fix: use flowctl opencode commands instead of Task subagent

gmickel · gmickel · commit 2aeb4f0dc910 · 2026-01-20T01:38:57.000+01:00
Skills were using Task tool with subagent_type opencode-reviewer
but OpenCode doesn't recognize agents from opencode.json as Task
subagent types. Updated to use flowctl opencode plan-review and
impl-review commands which internally call opencode run --agent.
diff --git a/.opencode/skill/flow-next-opencode-impl-review/SKILL.md b/.opencode/skill/flow-next-opencode-impl-review/SKILL.md
@@ -97,10 +97,9 @@ fi
 5. **Re-reviews MUST stay in SAME chat** - omit `--new-chat` after first review
 
 **For opencode backend:**
-1. Use the **task tool** with subagent_type `opencode-reviewer`
-2. Reviewer gathers context via tools (git diff/log, read files)
-3. Parse verdict from reviewer output
-4. Extract `session_id` from `<task_metadata>` and reuse it for re-reviews
+1. Use `$FLOWCTL opencode impl-review` command
+2. Pass `--base` for diff scope and `--receipt` for session continuity
+3. Parse verdict from command output
 
 **For all backends:**
 - If `REVIEW_RECEIPT_PATH` set: write receipt after review (any verdict)
@@ -135,30 +134,17 @@ Run backend detection from SKILL.md above. Then branch:
 
 ### OpenCode Backend
 
-Use the task tool with subagent_type `opencode-reviewer`.
-
-Prompt must require:
-- `git log main..HEAD --oneline`
-- `git diff main..HEAD --stat`
-- `git diff main..HEAD`
-- Read any changed files needed for correctness
-- No questions, no code changes, no TodoWrite
-- End with `<verdict>SHIP</verdict>` or `<verdict>NEEDS_WORK</verdict>` or `<verdict>MAJOR_RETHINK</verdict>`
-
-Parse verdict from the subagent response.
-Extract `session_id` from `<task_metadata>` and reuse for re-review.
-
-On NEEDS_WORK: fix code, commit, re-run review (same backend).
-
-Write receipt if `REVIEW_RECEIPT_PATH` set:
 ```bash
-mkdir -p "$(dirname "$RECEIPT_PATH")"
-ts="$(date -u +%Y-%m-%dT%H:%M:%SZ)"
-cat > "$RECEIPT_PATH" <<EOF
-{"type":"impl_review","id":"$TASK_ID","mode":"opencode","verdict":"<VERDICT>","timestamp":"$ts"}
-EOF
+TASK_ID="${1:-}"
+BASE_COMMIT="${BASE_COMMIT:-main}"
+RECEIPT_PATH="${REVIEW_RECEIPT_PATH:-/tmp/impl-review-receipt.json}"
+
+$FLOWCTL opencode impl-review "$TASK_ID" --base "$BASE_COMMIT" --receipt "$RECEIPT_PATH"
+# Output includes VERDICT=SHIP|NEEDS_WORK|MAJOR_RETHINK
 ```
 
