opencode-agile-agent 1.0.4 → 1.2.1

package/bin/validate-templates.js

@@ -10,59 +10,57 @@ const templatesDir = join(__dirname, '..', 'templates');
 
 // Required files for templates
 const requiredFiles = [
-  'AGENTS.template.md',
+  'opencode.json',
+  '.opencode/.gitignore',
   '.opencode/README.md',
   '.opencode/ARCHITECTURE.md',
   '.opencode/archive/README.md',
+  '.opencode/artifacts/README.md',
   '.opencode/config.template.json',
   '.opencode/package.json',
-  '.opencode/memory/project.md',
+  '.opencode/plugins/session-artifacts.ts',
+  '.opencode/agents/retrospective-writer.md',
+  '.opencode/agents/archiver.md',
   '.opencode/agents/context-gatherer.md',
-  '.opencode/agents/api-designer.md',
   '.opencode/agents/backend-specialist.md',
-  '.opencode/agents/code-archaeologist.md',
-  '.opencode/agents/database-architect.md',
   '.opencode/agents/debugger.md',
   '.opencode/agents/developer.md',
   '.opencode/agents/devops-engineer.md',
-  '.opencode/agents/documentation-writer.md',
-  '.opencode/agents/explorer-agent.md',
   '.opencode/agents/feature-lead.md',
   '.opencode/agents/frontend-specialist.md',
-  '.opencode/agents/game-developer.md',
-  '.opencode/agents/mobile-developer.md',
-  '.opencode/agents/orchestrator.md',
-  '.opencode/agents/penetration-tester.md',
   '.opencode/agents/performance-optimizer.md',
   '.opencode/agents/pr-reviewer.md',
-  '.opencode/agents/product-manager.md',
   '.opencode/agents/project-planner.md',
-  '.opencode/agents/qa-automation-engineer.md',
   '.opencode/agents/security-auditor.md',
-  '.opencode/agents/seo-specialist.md',
   '.opencode/agents/system-analyst.md',
   '.opencode/agents/test-engineer.md',
   '.opencode/commands/brainstorm.md',
+  '.opencode/commands/archive.md',
+  '.opencode/commands/assign-models.md',
+  '.opencode/commands/check-progress.md',
   '.opencode/commands/create.md',
-  '.opencode/commands/debug.md',
   '.opencode/commands/plan.md',
+  '.opencode/commands/reframe.md',
   '.opencode/commands/review.md',
+  '.opencode/commands/rubber-duck.md',
   '.opencode/commands/status.md',
   '.opencode/commands/test.md',
   '.opencode/skills/api-patterns/SKILL.md',
+  '.opencode/skills/archive-writing/SKILL.md',
+  '.opencode/skills/artifact-discipline/SKILL.md',
   '.opencode/skills/brainstorming/SKILL.md',
+  '.opencode/skills/clarify-first/SKILL.md',
   '.opencode/skills/clean-code/SKILL.md',
   '.opencode/skills/code-philosophy/SKILL.md',
   '.opencode/skills/context-archive/SKILL.md',
   '.opencode/skills/context-gathering/SKILL.md',
-  '.opencode/skills/frontend-design/SKILL.md',
-  '.opencode/skills/intelligent-routing/SKILL.md',
-  '.opencode/skills/parallel-agents/SKILL.md',
+'.opencode/skills/intelligent-routing/SKILL.md',
   '.opencode/skills/plan-writing/SKILL.md',
-  '.opencode/skills/redteam-validation/SKILL.md',
   '.opencode/skills/security-gate/SKILL.md',
   '.opencode/skills/systematic-debugging/SKILL.md',
   '.opencode/skills/testing-patterns/SKILL.md',
+  '.opencode/templates/review-summary.template.md',
+  '.opencode/templates/session-summary.template.md',
   '.opencode/rules/coding-standards.md',
   '.opencode/rules/git-conventions.md',
 ];

