feat(pi): support pi reviewer model routing
This commit is contained in:
Stefano Fiorini
@@ -6,6 +6,29 @@ This document covers the shared reviewer-runtime helpers used by the Pi workflow
|
||||
|
||||
It is intentionally separate from [PI-SUPERPOWERS.md](./PI-SUPERPOWERS.md). Superpowers are skill dependencies; reviewer-runtime is helper-script setup.
|
||||
|
||||
## Pi As A Reviewer CLI
|
||||
|
||||
Pi workflow skills may use `pi` itself as the reviewer CLI, even when the main workflow is already running in pi. In that case, the reviewer model is configured independently from the running agent model. This lets the operator run a workflow with one model while asking another pi-configured model, including provider-qualified model IDs, to review the plan or implementation.
|
||||
|
||||
The canonical isolated, read-only reviewer command is:
|
||||
|
||||
```bash
|
||||
pi --no-session --no-skills --no-prompt-templates --no-extensions --no-context-files --model "$REVIEWER_MODEL" --tools read,grep,find,ls -p "Read the review payload and return the requested verdict."
|
||||
```
|
||||
|
||||
Pi reviewer calls must stay isolated from the main workflow:
|
||||
|
||||
- Use `--no-session` so the reviewer does not continue or persist the workflow session.
|
||||
- Use `--no-skills --no-prompt-templates --no-extensions --no-context-files` so the reviewer does not load workflow skills, project context files, or package extensions that could re-enter `create-plan`, `implement-plan`, or `do-task`.
|
||||
- Use exactly `--tools read,grep,find,ls` for review. The pi reviewer command MUST NOT include `write`, `edit`, or `bash`; the reviewer reads payloads and diffs but never modifies files or runs commands.
|
||||
|
||||
If the reviewer subprocess exits non-zero because the provider, credentials, or model ID are unavailable, surface the captured stderr/status from `run-review.sh`, then ask the user for a configured reviewer model. Use `pi --list-models [search]` to inspect available configured models when needed.
|
||||
|
||||
Official references:
|
||||
|
||||
- `https://github.com/badlogic/pi-mono/tree/main/packages/coding-agent` documents pi providers, model selection, skills, extensions, and print mode.
|
||||
- Local `pi --help` for pi `0.70.0` confirms `--model <pattern>` supports `provider/id`, `--print, -p` runs non-interactively, `--tools, -t <tools>` allowlists tools, and the read-only example is `pi --tools read,grep,find,ls -p "Review the code in src/"`.
|
||||
|
||||
## Required Files
|
||||
|
||||
The workflow-heavy Pi skills expect these helper files from `skills/reviewer-runtime/pi/`:
|
||||
|
||||
@@ -71,17 +71,30 @@ If any dependency is missing, stop and return:
|
||||
|
||||
If the user already specified a reviewer CLI and model, use those values. Otherwise ask:
|
||||
|
||||
Reviewer CLI: `codex`, `claude`, `cursor`, `opencode`, `pi`, or `skip`
|
||||
|
||||
1. Which CLI should review the plan?
|
||||
- `codex`
|
||||
- `claude`
|
||||
- `cursor`
|
||||
- `opencode`
|
||||
- `pi`
|
||||
- `skip`
|
||||
2. Which model?
|
||||
3. Max review rounds? Default: `10`
|
||||
|
||||
Store `REVIEWER_CLI`, `REVIEWER_MODEL`, and `MAX_ROUNDS` for the review loop.
|
||||
|
||||
When `REVIEWER_CLI=pi`, the reviewer model is configured independently from the pi model running this workflow. Use any configured pi model string, including provider-qualified model IDs. If the reviewer model or provider is unavailable, surface the review helper stderr/status and ask for a configured model; use `pi --list-models [search]` to inspect configured models.
|
||||
|
||||
The pi reviewer command rendered into `/tmp/plan-review-${REVIEW_ID}.sh` must be isolated and read-only:
|
||||
|
||||
```bash
|
||||
pi --no-session --no-skills --no-prompt-templates --no-extensions --no-context-files --model "$REVIEWER_MODEL" --tools read,grep,find,ls -p "Read the file /tmp/plan-${REVIEW_ID}.md and review."
|
||||
```
|
||||
|
||||
The pi reviewer invocation must not load workflow skills and must not include `write`, `edit`, or `bash` tools.
|
||||
|
||||
### Phase 4: Design
|
||||
|
||||
- Load `brainstorming`.