+On NEEDS_WORK: fix code, commit, re-run (receipt enables session continuity).
+
 ### RepoPrompt Backend
 
 ```bash
@@ -188,7 +174,7 @@ If verdict is NEEDS_WORK, loop internally until SHIP:
 2. **Fix code** and run tests/lints
 3. **Commit fixes** (mandatory before re-review)
 4. **Re-review**:
-   - **OpenCode**: re-run reviewer subagent with updated diff
+   - **OpenCode**: `$FLOWCTL opencode impl-review "$TASK_ID" --base "$BASE_COMMIT" --receipt "$RECEIPT_PATH"`
    - **RP**: `$FLOWCTL rp chat-send --window "$W" --tab "$T" --message-file /tmp/re-review.md` (NO `--new-chat`)
 5. **Repeat** until `<verdict>SHIP</verdict>`
 
diff --git a/.opencode/skill/flow-next-opencode-impl-review/workflow.md b/.opencode/skill/flow-next-opencode-impl-review/workflow.md
@@ -71,22 +71,16 @@ Include:
 - Review criteria (correctness, security, performance, tests, risks)
 - Required verdict tag
 
-### Step 3: Execute review (subagent)
+### Step 3: Execute review via flowctl
 
-Use the **task** tool with subagent_type `opencode-reviewer`. The reviewer must gather context itself via tools, including Flow task/epic specs.
+Use the `flowctl opencode impl-review` command:
 
-**Task tool call** (example):
-```json
-{
-  "description": "Impl review",
-  "prompt": "You are the OpenCode reviewer. Review current branch vs main. Rules: no questions, no code changes, no TodoWrite. REQUIRED: set FLOWCTL to `.opencode/bin/flowctl`, then run `$FLOWCTL show <TASK_ID> --json` and `$FLOWCTL cat <TASK_ID>`. Then get epic id from task JSON and run `$FLOWCTL show <EPIC_ID> --json` and `$FLOWCTL cat <EPIC_ID>`. REQUIRED: run `git log main..HEAD --oneline` (fallback master), `git diff main..HEAD --stat`, `git diff main..HEAD`. Read any changed files needed for correctness. Then output issues grouped by severity and end with exactly one verdict tag: <verdict>SHIP</verdict> or <verdict>NEEDS_WORK</verdict> or <verdict>MAJOR_RETHINK</verdict>.",
-  "subagent_type": "opencode-reviewer"
-}
+```bash
+$FLOWCTL opencode impl-review "$TASK_ID" --base "$BASE_BRANCH" --receipt "$RECEIPT_PATH"
+# Output includes VERDICT=SHIP|NEEDS_WORK|MAJOR_RETHINK
 ```
 
-**After the task completes**:
-- Parse `VERDICT` from the subagent output.
-- Extract `session_id` from the `<task_metadata>` block (used for re-reviews).
+Parse `VERDICT` from the command output.
 
 If `VERDICT` is empty, output `<promise>RETRY</promise>` and stop.
 
@@ -107,7 +101,7 @@ EOF
 If `VERDICT=NEEDS_WORK`:
 1. Parse issues from output
 2. Fix code, commit, run tests
-3. Re-run Step 3 **with the same task session_id** (pass `session_id` to the task tool)
+3. Re-run: `$FLOWCTL opencode impl-review "$TASK_ID" --base "$BASE_BRANCH" --receipt "$RECEIPT_PATH"`
 4. Repeat until SHIP
 
 ---
diff --git a/.opencode/skill/flow-next-opencode-plan-review/SKILL.md b/.opencode/skill/flow-next-opencode-plan-review/SKILL.md
@@ -97,10 +97,9 @@ fi
 5. **Re-reviews MUST stay in SAME chat** - omit `--new-chat` after first review
 
 **For opencode backend:**
-1. Use the **task tool** with subagent_type `opencode-reviewer`
-2. Reviewer gathers context via tools (`flowctl show/cat`)
-3. Parse verdict from reviewer output
-4. Extract `session_id` from `<task_metadata>` and reuse it for re-reviews
+1. Use `$FLOWCTL opencode plan-review` command
+2. Pass `--receipt` for session continuity on re-reviews
+3. Parse verdict from command output
 
 **For all backends:**
 - If `REVIEW_RECEIPT_PATH` set: write receipt after review (any verdict)
@@ -133,18 +132,15 @@ Run backend detection from SKILL.md above. Then branch:
 
 ### OpenCode Backend
 
-Use the task tool with subagent_type `opencode-reviewer`.
-
-Prompt must require:
-- `flowctl show <EPIC_ID> --json`
-- `flowctl cat <EPIC_ID>`
-- No questions, no code changes, no TodoWrite
-- End with `<verdict>SHIP</verdict>` or `<verdict>NEEDS_WORK</verdict>` or `<verdict>MAJOR_RETHINK</verdict>`
+```bash
+EPIC_ID="${1:-}"
+RECEIPT_PATH="${REVIEW_RECEIPT_PATH:-/tmp/plan-review-receipt.json}"
 
-Parse verdict from the subagent response.
-Extract `session_id` from `<task_metadata>` and reuse for re-review.
+$FLOWCTL opencode plan-review "$EPIC_ID" --receipt "$RECEIPT_PATH"
+# Output includes VERDICT=SHIP|NEEDS_WORK|MAJOR_RETHINK
+```
 
-On NEEDS_WORK: fix plan via `$FLOWCTL epic set-plan`, then re-run review (same backend).
+On NEEDS_WORK: fix plan via `$FLOWCTL epic set-plan`, then re-run (receipt enables session continuity).
 