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "opencode-agile-agent",
-  "version": "1.0.4",
+  "version": "1.2.1",
   "description": "AI Agile Spec-Driven Development - Install .opencode agile agent system into any project",
   "keywords": [
     "ai",

package/templates/.opencode/ARCHITECTURE.md

@@ -1,82 +1,112 @@
-# OpenCode Agent Kit - Architecture
-
-## Overview
+# OpenCode Agent Kit — Architecture
 
-- 26 agents
-- 25 agents
-- 14 skills
-- 7 commands
-- Primary agent: feature-lead
-- All other agents are subagents and communicate through @ handoffs
+## Overview
+- 15 agents, 19 skills, 11 commands, 1 runtime plugin
+- Primary: feature-lead
+- All others are subagents communicating through @ handoffs
+- `session-artifacts` is the runtime spine for active feature state and handoff safety
 
 ## Request Lifecycle
-
 1. User request lands with @feature-lead.
-2. @context-gatherer maps the current project and active work.
-3. Discovery and planning create the compact context bundle.
-4. Build agents implement the approved spec.
-5. Quality agents review, test, and harden the change.
-6. Security-sensitive changes pass through @security-auditor and @penetration-tester before approval.
-7. Any failure loops back to the owning agent.
-8. Approved feature bundles are archived under .opencode/archive/<feature-slug>/.
-9. feature-lead synthesizes the final outcome.
-
-## Context Bundle
+2. @feature-lead **clarifies first** — iterates questions until scope, intent, and business value are clear. If user has no idea, offers default + trade-offs.
+3. `session_artifact_current` restores active feature state.
+4. @context-gatherer maps current project and active work.
+5. If new feature: create context bundle + status.yaml. If continuing: refresh the active artifact.
+6. Discovery and planning produce the compact context bundle.
+7. Build agents implement the approved spec from canonical handoff packets.
+8. Quality agents review, test, and harden.
+9. Security-sensitive changes pass through @security-auditor.
+10. Any failure loops back to the owning agent.
+11. When the flow may compact before the final gates, @retrospective-writer captures a compact checkpoint.
+12. After a meaningful failure is resolved, the agent may ask whether the lesson should become a reusable skill or rule.
+13. Approved work summaries archived under `.opencode/archive/<feature-slug>.md` through the artifact finalizer.
+14. @archiver writes the final archive record.
+15. feature-lead synthesizes the final outcome.
 
-- proposal.md: why, value, scope
-- goal.md: target outcome, constraints, default choice
-- spec.md: contract, data flow, edge cases, risks
+## Context Bundle
+Create only for new features. Skip if no active feature and user is continuing work.
+- brief.md: why, outcome, business context, scope, constraints, default choice
+- spec.md: contract, data flow, edge cases, risks, acceptance criteria
 - task.md: ordered checklist, dependencies, owners
-- important.md: facts, blockers, links, decisions
+- notes.md: facts, decisions, blockers, links
+- status.yaml: live execution state
 
-## Archive
+`status.yaml.status` allowed: `brainstorm`, `planning`, `implementation`, `verification`, `review`, `done`, `blocked`.
 
-- Archive completed bundles in `.opencode/archive/<feature-slug>/`.
-- Keep archive entries compact and aligned with the approved bundle.
+## Archive
+- Archive in `.opencode/archive/<feature-slug>.md`.
+- Include the shipped summary, changed surfaces, verification signal, and remaining follow-up.
+- Finalize only when `status.yaml` is `done`.
+- Finalize only from the primary agent.
 
