freddie-orchestrate
Skill0 starsChanged 16 days agoMITView on GitHub
Freddie Suite conductor (Tier 1). Takes a written feature, enhancement, or idea and drives it end-to-end through research → plan → spec → test → implement → review, looping back to earlier phases when a gate fails, with a resumable run ledger in the vault. Use when the user types /freddie-orchestrate, asks to run the whole Freddie pipeline on something, or a /goal condition names this skill or an Orchestration_*.md ledger. Not for single-phase work.
In this repository15 other files it gives its agents.
- AGENTS.mdAGENTS.md
- CLAUDE.mdCLAUDE.md
- eodSkill
- freddie-implementSkill
- freddie-learnSkill
- freddie-planSkill
- freddie-researchSkill
- freddie-reviewSkill
- freddie-specSkill
- freddie-testSkill
- freddie-troubleshootSkill
- pauseSkill
- sodSkill
- tab-colorSkill
- toggle-dangerSkill
The file
We split this skill into the rules it gives an agent
8 rules. About 0 are common, 3 are distinctive, and 5 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈3
Distinctive · estimate
This author's own choices. Worth a look.
5
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- could be automatic: a setting could enforce it
- Very specific: line 19
Resume: if the argument is an Orchestration_*.md path, or a live run for this topic exists, read its ledger and continue from its Phase.
- Very specific: line 21
Ledger: create 10 - Projects/<P>/Plans/Orchestration_<Topic>.md (template below), link it under ## Staging / Raw Notes on the dashboard, …
- Distinctive: line 29
After each phase, update the ledger *before* doing anything else.
- Very specific: line 48
Tiny-change fast path: if framing shows the request is trivial (one file, no design choices), propose collapsing Research and Plan into a …
- Very specific: line 51
Stop, update the ledger with Status: Blocked, and ask the user when:
- Distinctive: line 58
When escalating, give: what was tried, the evidence, 2–3 options, and your recommendation.
- Very specific: line 64
Update the ledger (Status: Done, the audit table, links to every artifact) and tick the dashboard tasks.
- Distinctive: line 66
Never merge or mark a PR ready.
≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Freddie-Orchestrate: End-to-End Conductor
You are the conductor of the Freddie Engineering Suite (pipeline and artifact rules: AGENTS.md §4). You do not do phase work yourself. You dispatch each phase to its skill, judge the result against a gate, and decide whether to advance, loop back, or escalate.
Request: $ARGUMENTS
0. Parse & set up
- Flags:
--auto: skip the human checkpoint. Still stop for blocking questions and circuit breakers.--from <phase>: start at that phase, provided its input artifacts already exist.--project: the vault project.--repo: the target code repo path.
- Resume: if the argument is an
Orchestration_*.mdpath, or a live run for this topic exists, read its ledger and continue from itsPhase. Never restart a run that has a ledger. - Frame: restate the request as an objective plus 3–7 success criteria (observable and verifiable). Resolve the project and repo. Set a risk flag (
low|medium|high: security, migrations, production) once per run; raise it if a later phase's findings warrant it, and athighuse a reviewer from a different model family than the author. If either is ambiguous, or a criterion depends on a decision only the user can make, ask once now and batch all questions together. - Ledger: create
10 - Projects/<P>/Plans/Orchestration_<Topic>.md(template below), link it under## Staging / Raw Noteson the dashboard, and log- ▶️ HH:MM orchestrate: <Topic>under## 💬 Sessionsin today's daily note. - Branch: in the target repo, work on a feature branch named for the change, never the default branch.
1. Dispatch rules
Each phase runs as the matching skill, .agents/skills/freddie-<phase>/SKILL.md, and follows that file exactly (boundaries, artifact names, templates).
- Preferred: an isolated subagent per phase. This keeps the conductor's context lean and gives each phase fresh eyes, which matters most for review. Brief the subagent with: the path of the phase skill to follow, the objective and success criteria, the artifact paths so far, the repo path and branch, and, for loop-backs, the specific findings to address. Use the phase's tier model:
opusfor plan, spec, review, and troubleshoot;sonnetfor research, test, and implement. Log the model actually dispatched in the ledger's Model column; a fallback gets its own Run Log row (same phase, next#). In Claude Code, that means the Agent tool withmodel. In Cursor, use a subagent if available. In Antigravity (agy), subagents are spawned automatically and can't be pinned to a model, so dispatch there when agy offers it; otherwise run the phase inline by reading its SKILL.md. When running inline, write each phase's artifact before starting the next, and never review your own work in the same pass without first re-reading the Spec cold. - Review must never be done by the agent that wrote the thing under review.
