumptions before drafting — Phase 2 applies Socratic pressure on stated assumptions and constraints, and a mandatory adversarial gate runs before the brief is written: the agent presents the strongest case against building the product and asks what would cause it to fail. Briefs are stronger as a result.
• The brainstorm agent now hard-blocks on incomplete briefs — if the Problem statement, Success Criteria, or Scope In are missing or unresolved, the brief won't be drafted until those gaps are closed. Cosmetic disagreements are noted as open questions; substantive ones block the output entirely.
• Scope inflation is now flagged throughout the brainstorm session — if the in-scope list grows to 5 or more items, the agent surfaces it once and asks what's truly essential.
• Harness now operates fully autonomously — it explores the codebase, decides what to encode, and acts without asking for confirmation at each step. It only stops in three explicit cases: the pattern can't be mechanized, encoding requires creating a new workflow file, or the trigger is too vague with no codebase signal to anchor it.
• The soul personality layer now applies to all agents with mode: "all" (brainstorm, planning, bug-finder, harness, gardener) — previously only Orion benefited from it
• The bug-finder agent now includes a pattern assessment in every output — flags whether the bug is a systemic pattern and recommends invoking harness when the root cause can be mechanically encoded
• The review-manager can now read any file directly — it no longer needs to spawn an explore sub-agent to read source files before reviewing

Fixed

• Lifecycle tools (project_state, mark_block_done, complete_plan, register_spec, check_artifacts) now return valid responses — previously the execute functions returned raw objects instead of strings, causing the OpenCode plugin API to silently discard their output
• Harness agent now has full bash, read, write, edit, glob, and grep permissions — previously it was registered with a restricted command allowlist and scoped file targets, which prevented it from running arbitrary lint commands or writing enforcement artifacts outside the predefined list.
• The harness agent no longer writes human-facing checklists to AGENTS.md — it now correctly identifies them as documentation and routes them to CI checks or docs/guiding-principles.md instead. An unwired script in the repo is also no longer treated as a valid enforcement artifact.
• Brainstorm agent now enforces a hard stop before responding to the user — the docs/briefs/ scan is mandatory regardless of how much context the user provides at session start, preventing the agent from skipping existing brief detection
• Planning agent write/edit permissions now correctly allow files directly in docs/exec-plans/ (not just subdirectories)
• Planning agent can now read AGENTS.md, README.md, and docs/** — the "*": "deny" in the read sub-object was blocking all file reads
• Lifecycle tools (project_state, register_spec, mark_block_done, complete_plan, check_artifacts) now work correctly when OpenCode passes worktree="/" — the plugin falls back to directory instead of treating the filesystem root as the project root
• Planning agent can now read project files and create exec-plans — permission rules were blocking read access and missing glob/grep tools needed for codebase exploration
• Removed invalid write permission key from all agent configs — OpenCode's permission system uses edit to govern all file modifications (write, edit, patch); the separate write key was silently ignored, causing new file creation to be blocked by the top-level "*": "deny" rule

Removed

• memory.md concept removed — the persistent project memory feature has been deprecated. The experimental.chat.system.transform hook and memory.md injections have been removed from the plugin. Only the scratchpad survives compaction. See why persistent agent memory is an anti-pattern.

v0.8.0

Added

• New bug-finder agent — a structured bug investigation orchestrator that forces root-cause analysis before any fix is applied. Prevents Orion from rushing to workarounds that mask symptoms and create code divergence.
• Persistent memory across sessions — Orion now maintains .opencode/memory.md, a project-level knowledge base that accumulates architecture decisions, conventions, and user preferences. The plugin injects it automatically into every session via experimental.chat.system.transform, so it's available from the first message without any tool call.
• The default soul directives are now in English — previously the built-in personality guidelines were in French, which was unexpected for non-French speakers. Disable with soul: false if you prefer a neutral voice.
• Orion now prefers registered user-defined agents over invented personas — when a project defines domain-specific agents (e.g., languages/typescript-pro, mcp/mcp-developer, web/react-specialist), Orion selects them instead of defaulting to a general + invented persona name

Changed

• The review-manager is now significantly faster for trivial changes — it instantly approves docs-only or formatting updates without spawning sub-agents, and only spins up a single code-reviewer for low-risk tweaks.
• The reviewer delegation prompt has been drastically slimmed down — the review-manager no longer wastes tokens re-explaining focus, stance, or formatting to specialized reviewers that already know their job.