 ## Skill Design
-
 - Skills are small, philosophy-first mental models.
-- Skill descriptions should state the trigger and the decision domain.
+- Skill descriptions state the trigger and decision domain.
 - Prefer the smallest skill that changes the next decision.
-
-## Role Map
-
+- `artifact-discipline` keeps runtime state out of prompt memory.
+- `session-closeout` leaves a durable checkpoint before compaction or pause.
+- Use `cross-model-regression` when risky changes need multiple specialist checks across different reviewer models.
+- Promote lessons from repeated or costly failures only when they create durable guidance.
+
+## Runtime Spine
+- Active work lives in `.opencode/artifacts/features/<feature-slug>/`.
+- `.opencode/artifacts/` is local runtime state and should be git-ignored.
+- Planning docs and runtime evidence live side-by-side but stay separate.
+- Every subagent should start from `session_artifact_handoff` instead of a document dump.
+- `session_artifact_repo_delta` guards against drift between artifact state and git state.
+- Archive remains the immutable shipped summary.
+
+## Flow Rules
+- The default command spine is `/brainstorm -> /plan -> /create -> /test -> /review -> /archive`.
+- The flow is resumable and non-linear. Any command can be re-entered when the work needs that lens.
+- Re-plan on scope drift, re-brainstorm on ambiguity, re-test on risky changes, and re-review at meaningful checkpoints.
+- Use @retrospective-writer when a failure may deserve a reusable lesson.
+
+## Role Map
+
 | Stage | Agents | Responsibility |
 |-------|--------|----------------|
-| Primary | feature-lead | Owns the request, tradeoffs, and the final gate loop |
-| Context | context-gatherer | Map the current project state and archive-ready history |
-| Coordination | project-planner, explorer-agent | Split, map, and sequence the work |
-| Advanced | orchestrator | Optional multi-domain synthesis pass |
-| Product | product-manager | Define problem, value, priorities, and release scope |
-| Spec | system-analyst, api-designer, database-architect | Create the compact context bundle and contracts |
-| Build | developer, frontend-specialist, backend-specialist, mobile-developer, game-developer, devops-engineer | Implement the approved spec |
-| Quality | test-engineer, security-auditor, penetration-tester, performance-optimizer, debugger, pr-reviewer | Verify, review, and harden |
-| Automation | qa-automation-engineer | Build and maintain test harnesses and repeatable validation |
-| Output | documentation-writer, seo-specialist | Explain and publish clearly |
-| Maintenance | code-archaeologist | Clean and simplify legacy code |
-
+| Primary | feature-lead | Own request, tradeoffs, final gate loop |
+| Archive | archiver | Write compact archive summaries for approved work |
+| Learning | retrospective-writer | Capture a reusable failure lesson and decide whether it should become a skill or rule |
+| Context | context-gatherer | Map project state and archive-ready history |
+| Coordination | project-planner | Split, map, sequence work |
+| Spec | system-analyst | Create compact context bundle and contracts |
+| Build | developer, frontend-specialist, backend-specialist, devops-engineer | Implement approved spec |
+| Quality | test-engineer, security-auditor, performance-optimizer, debugger, pr-reviewer | Verify, review, harden |
+
 ## Command Map
 
