tools-only fn-17-wti Centralize Review Backend Detection
Problem
Currently, 6 skill files detect rp-cli and codex availability at runtime using which rp-cli and which codex. This causes:
- LLM deviation - Agents sometimes check wrong binary names (
rp,repopromptinstead ofrp-cli) - Redundant checks - Same detection repeated in every skill invocation (12+ subprocess calls)
- Inconsistency - Ralph mode already handles this correctly via config
Solution
Move detection to /flow-next:setup (one-time), persist to .flow/config.json, remove runtime detection from skills.
Design Decisions
| Question | Decision |
|---|---|
| Tool unavailable at runtime but configured | Error with clear message: "rp-cli configured but not found. Run /flow-next:setup to reconfigure." |
| Setup when no tools available | Show options with note "neither detected", default to "none" |
| Config key | review.backend (nested, consistent with memory.enabled, planSync.enabled) |
| Re-run setup behavior | Pre-select current value, allow user to change |
| User skips question | Write "none" explicitly (no ambiguity) |
Priority Order (all skills)
Skills use this priority (first match wins):
--review=rp|codex|export|noneargument (explicit user choice)FLOW_REVIEW_BACKENDenv var (Ralph mode sets this).flow/config.json→review.backend(setup configured this)- ERROR - "No review backend configured. Run /flow-next:setup or pass --review=X"
NO auto-detection fallback. Trust explicit config or fail fast.
Scope
In scope
- Add review backend question to
/flow-next:setup - Add
review.backendto default config in flowctl.py - Remove runtime detection from 6 skill files
- Update error messages to guide users
Out of scope
- Changes to
/flow-next:ralph-init(already works correctly viaFLOW_REVIEW_BACKENDenv) - Per-command backend config (future enhancement)
- Tool path caching (future enhancement)
Files to Modify
| File | Change |
|---|---|
plugins/flow-next/scripts/flowctl.py |
Add review.backend to get_default_config() |
plugins/flow-next/skills/flow-next-setup/workflow.md |
Add review backend question to Step 6 |
plugins/flow-next/skills/flow-next-plan/SKILL.md |
Remove which detection |
plugins/flow-next/skills/flow-next-work/SKILL.md |
Remove which detection |
plugins/flow-next/skills/flow-next-plan-review/SKILL.md |
Remove which detection |
plugins/flow-next/skills/flow-next-plan-review/workflow.md |
Remove which detection |
plugins/flow-next/skills/flow-next-impl-review/SKILL.md |
Remove which detection |
plugins/flow-next/skills/flow-next-impl-review/workflow.md |
Remove which detection |
Quick Commands
# Run smoke tests
plugins/flow-next/scripts/smoke_test.sh
# Verify config default includes review.backend
python3 plugins/flow-next/scripts/flowctl.py init --json
cat .flow/config.json | jq '.review'
# Verify no 'which rp-cli' in skill files
grep -r "which rp-cli" plugins/flow-next/skills/ && echo "FAIL: detection still present" || echo "PASS: detection removed"
Acceptance
-
/flow-next:setupasks about review backend and writes to config -
flowctl initcreates config withreview.backend: null(or absent) - All 6 skill files no longer run
which rp-cliorwhich codex - Skills read from: flag → env → config → error (no auto-detect)
- Error message is clear when no backend configured
- Ralph mode unaffected (FLOW_REVIEW_BACKEND env still works)
- Smoke tests pass
- README docs updated (remove "auto-detect" from priority)
References
- Gap analysis identified edge cases and priority questions
- Ralph mode uses
FLOW_REVIEW_BACKENDenv (lines 798, 816 of ralph.sh) - Config pattern in flowctl.py lines 76-78, 107-149