Fixed

• Reviewer agents (requirements, code, security) now have a hardcoded skeptical stance in their system prompts — this counteracts the default LLM approval bias where agents would spot real issues but rationalize them away instead of flagging them.

Removed

• sequential-thinking has been removed — modern models decompose complex workflows natively, making the explicit planning tool unnecessary friction
• The "What you MUST NOT do" tool list has been removed — the Cardinal Rule and Anti-Patterns section already cover this constraint more effectively
• The delegation prompt template (the 5-section ## Context / ## Task / ## Files / ## Constraints / ## Deliverable scaffold) has been removed — modern models structure delegations well without an explicit template, and the surrounding prose conveys the substance
• In-Flight Delegations tracking from the scratchpad template has been removed — the compaction hook already preserves the full scratchpad, making the urgent task_id recording instruction redundant
• The Self-Evaluation numbered checklist has been collapsed to prose — the core checks (original request coverage, multi-agent coherence, scope drift, side effects) are preserved but without the mechanical 6-item format
• The < 1-2 delegations away interruption threshold has been removed from the scope-switching protocol — the metric was unmeasurable and the principle (park state, switch, return) stands without it
• The "When NOT to Prune" subsection has been removed from the Context Management section — all three bullets described actions a model wouldn't take anyway
• The "Max Retries" column has been removed from the error handling retry table — the "2 total attempts → escalate" rule in prose is the one that matters, the per-cause counters were redundant and contradictory

v0.8.0-beta.8

Added

• New bug-finder agent — a structured bug investigation orchestrator that forces root-cause analysis before any fix is applied. Prevents Orion from rushing to workarounds that mask symptoms and create code divergence.
• Persistent memory across sessions — Orion now maintains .opencode/memory.md, a project-level knowledge base that accumulates architecture decisions, conventions, and user preferences. The plugin injects it automatically into every session via experimental.chat.system.transform, so it's available from the first message without any tool call.
• The default soul directives are now in English — previously the built-in personality guidelines were in French, which was unexpected for non-French speakers. Disable with soul: false if you prefer a neutral voice.
• Orion now prefers registered user-defined agents over invented personas — when a project defines domain-specific agents (e.g., languages/typescript-pro, mcp/mcp-developer, web/react-specialist), Orion selects them instead of defaulting to a general + invented persona name

Changed

• The review-manager is now significantly faster for trivial changes — it instantly approves docs-only or formatting updates without spawning sub-agents, and only spins up a single code-reviewer for low-risk tweaks.
• The reviewer delegation prompt has been drastically slimmed down — the review-manager no longer wastes tokens re-explaining focus, stance, or formatting to specialized reviewers that already know their job.

Fixed

• Reviewer agents (requirements, code, security) now have a hardcoded skeptical stance in their system prompts — this counteracts the default LLM approval bias where agents would spot real issues but rationalize them away instead of flagging them.

Removed

• sequential-thinking has been removed — modern models decompose complex workflows natively, making the explicit planning tool unnecessary friction
• The "What you MUST NOT do" tool list has been removed — the Cardinal Rule and Anti-Patterns section already cover this constraint more effectively
• The delegation prompt template (the 5-section ## Context / ## Task / ## Files / ## Constraints / ## Deliverable scaffold) has been removed — modern models structure delegations well without an explicit template, and the surrounding prose conveys the substance
• In-Flight Delegations tracking from the scratchpad template has been removed — the compaction hook already preserves the full scratchpad, making the urgent task_id recording instruction redundant
• The Self-Evaluation numbered checklist has been collapsed to prose — the core checks (original request coverage, multi-agent coherence, scope drift, side effects) are preserved but without the mechanical 6-item format
• The < 1-2 delegations away interruption threshold has been removed from the scope-switching protocol — the metric was unmeasurable and the principle (park state, switch, return) stands without it
• The "When NOT to Prune" subsection has been removed from the Context Management section — all three bullets described actions a model wouldn't take anyway
• The "Max Retries" column has been removed from the error handling retry table — the "2 total attempts → escalate" rule in prose is the one that matters, the per-cause counters were redundant and contradictory

v0.8.0-beta.7

Added

• New bug-finder agent — a structured bug investigation orchestrator that forces root-cause analysis before any fix is applied. Prevents Orion from rushing to workarounds that mask symptoms and create code divergence.
• Persistent memory across sessions — Orion now maintains .opencode/memory.md, a project-level knowledge base that accumulates architecture decisions, conventions, and user preferences. The plugin injects it automatically into every session via experimental.chat.system.transform, so it's available from the first message without any tool call.
• The default soul directives are now in English — previously the built-in personality guidelines were in French, which was unexpected for non-French speakers. Disable with soul: false if you prefer a neutral voice.