 | Command | Entry | Exit |
 |---------|-------|------|
-| brainstorm | @feature-lead receives the request. | One recommended path is clear. |
-| create | @feature-lead receives the feature request. | Implementation matches the spec. |
-| debug | @feature-lead scopes the bug and the impact. | The original failure no longer reproduces. |
-| plan | @feature-lead receives the request. | The planning bundle is short and complete. |
-| review | @feature-lead hands the change to @pr-reviewer. | Either APPROVED or CHANGES REQUESTED is clear. |
-| status | @feature-lead asks for the current project state. | The current health is easy to understand. |
-| test | @feature-lead defines the test scope. | The important paths are covered. |
-
-## Gate Rules
-
-- Do not hand off to build until the spec bundle is clear.
-- Do not hand off to review until implementation is complete.
-- Do not ship until the review gate is clean.
-- Loop back to the owning agent when a gate fails.
-
-## Decision Rule
-
-- Prefer defaults when the tradeoff is low risk.
-- Escalate to the user only when scope, security, or architecture changes materially.
-- Keep all context compact enough for a subagent to finish without asking again.
+| archive | @archiver closes approved work | Archive path and saved summary are explicit |
+| brainstorm | @feature-lead receives request or ambiguity returns | One recommended path is clear |
+| check-progress | @feature-lead needs a git-aware status read | Current state and next step are explicit |
+| create | @feature-lead receives feature request | Implementation matches spec |
+| plan | @feature-lead receives request or scope drift | Planning bundle is short and complete |
+| review | @feature-lead hands change to @pr-reviewer at any meaningful checkpoint | APPROVED or CHANGES REQUESTED |
+| reframe | @feature-lead detects a framing mismatch | Better direction and next step are explicit |
+| rubber-duck | @feature-lead needs to reason before changing code | Assumptions and next check are explicit |
+| status | @feature-lead asks for project state | Current health is easy to understand |
+| test | @feature-lead defines test scope | Important paths are covered |
+
+## Gate Rules
+- Do not hand off to build until spec bundle is clear.
+- Do not hand off to review until implementation is complete.
+- Do not ship until review gate is clean.
+- Do not archive until repo delta is reconciled.
+- Loop back to owning agent when a gate fails.
+
+## Decision Rules
+- Prefer defaults when tradeoff is low risk AND business intent is clear.
+- Escalate to user when scope, security, or architecture changes materially.
+- Keep context compact enough for a subagent to finish without asking again.
+- Pass intent in prompts and retrieve documents through tools.
+
+## Clarification Rules
+- **Iterate until clear**: Keep asking until scope, intent, and business value are unambiguous.
+- **No hallucination**: If you don't know, ask. Never assume user intent.
+- **Default + trade-offs**: If user has no idea, propose a default path with trade-offs. Let them choose.
+- **Business before technical**: Capture who, what value, and success criteria before any technical planning.
+- **State assumptions**: Flag what you inferred and what is unverified.
+- **Compact but complete**: Keep context bundle output compact, but never skip user requirements.

package/templates/.opencode/README.md

@@ -1,14 +1,42 @@
-# OpenCode Agent Kit
-
-Spec-driven multi-agent system for OpenCode.
-
----
-
+# OpenCode Agent Kit
+
+Spec-driven multi-agent system for OpenCode.
+
+---
+
 ## Included
 
-- 25 agents
-- 14 skills
-- 7 commands
+- 15 agents
+- 15 skills
+- 11 commands
+- 1 runtime plugin
+
+## Plannotator (Optional)
+
+- This kit can integrate with Plannotator for browser-based plan review.
+- If your project has `opencode.json` with `"plugin": ["@plannotator/opencode@latest"]`, the `submit_plan` tool becomes available (by default for OpenCode's built-in `plan` agent).
+- The template `opencode.json` included with this repo configures Plannotator in `plan-agent` mode and allows `feature-lead` to call `submit_plan` during `/plan`.
+- Optional slash commands (installs globally):
+  - macOS/Linux/WSL: `curl -fsSL https://plannotator.ai/install.sh | bash`
+  - Windows PowerShell: `irm https://plannotator.ai/install.ps1 | iex`
+
+## Runtime Spine
+
+- `session-artifacts` keeps live feature state in `.opencode/artifacts/features/<feature-slug>/`.
+- `.opencode/artifacts/` is local runtime state and should stay out of git.
+- Agents should retrieve active state through artifact tools instead of carrying full bundle context.
+- Archive writes a concise summary record to `.opencode/archive/<feature-slug>.md`, and finalization runs through the artifact plugin.
+
+## Flow Shape
+
+- The default spine is `/brainstorm -> /plan -> /create -> /test -> /review -> /archive`.
+- The flow is not rigid. Commands can be reused mid-stream when the work needs them.
+- Common examples:
+  - Use `/brainstorm` again during implementation when a requirement turns ambiguous.
+  - Use `/plan` again when scope changes materially.
+  - Use `/test` before `/create` is fully done when a risky slice needs early proof.
+  - Use `/review` on an intermediate slice before the full feature is complete.
+  - Use `/archive` to preserve one finished slice even if the larger feature continues.
 