|
||||
|
||||
@@ -95,12 +95,24 @@ Use current-branch execution by default. Only switch to a worktree when the prom
|
||||
|
||||
If the user already specified reviewer settings, use them. Otherwise ask:
|
||||
|
||||
1. Reviewer CLI: `codex`, `claude`, `cursor`, `opencode`, or `skip`
|
||||
Reviewer CLI: `codex`, `claude`, `cursor`, `opencode`, `pi`, or `skip`
|
||||
|
||||
1. Which CLI should review the plan and implementation?
|
||||
2. Reviewer model
|
||||
3. Max rounds, default `10`
|
||||
|
||||
Store `REVIEWER_CLI`, `REVIEWER_MODEL`, and `MAX_ROUNDS`.
|
||||
|
||||
When `REVIEWER_CLI=pi`, the reviewer model is configured independently from the pi model running this workflow. Use any configured pi model string, including provider-qualified model IDs. If the reviewer model or provider is unavailable, surface the review helper stderr/status and ask for a configured model; use `pi --list-models [search]` to inspect configured models.
|
||||
|
||||
The pi reviewer command rendered into `/tmp/do-task-${REVIEW_KIND}-review-${REVIEW_ID}.sh` must be isolated and read-only:
|
||||
|
||||
```bash
|
||||
pi --no-session --no-skills --no-prompt-templates --no-extensions --no-context-files --model "$REVIEWER_MODEL" --tools read,grep,find,ls -p "Read the file /tmp/do-task-${REVIEW_KIND}-${REVIEW_ID}.md and review."
|
||||
```
|
||||
|
||||
The pi reviewer invocation must not load workflow skills and must not include `write`, `edit`, or `bash` tools.
|
||||
|
||||
### Phase 4: Initialize `task-plan.md`
|
||||
|
||||
1. Compute `ai_plan/YYYY-MM-DD-<slug>/`
|
||||
|
||||
@@ -9,7 +9,7 @@
|
||||
| Created | YYYY-MM-DD |
|
||||
| Slug | YYYY-MM-DD-<slug> |
|
||||
| Runtime | pi |
|
||||
| Reviewer CLI | codex \| claude \| cursor \| opencode |
|
||||
| Reviewer CLI | codex \| claude \| cursor \| opencode \| pi |
|
||||
| Reviewer Model | <model> |
|
||||
| MAX_ROUNDS | 10 |
|
||||
| Branch Strategy | current-branch \| worktree |
|
||||
|
||||
@@ -78,12 +78,24 @@ If any dependency is missing, stop and return:
|
||||
|
||||
If the user already provided reviewer settings, use them. Otherwise ask:
|
||||
|
||||
1. Reviewer CLI: `codex`, `claude`, `cursor`, `opencode`, or `skip`
|
||||
Reviewer CLI: `codex`, `claude`, `cursor`, `opencode`, `pi`, or `skip`
|
||||
|
||||
1. Which CLI should review milestone implementations?
|
||||
2. Reviewer model
|
||||
3. Max rounds, default `10`
|
||||
|
||||
Store `REVIEWER_CLI`, `REVIEWER_MODEL`, and `MAX_ROUNDS`.
|
||||
|
||||
When `REVIEWER_CLI=pi`, the reviewer model is configured independently from the pi model running this workflow. Use any configured pi model string, including provider-qualified model IDs. If the reviewer model or provider is unavailable, surface the review helper stderr/status and ask for a configured model; use `pi --list-models [search]` to inspect configured models.
|
||||
|
||||
The pi reviewer command rendered into `/tmp/milestone-review-${REVIEW_ID}.sh` must be isolated and read-only:
|
||||
|
||||
```bash
|
||||
pi --no-session --no-skills --no-prompt-templates --no-extensions --no-context-files --model "$REVIEWER_MODEL" --tools read,grep,find,ls -p "Read the file /tmp/milestone-${REVIEW_ID}.md and review."
|
||||
```
|
||||
|
||||
The pi reviewer invocation must not load workflow skills and must not include `write`, `edit`, or `bash` tools.
|
||||
|
||||
### Phase 3: Set Up Workspace
|
||||
|
||||
1. Load `using-git-worktrees`
|
||||
|
||||
@@ -43,8 +43,17 @@ find skills/atlassian/pi/scripts -type f -print -quit | grep -q .