Changed

• The review-manager is now significantly faster for trivial changes — it instantly approves docs-only or formatting updates without spawning sub-agents, and only spins up a single code-reviewer for low-risk tweaks.
• The reviewer delegation prompt has been drastically slimmed down — the review-manager no longer wastes tokens re-explaining focus, stance, or formatting to specialized reviewers that already know their job.

Fixed

• Reviewer agents (requirements, code, security) now have a hardcoded skeptical stance in their system prompts — this counteracts the default LLM approval bias where agents would spot real issues but rationalize them away instead of flagging them.

Removed

• sequential-thinking has been removed — modern models decompose complex workflows natively, making the explicit planning tool unnecessary friction
• The "What you MUST NOT do" tool list has been removed — the Cardinal Rule and Anti-Patterns section already cover this constraint more effectively
• The delegation prompt template (the 5-section ## Context / ## Task / ## Files / ## Constraints / ## Deliverable scaffold) has been removed — modern models structure delegations well without an explicit template, and the surrounding prose conveys the substance
• In-Flight Delegations tracking from the scratchpad template has been removed — the compaction hook already preserves the full scratchpad, making the urgent task_id recording instruction redundant
• The Self-Evaluation numbered checklist has been collapsed to prose — the core checks (original request coverage, multi-agent coherence, scope drift, side effects) are preserved but without the mechanical 6-item format
• The < 1-2 delegations away interruption threshold has been removed from the scope-switching protocol — the metric was unmeasurable and the principle (park state, switch, return) stands without it
• The "When NOT to Prune" subsection has been removed from the Context Management section — all three bullets described actions a model wouldn't take anyway
• The "Max Retries" column has been removed from the error handling retry table — the "2 total attempts → escalate" rule in prose is the one that matters, the per-cause counters were redundant and contradictory

v0.8.0-beta.6

Added

• New bug-finder agent — a structured bug investigation orchestrator that forces root-cause analysis before any fix is applied. Prevents Orion from rushing to workarounds that mask symptoms and create code divergence.
• Persistent memory across sessions — Orion now maintains .opencode/memory.md, a project-level knowledge base that accumulates architecture decisions, conventions, and user preferences. The plugin injects it automatically into every session via experimental.chat.system.transform, so it's available from the first message without any tool call.
• The default soul directives are now in English — previously the built-in personality guidelines were in French, which was unexpected for non-French speakers. Disable with soul: false if you prefer a neutral voice.

Changed

• The review-manager is now significantly faster for trivial changes — it instantly approves docs-only or formatting updates without spawning sub-agents, and only spins up a single code-reviewer for low-risk tweaks.
• The reviewer delegation prompt has been drastically slimmed down — the review-manager no longer wastes tokens re-explaining focus, stance, or formatting to specialized reviewers that already know their job.

Fixed

• Reviewer agents (requirements, code, security) now have a hardcoded skeptical stance in their system prompts — this counteracts the default LLM approval bias where agents would spot real issues but rationalize them away instead of flagging them.

Removed

• sequential-thinking has been removed — modern models decompose complex workflows natively, making the explicit planning tool unnecessary friction
• The "What you MUST NOT do" tool list has been removed — the Cardinal Rule and Anti-Patterns section already cover this constraint more effectively
• The delegation prompt template (the 5-section ## Context / ## Task / ## Files / ## Constraints / ## Deliverable scaffold) has been removed — modern models structure delegations well without an explicit template, and the surrounding prose conveys the substance
• In-Flight Delegations tracking from the scratchpad template has been removed — the compaction hook already preserves the full scratchpad, making the urgent task_id recording instruction redundant
• The Self-Evaluation numbered checklist has been collapsed to prose — the core checks (original request coverage, multi-agent coherence, scope drift, side effects) are preserved but without the mechanical 6-item format
• The < 1-2 delegations away interruption threshold has been removed from the scope-switching protocol — the metric was unmeasurable and the principle (park state, switch, return) stands without it
• The "When NOT to Prune" subsection has been removed from the Context Management section — all three bullets described actions a model wouldn't take anyway
• The "Max Retries" column has been removed from the error handling retry table — the "2 total attempts → escalate" rule in prose is the one that matters, the per-cause counters were redundant and contradictory