 ### RepoPrompt Backend
 
diff --git a/.opencode/skill/flow-next-opencode-plan-review/workflow.md b/.opencode/skill/flow-next-opencode-plan-review/workflow.md
@@ -65,22 +65,16 @@ Include:
 - Review criteria (completeness, feasibility, clarity, architecture, risks, scope, testability)
 - Required verdict tag
 
-### Step 3: Execute review (subagent)
+### Step 3: Execute review via flowctl
 
-Use the **task** tool with subagent_type `opencode-reviewer`. The reviewer must gather context itself via tools, including Flow plan/spec.
+Use the `flowctl opencode plan-review` command:
 
-**Task tool call** (example):
-```json
-{
-  "description": "Plan review",
-  "prompt": "You are the OpenCode reviewer. Review the plan for EPIC_ID. Rules: no questions, no code changes, no TodoWrite. REQUIRED: set FLOWCTL to `.opencode/bin/flowctl`, then run `$FLOWCTL show <EPIC_ID> --json` and `$FLOWCTL cat <EPIC_ID>`. Review for completeness, feasibility, clarity, architecture, risks (incl security), scope, and testability. End with exactly one verdict tag: <verdict>SHIP</verdict> or <verdict>NEEDS_WORK</verdict> or <verdict>MAJOR_RETHINK</verdict>.",
-  "subagent_type": "opencode-reviewer"
-}
+```bash
+$FLOWCTL opencode plan-review "$EPIC_ID" --receipt "$RECEIPT_PATH"
+# Output includes VERDICT=SHIP|NEEDS_WORK|MAJOR_RETHINK
 ```
 
-**After the task completes**:
-- Parse `VERDICT` from the subagent output.
-- Extract `session_id` from the `<task_metadata>` block (used for re-reviews).
+Parse `VERDICT` from the command output.
 
 If `VERDICT` is empty, output `<promise>RETRY</promise>` and stop.
 
@@ -110,7 +104,7 @@ EOF
 If `VERDICT=NEEDS_WORK`:
 1. Parse issues from output
 2. Fix plan via `$FLOWCTL epic set-plan`
-3. Re-run Step 3 **with the same task session_id** (pass `session_id` to the task tool)
+3. Re-run: `$FLOWCTL opencode plan-review "$EPIC_ID" --receipt "$RECEIPT_PATH"`
 4. Repeat until SHIP
 
 ---
diff --git a/.opencode/skill/flow-next-opencode-ralph-init/templates/prompt_plan.md b/.opencode/skill/flow-next-opencode-ralph-init/templates/prompt_plan.md
@@ -17,7 +17,7 @@ Steps:
 
 Ralph mode rules (must follow):
 - If PLAN_REVIEW=rp: use `flowctl rp` wrappers (setup-review, select-add, prompt-get, chat-send).
-- If PLAN_REVIEW=opencode: use the task tool with subagent_type `opencode-reviewer`.
+- If PLAN_REVIEW=opencode: use `scripts/ralph/flowctl opencode plan-review` command.
 - Write receipt via bash heredoc (no Write tool) if REVIEW_RECEIPT_PATH is set.
 - Do NOT run /flow-next:* as shell commands.
 - If any rule is violated, output `<promise>RETRY</promise>` and stop.
@@ -26,7 +26,7 @@ Ralph mode rules (must follow):
    - Call the skill tool: flow-next-opencode-plan-review.
    - Follow the workflow in the skill using the exact arguments above.
    - Do NOT stop after loading the skill.
-   - For opencode: run reviewer via task tool and require `<verdict>` tag.
+   - For opencode: run `flowctl opencode plan-review` and require `<verdict>` tag.
    - For rp: use flowctl rp wrappers (no --json, no --new-chat on re-review).
    - For export: follow export flow in skill.
    - If PLAN_REVIEW=none:
