feat: add retrospect-workflow for self-improvement
This commit is contained in:
@@ -0,0 +1,220 @@
|
||||
name: "retrospect-workflow"
|
||||
description: "Post-execution retrospective: analyze a completed thread, find inefficiencies, and improve the workflow definition."
|
||||
roles:
|
||||
analyst:
|
||||
description: "Scans thread execution for anomalies and produces a findings report"
|
||||
goal: "You are a workflow execution analyst. You review completed thread data to find inefficiencies, wasted effort, and procedure gaps."
|
||||
capabilities:
|
||||
- data-analysis
|
||||
procedure: |
|
||||
You receive a completed thread ID in your task prompt.
|
||||
|
||||
Phase 0 — Validation (must pass before any analysis):
|
||||
1. Run `uwf step list <thread-id>` to get thread metadata including the workflow hash
|
||||
2. Run `uwf workflow show <workflow-hash>` to get the workflow name
|
||||
3. Verify the workflow exists locally: check `.workflows/<name>.yaml` in the current repo
|
||||
- If NOT found: output $status=wrong_project with the workflow name. Do NOT proceed.
|
||||
4. Compare the thread's workflow hash against the current registered version:
|
||||
- Run `uwf workflow show <name>` to get the current hash
|
||||
- If hashes differ: the thread ran on an older version. Note this — you will need to diff versions after analysis.
|
||||
|
||||
Phase 1 — Overview scan:
|
||||
5. From the step list, compute a health signal for each step:
|
||||
- Duration: flag if >2x the median of other steps
|
||||
- Output tokens: flag if >2x the median
|
||||
- Status flow: flag non-happy-path transitions (rejected, fix_code, fix_spec, hook_failed)
|
||||
- Step count: flag if the same role appears more than expected (indicates loops)
|
||||
6. If no anomalies found AND versions match: output $status=clean
|
||||
7. If no anomalies found BUT versions differ:
|
||||
- Diff the two workflow versions to check if any procedure changes are relevant
|
||||
- If the current version already addresses potential concerns: output $status=clean with a note
|
||||
- Otherwise: proceed to Phase 2
|
||||
|
||||
Phase 2 — Targeted deep-dive (only for flagged steps):
|
||||
8. For each flagged step, run `uwf step show <hash>` to get the detail with turns
|
||||
9. Analyze the turn sequence for:
|
||||
- Repeated tool calls with the same or similar input (blind retries)
|
||||
- Tool errors followed by no strategy change (same approach retried)
|
||||
- Unnecessary exploration (reading files or running commands unrelated to the task)
|
||||
- Hallucinated commands or flags (commands that don't exist or wrong syntax)
|
||||
- Excessive turns before reaching the goal
|
||||
10. For each finding, record:
|
||||
- Which role and step hash
|
||||
- What happened (specific turn indices and commands)
|
||||
- Root cause hypothesis (procedure gap, missing pitfall, unclear instruction)
|
||||
- Suggested fix (what to add/change in the procedure)
|
||||
11. If versions differ: compare findings against the version diff.
|
||||
Mark any finding that is already fixed in the current version as "resolved_in_current".
|
||||
Only report findings that are NOT yet addressed.
|
||||
|
||||
Output a structured findings report. Set $status=clean if nothing actionable, $status=findings if unresolved issues exist, or $status=wrong_project if the workflow doesn't belong here.
|
||||
output: "A findings report with per-issue root cause and suggested procedure fixes. Set $status to clean or findings (with report hash)."
|
||||
frontmatter:
|
||||
oneOf:
|
||||
- properties:
|
||||
$status: { const: "clean" }
|
||||
summary: { type: string }
|
||||
required: [$status, summary]
|
||||
- properties:
|
||||
$status: { const: "findings" }
|
||||
report: { type: string }
|
||||
targetWorkflow: { type: string }
|
||||
required: [$status, report, targetWorkflow]
|
||||
- properties:
|
||||
$status: { const: "wrong_project" }
|
||||
workflowName: { type: string }
|
||||
required: [$status, workflowName]
|
||||
proposer:
|
||||
description: "Translates findings into concrete workflow edits"
|
||||
goal: "You are a workflow improvement proposer. You read the analyst's findings and produce specific, minimal edits to the workflow YAML."
|
||||
capabilities:
|
||||
- planning
|
||||
procedure: |
|
||||
1. Read the analyst's findings report from your task prompt
|
||||
2. Locate the target workflow YAML:
|
||||
- Workflow definitions live in the WORKFLOW ENGINE repo (where `uwf` is developed), NOT in the repo that was analyzed.