|
||||
for file in skills/create-plan/pi/SKILL.md skills/do-task/pi/SKILL.md skills/implement-plan/pi/SKILL.md; do
|
||||
grep -q 'docs/PI-SUPERPOWERS.md' "$file"
|
||||
grep -q 'docs/PI-COMMON-REVIEWER.md' "$file"
|
||||
grep -q 'Reviewer CLI: `codex`, `claude`, `cursor`, `opencode`, `pi`, or `skip`' "$file"
|
||||
grep -q 'pi --no-session --no-skills --no-prompt-templates --no-extensions --no-context-files' "$file"
|
||||
grep -q -- '--tools read,grep,find,ls -p' "$file"
|
||||
grep -q 'pi --list-models \[search\]' "$file"
|
||||
done
|
||||
|
||||
grep -q 'reviewer model is configured independently' docs/PI-COMMON-REVIEWER.md
|
||||
grep -q 'provider-qualified model IDs' docs/PI-COMMON-REVIEWER.md
|
||||
grep -q 'MUST NOT include `write`, `edit`, or `bash`' docs/PI-COMMON-REVIEWER.md
|
||||
grep -q 'Reviewer CLI | codex \\| claude \\| cursor \\| opencode \\| pi' skills/do-task/pi/templates/task-plan.md
|
||||
|
||||
grep -q 'pi-package/skills/atlassian/scripts' skills/atlassian/pi/SKILL.md
|
||||
grep -q 'pi-package/skills/web-automation/scripts' skills/web-automation/pi/SKILL.md
|
||||
grep -q 'local checkout package install keeps the runtime in `pi-package/skills/<skill>/scripts`' docs/PI.md
|
||||
|
||||
@@ -13,12 +13,33 @@ WORKFLOW_FILES=(
|
||||
for file in "${WORKFLOW_FILES[@]}"; do
|
||||
test -f "$file"
|
||||
grep -q 'docs/PI-SUPERPOWERS.md' "$file"
|
||||
grep -q 'Reviewer CLI: `codex`, `claude`, `cursor`, `opencode`, `pi`, or `skip`' "$file"
|
||||
grep -q 'pi --no-session --no-skills --no-prompt-templates --no-extensions --no-context-files' "$file"
|
||||
grep -q -- '--tools read,grep,find,ls -p' "$file"
|
||||
grep -q 'pi --list-models \[search\]' "$file"
|
||||
done
|
||||
|
||||
grep -q 'reviewer model is configured independently' docs/PI-COMMON-REVIEWER.md
|
||||
grep -q 'provider-qualified model IDs' docs/PI-COMMON-REVIEWER.md
|
||||
grep -q 'MUST NOT include `write`, `edit`, or `bash`' docs/PI-COMMON-REVIEWER.md
|
||||
|
||||
if command -v pi >/dev/null 2>&1; then
|
||||
PI_HELP=$(pi --help 2>&1)
|
||||
grep -q -- '--model <pattern>' <<<"$PI_HELP"
|
||||
grep -q -- '--print, -p' <<<"$PI_HELP"
|
||||
grep -q -- '--tools, -t <tools>' <<<"$PI_HELP"
|
||||
grep -q -- '--no-session' <<<"$PI_HELP"
|
||||
grep -q -- '--no-skills' <<<"$PI_HELP"
|
||||
grep -q -- '--no-prompt-templates' <<<"$PI_HELP"
|
||||
grep -q -- '--no-extensions' <<<"$PI_HELP"
|
||||
grep -q -- '--no-context-files' <<<"$PI_HELP"
|
||||
fi
|
||||
|
||||
test -f skills/create-plan/pi/templates/continuation-runbook.md
|
||||
test -f skills/create-plan/pi/templates/milestone-plan.md
|
||||
test -f skills/create-plan/pi/templates/story-tracker.md
|
||||
test -f skills/do-task/pi/templates/task-plan.md
|
||||
grep -q 'Reviewer CLI | codex \\| claude \\| cursor \\| opencode \\| pi' skills/do-task/pi/templates/task-plan.md
|
||||
test -x skills/reviewer-runtime/pi/run-review.sh
|
||||
test -x skills/reviewer-runtime/pi/notify-telegram.sh
|
||||
|
||||
|
||||
@@ -71,17 +71,30 @@ If any dependency is missing, stop and return:
|
||||
|
||||
If the user already specified a reviewer CLI and model, use those values. Otherwise ask:
|
||||
|
||||
Reviewer CLI: `codex`, `claude`, `cursor`, `opencode`, `pi`, or `skip`
|
||||
|
||||
1. Which CLI should review the plan?
|
||||
- `codex`
|
||||
- `claude`
|
||||
- `cursor`
|
||||
- `opencode`
|
||||
- `pi`
|
||||
- `skip`
|
||||
2. Which model?
|
||||
3. Max review rounds? Default: `10`
|
||||
|
||||
Store `REVIEWER_CLI`, `REVIEWER_MODEL`, and `MAX_ROUNDS` for the review loop.
|
||||
|
||||
When `REVIEWER_CLI=pi`, the reviewer model is configured independently from the pi model running this workflow. Use any configured pi model string, including provider-qualified model IDs. If the reviewer model or provider is unavailable, surface the review helper stderr/status and ask for a configured model; use `pi --list-models [search]` to inspect configured models.