 ## Skill Design
 
@@ -19,92 +47,116 @@ Spec-driven multi-agent system for OpenCode.
 ## Context Skills
 
 - `context-gathering` maps the current project and active work.
-- `context-archive` stores completed bundles in `.opencode/archive/<feature-slug>/`.
+- `context-archive` stores completed work summaries in `.opencode/archive/<feature-slug>.md`.
+- `archive-writing` turns approved work into a compact archive note.
+- `cross-model-regression` coordinates multi-agent regression sweeps across different reviewer models.
+- `failure-learning` turns verified mistakes into reusable skills or rules when the lesson is worth keeping.
 - `security-gate` decides when a change needs a security gate or redteam phase.
 - `redteam-validation` simulates attacker behavior and proves exploitability.
-- `orchestrator` is optional; default routing stays with `feature-lead` and the owning specialists.
 
 ## Commands
 
 - Custom slash commands live in `.opencode/commands/`.
 - Each command file uses Markdown frontmatter plus a prompt body, matching OpenCode's command format.
-- The current command set is `brainstorm`, `create`, `debug`, `plan`, `review`, `status`, and `test`.
+- The current command set is `archive`, `assign-models`, `brainstorm`, `check-progress`, `create`, `plan`, `reframe`, `review`, `rubber-duck`, `status`, and `test`.
 
 ## Primary Flow
 
 1. @feature-lead receives the request.
-2. @context-gatherer maps the current project state.
-3. @project-planner and @system-analyst create the compact bundle: proposal.md, goal.md, spec.md, task.md, and important.md.
-4. @developer implements the approved spec.
-5. @test-engineer, @security-auditor, @penetration-tester, and @pr-reviewer close the loop.
-6. @feature-lead archives the completed bundle.
-
+2. `session_artifact_current` restores the live feature state.
+3. @context-gatherer maps the current project state.
+4. @project-planner and @system-analyst create or refresh the compact bundle: brief.md, spec.md, task.md, notes.md, and status.yaml.
+5. @developer implements the approved spec from a canonical handoff packet.
+6. @test-engineer, @security-auditor, and @pr-reviewer close the loop.
+7. When a failure exposed a reusable lesson, ask whether to promote it into a skill or rule.
+8. If the session may compact before the final gate, @retrospective-writer captures a checkpoint.
+9. @archiver writes the completed work summary through the artifact plugin.
+
 ## Context Bundle
 
-- proposal.md: why, value, scope
-- goal.md: target outcome, constraints, default choice
-- spec.md: contract, data flow, edge cases, risks
+- Templates live in `.opencode/templates/` for quick bundle creation.
+- brief.md: why, outcome, scope, constraints, default choice
+- spec.md: contract, data flow, edge cases, risks, acceptance criteria
 - task.md: ordered checklist, dependencies, owners
-- important.md: facts, blockers, links, decisions
+- notes.md: facts, decisions, blockers, links
+- status.yaml: live execution state
+
+- `status.yaml` is the live execution artifact; the markdown files stay as stable planning/reference context.
+- `status.yaml.status` allowed values: `brainstorm`, `planning`, `implementation`, `verification`, `review`, `done`, `blocked`.
 
 ## Archive
 
-- Completed bundles live in `.opencode/archive/<feature-slug>/`.
+- Completed work summaries live in `.opencode/archive/<feature-slug>.md`.
+- Live artifacts in `.opencode/artifacts/` should not be committed.
 - Keep the archive copy approved, compact, and read-only in practice.
-
+- Archive only when `status.yaml` is `done`.
+- Archive the outcome, not the full live bundle.
+- Only the archive flow should finalize the archive summary.
+
 ## Agent Groups
 