|
||||
- Find it via: `uwf workflow show <targetWorkflow> --format yaml` to read the current definition
|
||||
- The physical file is `.workflows/<targetWorkflow>.yaml` in the workflow engine repo
|
||||
- Use `git rev-parse --show-toplevel` in the current directory to find the workflow engine repo root
|
||||
3. Read the current workflow YAML to understand existing procedures
|
||||
4. For each finding, draft a minimal edit:
|
||||
- Prefer adding a pitfall note or clarifying instruction over restructuring
|
||||
- If a procedure step is ambiguous, make it explicit
|
||||
- If a tool usage pattern is wrong, add a "Do NOT" or "IMPORTANT" note
|
||||
- Keep edits surgical — don't rewrite procedures that work fine
|
||||
5. Check if existing tests need updating (search for test files referencing the workflow)
|
||||
6. Produce a change plan as CAS text node via `uwf cas put-text "<plan>"`
|
||||
|
||||
The plan should list each edit with:
|
||||
- File path
|
||||
- What to change (old text → new text, or addition)
|
||||
- Why (linked to which finding)
|
||||
- Any test updates needed
|
||||
output: "A change plan stored in CAS. Set $status to ready (with plan hash and repoPath) or no_action (if findings don't warrant changes)."
|
||||
frontmatter:
|
||||
oneOf:
|
||||
- properties:
|
||||
$status: { const: "ready" }
|
||||
plan: { type: string }
|
||||
repoPath: { type: string }
|
||||
required: [$status, plan, repoPath]
|
||||
- properties:
|
||||
$status: { const: "no_action" }
|
||||
reason: { type: string }
|
||||
required: [$status, reason]
|
||||
developer:
|
||||
description: "Applies the proposed workflow edits"
|
||||
goal: "You are a developer agent. You apply workflow YAML edits and update related tests."
|
||||
capabilities:
|
||||
- coding
|
||||
procedure: |
|
||||
IMPORTANT: Always work in a git worktree, NEVER modify the main working directory directly.
|
||||
The workflow definitions live in THIS repo (the workflow engine), not the repo that was analyzed.
|
||||
|
||||
Before starting any work, set up an isolated worktree:
|
||||
1. Use `git rev-parse --show-toplevel` to find the repo root (do NOT use repoPath from proposer — that's the analyzed repo)
|
||||
2. `git fetch origin` to get latest refs
|
||||
3. `git worktree add .worktrees/retrospect/<short-slug> -b retrospect/<short-slug> origin/main`
|
||||
4. `cd .worktrees/retrospect/<short-slug> && bun install`
|
||||
5. ALL subsequent work must happen inside the worktree directory.
|
||||
|
||||
Then apply changes:
|
||||
6. Read the change plan from CAS: `uwf cas get <plan hash>`
|
||||
7. Apply each edit from the plan to the workflow YAML
|
||||
8. Update or add tests as specified in the plan
|
||||
9. Run `bun run build` and `bun test` to verify
|
||||
10. Run `bun run check` for lint
|
||||
11. Commit with message: `improve: <workflow-name> — <brief summary>`
|
||||
output: "List all files changed and provide a summary. Set $status to done (with branch/worktree), or failed (with reason)."
|
||||
frontmatter:
|
||||
oneOf:
|
||||
- properties:
|
||||
$status: { const: "done" }
|
||||
branch: { type: string }
|
||||
worktree: { type: string }
|
||||
required: [$status, branch, worktree]
|
||||
- properties:
|
||||
$status: { const: "failed" }
|
||||
reason: { type: string }
|
||||
required: [$status, reason]
|
||||
reviewer:
|
||||
description: "Reviews the workflow edits for correctness"
|
||||
goal: "You are a reviewer. You verify that workflow edits are minimal, correct, and actually address the findings."
|
||||
capabilities:
|
||||
- code-review
|
||||
procedure: |
|
||||
The worktree path is provided in your task prompt. cd into it first.
|
||||
|
||||
Review criteria:
|
||||
1. Each edit must trace back to a specific finding — no drive-by changes
|
||||
2. Edits should be minimal — don't rewrite working procedures
|
||||
3. New pitfall notes or instructions must be clear and actionable
|
||||
4. Tests must be updated if assertions changed
|
||||
5. `bun run build` and `bun test` must pass
|
||||
6. `bunx biome check` must pass
|
||||
|
||||
IMPORTANT: `tea pr create` must run from the MAIN repo directory (not a worktree), because tea cannot detect the repo from worktree `.git` files.
