Imported from BrightWW/wiwynn-fw-agent (.opencode/agents/AGENTS.md). Install upstream with npx skills add BrightWW/wiwynn-fw-agent --skill agents. Copyright stays with the author.

agents/ — Firmware Pipeline Agents

OVERVIEW

4 specialized agents. Each has YAML frontmatter (name, description, model). Loaded by /fw-dev via task() calls — not invoked directly by users (except fw-reviewer-sonnet which can be called standalone for review loops).

AGENT TABLE

AGENT DETAILS

fw-issue-analyst

• Input: JIRA key (e.g. GC20T5T7-121) or GitHub Issue URL
• Output: Structured analysis report — platform, root cause, scope, image extraction results
• Key behavior: MUST complete image extraction (structured_events, confidence, unknowns) before pipeline continues
• Vision: Processes JIRA attachments / GitHub issue images

fw-code-analyst

• Input: Analysis report from fw-issue-analyst
• Output: Modification plan — target files, functions, line numbers, what to change and why
• Key behavior: 3-step search (search_github → grep → fetch); never fetches large files blindly

fw-coder

• Input (Step 3): Modification plan → Output: Unified diff (git-apply ready)
• Input (Step 4): Staged diff → Output: EF1900-compliant commit message with [AI Agent] section
• Input (Step 7): Review feedback → Output: Fixed unified diff
• Reasoning switch: high for code writing (Step 3/7), medium for commit message (Step 4)

fw-reviewer-sonnet

• Input: PR diff + commit message + original issue context
• Output: APPROVE or REQUEST_CHANGES with 7-dimension analysis
• 7 dimensions: Coding Style, Error Handling, Memory Safety, Commit Message, Platform Isolation, Logic Correctness, Test Coverage
• Speed: Sonnet chosen for fast iterative loop turnaround

ANTI-PATTERNS

• DO NOT invoke fw-coder with subtask: true — fw-coder.md explicitly sets subtask: false
• DO NOT skip fw-issue-analyst image extraction — pipeline may proceed on incomplete context
• DO NOT reuse a previous agent session for a new JIRA issue — always spawn fresh task()