-| Group | Agents |
-|-------|--------|
-| Context | `context-gatherer` |
-| Primary | `feature-lead` |
-| Coordination | `project-planner`, `explorer-agent` |
-| Advanced | `orchestrator` |
-| Product | `product-manager` |
-| Spec | `system-analyst`, `api-designer`, `database-architect` |
-| Build | `developer`, `frontend-specialist`, `backend-specialist`, `mobile-developer`, `game-developer`, `devops-engineer` |
-| Quality | `test-engineer`, `security-auditor`, `penetration-tester`, `performance-optimizer`, `debugger`, `pr-reviewer` |
-| Automation | `qa-automation-engineer` |
-| Output | `documentation-writer`, `seo-specialist` |
-| Maintenance | `code-archaeologist` |
-
+| Group        | Agents                                                                                                            |
+| ------------ | ----------------------------------------------------------------------------------------------------------------- |
+| Archive      | `archiver`                                                                                                        |
+| Context      | `context-gatherer`                                                                                                |
+| Primary      | `feature-lead`                                                                                                    |
+| Coordination | `project-planner`                                                                                                 |
+| Spec         | `system-analyst`                                                                                                  |
+| Build        | `developer`, `frontend-specialist`, `backend-specialist`, `devops-engineer`                                      |
+| Quality      | `test-engineer`, `security-auditor`, `performance-optimizer`, `debugger`, `pr-reviewer`                         |
+| Learning     | `retrospective-writer`                                                                                            |
+
 ## Command Catalog
 
-| Command | Use When |
-|---------|----------|
-| brainstorm | Explore options and clarify requirements before planning. |
-| create | Build new features, components, or project slices with a spec-driven flow. |
-| debug | Diagnose and fix bugs using a root-cause approach. |
-| plan | Create structured task breakdowns and spec artifacts. |
-| review | Review code against the spec, standards, and quality gates. |
-| status | Check project health, docs, and operating readiness. |
-| test | Run and generate tests for the codebase. |
-
+| Command        | Use When                                                                                       |
+| -------------- | ---------------------------------------------------------------------------------------------- |
+| archive        | Save an approved work summary into `.opencode/archive/<feature-slug>.md` for future reference. |
+| assign-models  | Map available models to each agent, confirm the routing, and update the active OpenCode config. |
+| brainstorm     | Explore options and clarify requirements before planning, or reopen ambiguity mid-flow.        |
+| check-progress | Check current status, visible changes, remaining work, and the next best step.                |
+| create         | Build new features, components, or project slices with a spec-driven flow.                    |
+| plan           | Create or refresh structured task breakdowns and spec artifacts.                              |
+| reframe        | Reset the framing when the output is off-target, unclear, or stuck repeating the same mistake. |
+| review         | Review code against the spec, standards, and quality gates at any meaningful checkpoint.      |
+| rubber-duck    | Think out loud, challenge assumptions, and isolate the real problem before changing anything. |
+| status         | Check project health, docs, and operating readiness.                                          |
+| test           | Run and generate tests for the codebase whenever proof is needed.                             |
+
 ## Skills
 
 - context-gathering
 - context-archive
+- archive-writing
+- cross-model-regression
+- artifact-discipline
 - clean-code
 - brainstorming
+- clarify-first
 - plan-writing
-- parallel-agents
-- intelligent-routing
-- frontend-design
-- api-patterns
+- intelligent-routing
+- frontend-design
+- api-patterns
+- session-closeout
 - testing-patterns
 - systematic-debugging
 - code-philosophy
 - security-gate
 - redteam-validation
-
-## Routing Examples
-
-- Add JWT auth -> @feature-lead -> @system-analyst -> @backend-specialist + @security-auditor + @penetration-tester -> @test-engineer
-- Fix a UI bug -> @feature-lead -> @frontend-specialist -> @test-engineer -> @pr-reviewer
+- failure-learning
+
+## Repo Delta Guard
+
+- `session_artifact_repo_delta` compares artifact-tracked files with actual git working tree changes.
+- Use it in `status`, `review`, and `archive` to catch drift before summaries or approval claims go stale.
+
+## Routing Examples
+
+- Add JWT auth -> @feature-lead -> @system-analyst -> @backend-specialist + @security-auditor -> @test-engineer
+- Fix a UI bug -> @feature-lead -> @frontend-specialist -> @test-engineer -> @pr-reviewer
 - Ship a multi-domain feature -> @feature-lead -> @project-planner -> specialist agents -> @pr-reviewer
-
-## Rules
-
-- Use the safest default when the downside is small.
-- Ask only when scope, security, or architecture changes materially.
-- Keep every handoff compact and explicit.
+
+## Rules
+
+- Use the safest default when the downside is small.
+- Ask only when scope, security, or architecture changes materially.
+- Keep every handoff compact and explicit.