v0.8.0-beta.5

Added

• New bug-finder agent — a structured bug investigation orchestrator that forces root-cause analysis before any fix is applied. Prevents Orion from rushing to workarounds that mask symptoms and create code divergence.
• Persistent memory across sessions — Orion now maintains .opencode/memory.md, a project-level knowledge base that accumulates architecture decisions, conventions, and user preferences. The plugin injects it automatically into every session via experimental.chat.system.transform, so it's available from the first message without any tool call.
• The default soul directives are now in English — previously the built-in personality guidelines were in French, which was unexpected for non-French speakers. Disable with soul: false if you prefer a neutral voice.

Removed

• sequential-thinking has been removed — modern models decompose complex workflows natively, making the explicit planning tool unnecessary friction
• The "What you MUST NOT do" tool list has been removed — the Cardinal Rule and Anti-Patterns section already cover this constraint more effectively
• The delegation prompt template (the 5-section ## Context / ## Task / ## Files / ## Constraints / ## Deliverable scaffold) has been removed — modern models structure delegations well without an explicit template, and the surrounding prose conveys the substance
• In-Flight Delegations tracking from the scratchpad template has been removed — the compaction hook already preserves the full scratchpad, making the urgent task_id recording instruction redundant
• The Self-Evaluation numbered checklist has been collapsed to prose — the core checks (original request coverage, multi-agent coherence, scope drift, side effects) are preserved but without the mechanical 6-item format
• The < 1-2 delegations away interruption threshold has been removed from the scope-switching protocol — the metric was unmeasurable and the principle (park state, switch, return) stands without it
• The "When NOT to Prune" subsection has been removed from the Context Management section — all three bullets described actions a model wouldn't take anyway
• The "Max Retries" column has been removed from the error handling retry table — the "2 total attempts → escalate" rule in prose is the one that matters, the per-cause counters were redundant and contradictory

v0.8.0-beta.4

Added

• New bug-finder agent — a structured bug investigation orchestrator that forces root-cause analysis before any fix is applied. Prevents Orion from rushing to workarounds that mask symptoms and create code divergence.
• Persistent memory across sessions — Orion now maintains .opencode/memory.md, a project-level knowledge base that accumulates architecture decisions, conventions, and user preferences. The plugin injects it automatically into every session via experimental.chat.system.transform, so it's available from the first message without any tool call.
• The default soul directives are now in English — previously the built-in personality guidelines were in French, which was unexpected for non-French speakers. Disable with soul: false if you prefer a neutral voice.

Removed

• sequential-thinking has been removed — modern models decompose complex workflows natively, making the explicit planning tool unnecessary friction
• The "What you MUST NOT do" tool list has been removed — the Cardinal Rule and Anti-Patterns section already cover this constraint more effectively
• The delegation prompt template (the 5-section ## Context / ## Task / ## Files / ## Constraints / ## Deliverable scaffold) has been removed — modern models structure delegations well without an explicit template, and the surrounding prose conveys the substance
• In-Flight Delegations tracking from the scratchpad template has been removed — the compaction hook already preserves the full scratchpad, making the urgent task_id recording instruction redundant
• The Self-Evaluation numbered checklist has been collapsed to prose — the core checks (original request coverage, multi-agent coherence, scope drift, side effects) are preserved but without the mechanical 6-item format
• The < 1-2 delegations away interruption threshold has been removed from the scope-switching protocol — the metric was unmeasurable and the principle (park state, switch, return) stands without it
• The "When NOT to Prune" subsection has been removed from the Context Management section — all three bullets described actions a model wouldn't take anyway
• The "Max Retries" column has been removed from the error handling retry table — the "2 total attempts → escalate" rule in prose is the one that matters, the per-cause counters were redundant and contradictory

v0.8.0-beta.3

Added

• Persistent memory across sessions — Orion now maintains .opencode/memory.md, a project-level knowledge base that accumulates architecture decisions, conventions, and user preferences. The plugin injects it automatically into every session via experimental.chat.system.transform, so it's available from the first message without any tool call.
• The default soul directives are now in English — previously the built-in personality guidelines were in French, which was unexpected for non-French speakers. Disable with soul: false if you prefer a neutral voice.