|
||||
output: "Explain your decision. Set $status to approved (with branch/worktree) or rejected (with comments)."
|
||||
frontmatter:
|
||||
oneOf:
|
||||
- properties:
|
||||
$status: { const: "approved" }
|
||||
branch: { type: string }
|
||||
worktree: { type: string }
|
||||
required: [$status, branch, worktree]
|
||||
- properties:
|
||||
$status: { const: "rejected" }
|
||||
comments: { type: string }
|
||||
worktree: { type: string }
|
||||
required: [$status, comments, worktree]
|
||||
committer:
|
||||
description: "Commits and creates PR"
|
||||
goal: "You are a committer agent. You create a clean commit and push a PR."
|
||||
capabilities: []
|
||||
procedure: |
|
||||
The worktree path, branch name, and repo info are provided in your task prompt.
|
||||
cd into the worktree first.
|
||||
|
||||
Note: You inherit the developer's worktree and branch. Do NOT create a new branch.
|
||||
1. Stage all changes: `git add -A`
|
||||
2. Commit with a descriptive message: `git commit -m "improve: <workflow> — <summary>"`
|
||||
3. Push the branch: `git push -u origin <branch-name>`
|
||||
- If push hook fails: capture the error log in your output, mark hook_failed
|
||||
4. On push success: create a PR via `tea pr create --title "..." --description "..."`
|
||||
- IMPORTANT: `tea pr create` must run from the MAIN repo directory (not a worktree), because tea cannot detect the repo from worktree `.git` files. cd to the repo root first.
|
||||
- Do NOT pass `--repo` — let tea auto-detect from the main repo's git remote.
|
||||
- PR description must include: What / Why / Findings / Changes sections
|
||||
- On tea failure: capture stderr/stdout, include PR details for manual creation, mark hook_failed
|
||||
5. After PR creation, clean up the worktree:
|
||||
- cd to the repo root (parent of .worktrees)
|
||||
- `git worktree remove <worktree-path>`
|
||||
output: "Include PR URL on success or error log on failure. Set $status to committed (with prUrl) or hook_failed (with error)."
|
||||
frontmatter:
|
||||
oneOf:
|
||||
- properties:
|
||||
$status: { const: "committed" }
|
||||
prUrl: { type: string }
|
||||
required: [$status, prUrl]
|
||||
- properties:
|
||||
$status: { const: "hook_failed" }
|
||||
error: { type: string }
|
||||
required: [$status, error]
|
||||
graph:
|
||||
$START:
|
||||
_: { role: "analyst", prompt: "Analyze completed thread {{{threadId}}} for execution anomalies." }
|
||||
analyst:
|
||||
clean: { role: "$END", prompt: "No issues found. Thread executed cleanly." }
|
||||
findings: { role: "proposer", prompt: "Findings report: {{{report}}}. Target workflow: {{{targetWorkflow}}}. Propose minimal edits." }
|
||||
wrong_project: { role: "$END", prompt: "Thread uses workflow '{{{workflowName}}}' which does not exist in this project. Run retrospect from the correct repo." }
|
||||
proposer:
|
||||
no_action: { role: "$END", prompt: "No actionable changes needed: {{{reason}}}." }
|
||||
ready: { role: "developer", prompt: "Apply the change plan (CAS hash: {{{plan}}}) to the workflow definitions in this repo." }
|
||||
developer:
|
||||
done: { role: "reviewer", prompt: "Review workflow edits on branch {{{branch}}} at {{{worktree}}}." }
|
||||
failed: { role: "$END", prompt: "Developer failed: {{{reason}}}. Ending workflow." }
|
||||
reviewer:
|
||||
rejected: { role: "developer", prompt: "Reviewer rejected: {{{comments}}}. Fix the issues in {{{worktree}}}." }
|
||||
approved: { role: "committer", prompt: "Approved. Commit and push branch {{{branch}}} from {{{worktree}}}." }
|
||||
committer:
|
||||
hook_failed: { role: "developer", prompt: "Push hook failed: {{{error}}}. Fix and re-submit." }
|
||||
committed: { role: "$END", prompt: "PR created: {{{prUrl}}}. Workflow improved." }
|
||||
Reference in New Issue
Block a user