package/templates/.opencode/agents/archiver.md

@@ -0,0 +1,45 @@
+---
+name: archiver
+description: Convert approved work into a compact archive summary with clear outcome, evidence, and follow-up.
+mode: subagent
+temperature: 0.1
+top_p: 0.82
+steps: 30
+permission:
+  task:
+    "*": deny
+tools:
+  read: true
+  grep: true
+  glob: true
+  write: false
+  edit: false
+skills:
+- context-archive
+- archive-writing
+- artifact-discipline
+---
+
+# Archiver
+
+## Role
+Turn finished work into a durable archive summary. Keep it compact, specific, and useful to the next human or agent.
+
+## @ Awareness
+- @feature-lead → archive gate owner
+- @pr-reviewer → review evidence or safety concerns
+- @test-engineer → verification evidence or coverage gaps
+- @system-analyst → scope or decision record needs clarification
+
+## Working Loop
+1. Start from `session_artifact_current`.
+2. Pull `session_artifact_review_packet`, `session_artifact_changed_files`, and `session_artifact_repo_delta` before writing.
+3. If artifact state and git state disagree, call out the mismatch before summarizing.
+4. Summarize what shipped, what changed, what was verified, and what follow-up remains.
+5. Hand the final summary back to the caller for `session_artifact_finalize`.
+
+## Guardrails
+- Do not archive before approval.
+- Do not copy the full live bundle into the archive summary.
+- Do not hide unresolved follow-up or verification gaps.
+- Prefer concise, evidence-backed summaries over narrative history.

package/templates/.opencode/agents/backend-specialist.md

@@ -1,46 +1,43 @@
----
-name: backend-specialist
-description: Subagent for APIs, services, business logic, and integration boundaries.
-mode: subagent
-tools:
-  read: true
-  grep: true
-  glob: true
-  bash: true
-  write: true
-  edit: true
-skills:
-  - clean-code
-  - api-patterns
-  - code-philosophy
-  - testing-patterns
----
-
-# Backend Specialist
-
-## Role
-- Build explicit APIs and service logic.
-- Validate inputs and keep contracts predictable.
-
-## @ Awareness
-- Call @database-architect for schema or migration changes.
-- Call @security-auditor for auth, permissions, or data exposure.
-- Call @test-engineer for contract and integration coverage.
-
-## Context Bundle
-- proposal.md: why, value, scope
-- goal.md: target outcome, constraints, default choice
-- spec.md: contract, data flow, edge cases, risks
-- task.md: ordered checklist, dependencies, owners
-- important.md: facts, blockers, links, decisions
-
-## Working Loop
-1. Read the assigned context.
-2. Solve the local problem in your domain.
-3. Expose tradeoffs and the recommended default.
-4. Hand off to the next owning agent.
-5. Stop when the exit gate is satisfied.
-
-## Guardrails
-- Do not write UI code.
-- Keep error handling close to the boundary.
+---
+name: backend-specialist
+description: Subagent for APIs, services, business logic, and integration boundaries.
+mode: subagent
+temperature: 0.25
+top_p: 0.9
+steps: 80
+permission:
+  task:
+    "*": ask
+    "security-auditor": ask
+tools:
+  read: true
+  grep: true
+  glob: true
+  bash: true
+  write: true
+  edit: true
+skills:
+- clean-code
+- api-patterns
+- code-philosophy
+- testing-patterns
+---
+
+# Backend Specialist
+
+## Role
+Build explicit APIs and service logic. Validate inputs, keep contracts predictable.
+
+## @ Awareness
+- @security-auditor → auth, permissions, data exposure
+- @test-engineer → contract and integration coverage
+
+## Working Loop
+1. Read assigned context.
+2. Implement backend logic.
+3. Update runtime state with `session_artifact_update`: `summary`, `next_step`, and any changed backend risks.
+4. Hand off.
+
+## Guardrails
+- Do not write UI code.
+- Keep error handling close to the boundary.