Removed

• sequential-thinking has been removed — modern models decompose complex workflows natively, making the explicit planning tool unnecessary friction
• The "What you MUST NOT do" tool list has been removed — the Cardinal Rule and Anti-Patterns section already cover this constraint more effectively
• The delegation prompt template (the 5-section ## Context / ## Task / ## Files / ## Constraints / ## Deliverable scaffold) has been removed — modern models structure delegations well without an explicit template, and the surrounding prose conveys the substance
• In-Flight Delegations tracking from the scratchpad template has been removed — the compaction hook already preserves the full scratchpad, making the urgent task_id recording instruction redundant
• The Self-Evaluation numbered checklist has been collapsed to prose — the core checks (original request coverage, multi-agent coherence, scope drift, side effects) are preserved but without the mechanical 6-item format
• The < 1-2 delegations away interruption threshold has been removed from the scope-switching protocol — the metric was unmeasurable and the principle (park state, switch, return) stands without it
• The "When NOT to Prune" subsection has been removed from the Context Management section — all three bullets described actions a model wouldn't take anyway
• The "Max Retries" column has been removed from the error handling retry table — the "2 total attempts → escalate" rule in prose is the one that matters, the per-cause counters were redundant and contradictory

v0.8.0-beta.2

Added

• Persistent memory across sessions — Orion now maintains .opencode/memory.md, a project-level knowledge base that accumulates architecture decisions, conventions, and user preferences. The plugin injects it automatically into every session via experimental.chat.system.transform, so it's available from the first message without any tool call.

Removed

• sequential-thinking has been removed — modern models decompose complex workflows natively, making the explicit planning tool unnecessary friction
• The "What you MUST NOT do" tool list has been removed — the Cardinal Rule and Anti-Patterns section already cover this constraint more effectively
• The delegation prompt template (the 5-section ## Context / ## Task / ## Files / ## Constraints / ## Deliverable scaffold) has been removed — modern models structure delegations well without an explicit template, and the surrounding prose conveys the substance
• In-Flight Delegations tracking from the scratchpad template has been removed — the compaction hook already preserves the full scratchpad, making the urgent task_id recording instruction redundant
• The Self-Evaluation numbered checklist has been collapsed to prose — the core checks (original request coverage, multi-agent coherence, scope drift, side effects) are preserved but without the mechanical 6-item format
• The < 1-2 delegations away interruption threshold has been removed from the scope-switching protocol — the metric was unmeasurable and the principle (park state, switch, return) stands without it
• The "When NOT to Prune" subsection has been removed from the Context Management section — all three bullets described actions a model wouldn't take anyway
• The "Max Retries" column has been removed from the error handling retry table — the "2 total attempts → escalate" rule in prose is the one that matters, the per-cause counters were redundant and contradictory

v0.8.0-beta.1

Removed

• sequential-thinking has been removed — modern models decompose complex workflows natively, making the explicit planning tool unnecessary friction
• The "What you MUST NOT do" tool list has been removed — the Cardinal Rule and Anti-Patterns section already cover this constraint more effectively
• The delegation prompt template (the 5-section ## Context / ## Task / ## Files / ## Constraints / ## Deliverable scaffold) has been removed — modern models structure delegations well without an explicit template, and the surrounding prose conveys the substance
• In-Flight Delegations tracking from the scratchpad template has been removed — the compaction hook already preserves the full scratchpad, making the urgent task_id recording instruction redundant
• The Self-Evaluation numbered checklist has been collapsed to prose — the core checks (original request coverage, multi-agent coherence, scope drift, side effects) are preserved but without the mechanical 6-item format
• The < 1-2 delegations away interruption threshold has been removed from the scope-switching protocol — the metric was unmeasurable and the principle (park state, switch, return) stands without it
• The "When NOT to Prune" subsection has been removed from the Context Management section — all three bullets described actions a model wouldn't take anyway
• The "Max Retries" column has been removed from the error handling retry table — the "2 total attempts → escalate" rule in prose is the one that matters, the per-cause counters were redundant and contradictory