|
||||
|
||||
The pi reviewer command rendered into `/tmp/plan-review-${REVIEW_ID}.sh` must be isolated and read-only:
|
||||
|
||||
```bash
|
||||
pi --no-session --no-skills --no-prompt-templates --no-extensions --no-context-files --model "$REVIEWER_MODEL" --tools read,grep,find,ls -p "Read the file /tmp/plan-${REVIEW_ID}.md and review."
|
||||
```
|
||||
|
||||
The pi reviewer invocation must not load workflow skills and must not include `write`, `edit`, or `bash` tools.
|
||||
|
||||
### Phase 4: Design
|
||||
|
||||
- Load `brainstorming`.
|
||||
|
||||
@@ -95,12 +95,24 @@ Use current-branch execution by default. Only switch to a worktree when the prom
|
||||
|
||||
If the user already specified reviewer settings, use them. Otherwise ask:
|
||||
|
||||
1. Reviewer CLI: `codex`, `claude`, `cursor`, `opencode`, or `skip`
|
||||
Reviewer CLI: `codex`, `claude`, `cursor`, `opencode`, `pi`, or `skip`
|
||||
|
||||
1. Which CLI should review the plan and implementation?
|
||||
2. Reviewer model
|
||||
3. Max rounds, default `10`
|
||||
|
||||
Store `REVIEWER_CLI`, `REVIEWER_MODEL`, and `MAX_ROUNDS`.
|
||||
|
||||
When `REVIEWER_CLI=pi`, the reviewer model is configured independently from the pi model running this workflow. Use any configured pi model string, including provider-qualified model IDs. If the reviewer model or provider is unavailable, surface the review helper stderr/status and ask for a configured model; use `pi --list-models [search]` to inspect configured models.
|
||||
|
||||
The pi reviewer command rendered into `/tmp/do-task-${REVIEW_KIND}-review-${REVIEW_ID}.sh` must be isolated and read-only:
|
||||
|
||||
```bash
|
||||
pi --no-session --no-skills --no-prompt-templates --no-extensions --no-context-files --model "$REVIEWER_MODEL" --tools read,grep,find,ls -p "Read the file /tmp/do-task-${REVIEW_KIND}-${REVIEW_ID}.md and review."
|
||||
```
|
||||
|
||||
The pi reviewer invocation must not load workflow skills and must not include `write`, `edit`, or `bash` tools.
|
||||
|
||||
### Phase 4: Initialize `task-plan.md`
|
||||
|
||||
1. Compute `ai_plan/YYYY-MM-DD-<slug>/`
|
||||
|
||||
@@ -9,7 +9,7 @@
|
||||
| Created | YYYY-MM-DD |
|
||||
| Slug | YYYY-MM-DD-<slug> |
|
||||
| Runtime | pi |
|
||||
| Reviewer CLI | codex \| claude \| cursor \| opencode |
|
||||
| Reviewer CLI | codex \| claude \| cursor \| opencode \| pi |
|
||||
| Reviewer Model | <model> |
|
||||
| MAX_ROUNDS | 10 |
|
||||
| Branch Strategy | current-branch \| worktree |
|
||||
|
||||
@@ -78,12 +78,24 @@ If any dependency is missing, stop and return:
|
||||
|
||||
If the user already provided reviewer settings, use them. Otherwise ask:
|
||||
|
||||
1. Reviewer CLI: `codex`, `claude`, `cursor`, `opencode`, or `skip`
|
||||
Reviewer CLI: `codex`, `claude`, `cursor`, `opencode`, `pi`, or `skip`
|
||||
|
||||
1. Which CLI should review milestone implementations?
|
||||
2. Reviewer model
|
||||
3. Max rounds, default `10`
|
||||
|
||||
Store `REVIEWER_CLI`, `REVIEWER_MODEL`, and `MAX_ROUNDS`.
|
||||
|
||||
When `REVIEWER_CLI=pi`, the reviewer model is configured independently from the pi model running this workflow. Use any configured pi model string, including provider-qualified model IDs. If the reviewer model or provider is unavailable, surface the review helper stderr/status and ask for a configured model; use `pi --list-models [search]` to inspect configured models.
|
||||
|
||||
The pi reviewer command rendered into `/tmp/milestone-review-${REVIEW_ID}.sh` must be isolated and read-only:
|
||||
|
||||
```bash
|
||||
pi --no-session --no-skills --no-prompt-templates --no-extensions --no-context-files --model "$REVIEWER_MODEL" --tools read,grep,find,ls -p "Read the file /tmp/milestone-${REVIEW_ID}.md and review."
|
||||
```
|
||||
|
||||
The pi reviewer invocation must not load workflow skills and must not include `write`, `edit`, or `bash` tools.
|
||||
|
||||
### Phase 3: Set Up Workspace
|
||||
|
||||
1. Load `using-git-worktrees`
|
||||
|
||||
Reference in New Issue
Block a user