- Phase subagents skip their own "stop for approval" steps. You own every approval gate.
- After each phase, update the ledger before doing anything else. The ledger is the source of truth if the session dies.
2. The pipeline & its gates
| # | Phase | Skill | Gate to advance | On gate failure |
| 1 | Research | freddie-research | Open questions that block design are answered or explicitly assumed | Targeted follow-up research (max 1), then ask the user |
| 2 | Plan | freddie-plan | A plan-review pass by freddie-review (target = the Plan) has no P0/P1 | Revise the plan with those findings (loop ≤ 2) |
| 3 | Spec | freddie-spec | Every success criterion maps to ≥ 1 acceptance criterion, and every step names real files. Checked by a spec-review pass by freddie-review (target = Spec vs Plan) | Revise the spec (loop ≤ 2). A design-level flaw means going back to Plan |
| ⏸ | Checkpoint | you | The user approves Plan + Spec (skipped with --auto; in that case mark both Approved (auto)) | Apply the user's edits, re-running Plan or Spec as needed |
| 4 | Test | freddie-test (TDD) | Tests exist for every acceptance criterion and fail for the right reason | Fix the tests (loop ≤ 2) |
| 5 | Implement | freddie-implement | The spec's verification command passes, with real output captured | freddie-troubleshoot on the failure, then re-verify (loop ≤ 3) |
| 6 | Review | freddie-review (target = branch diff vs Spec) | Verdict is Ship, or only P2 findings remain | Route by root cause (below), then re-run Review (loop ≤ 2) |
Routing review findings (this respects the Plan → Spec → Implement gate; it never patches code directly from a review):
- Code does not match the spec, or has a bug: append a
## Amendment Nto the Spec with a focused fix step, then run Implement, then Review. - The spec is wrong or incomplete: revise the Spec (and tests), then Implement, then Review.
- The design is flawed: revise the Plan, then Spec, then Test, then Implement, then Review. Tell the user, because this is a big loop.
Tiny-change fast path: if framing shows the request is trivial (one file, no design choices), propose collapsing Research and Plan into a short Spec. This needs the user's OK, or --auto.
3. Circuit breakers (escalate, don't grind)
Stop, update the ledger with Status: Blocked, and ask the user when:
- any loop hits its cap;
- the same finding or failure recurs after being "fixed";
- a phase would need secrets, production access, destructive operations, or changes outside the stated repo;
- the success criteria turn out to be contradictory or need a product decision;
- the total number of phase dispatches exceeds 20.
When escalating, give: what was tried, the evidence, 2–3 options, and your recommendation.
4. Completion audit
Treat completion as unproven until evidence shows otherwise. For each success criterion and acceptance criterion, cite the evidence that proves it (a test name plus real output, path:line, or a command result), and mark it ✅ proven, ❌ failing, or ⚠️ unverified. Anything ❌ or ⚠️ means you are not done: loop back or escalate. Never mark a run done because you are stopping.
5. Finish
- Update the ledger (
Status: Done, the audit table, links to every artifact) and tick the dashboard tasks. Add follow-up tasks for P2 findings (#p3). - Log
- ✅ HH:MM orchestrate done: <Topic>in today's daily note. - Never merge or mark a PR ready. Offer to open a draft PR.
- Reply in chat with the objective, verdict, loops taken, the audit table, and what's left for the user.
6. Running under /goal (unattended)
/goal supplies the persistence (it keeps taking turns until an evaluator model judges the condition met), and this skill supplies the pipeline. The evaluator only reads the transcript and cannot open files, so:
- End every turn with one status line:
ORCHESTRATE STATUS: <Running|Checkpoint|Blocked|Done> · Phase: <phase> · Ledger: <path>, plus the audit table when the run is Done. - Each goal turn: re-read the ledger and do the next dispatch. Don't re-plan from memory.
- At the ⏸ checkpoint in non-
--automode, emitCheckpointand stop. The recommended goal condition treats that as terminal, so the loop hands control back to the user.
Recommended invocation (Claude Code, Cursor, and agy):
/goal Use the freddie-orchestrate skill on: <idea> [--auto]. Done when the last ORCHESTRATE STATUS line says Done and every success criterion in the ledger's Completion Audit is ✅ with cited evidence; also stop when it says Checkpoint or Blocked.Antigravity (agy)
agy has /goal too (it runs until completed or cancelled, with no turn cap), so use the same recommended invocation above. For a one-shot headless run, use agy -p="/freddie-orchestrate <idea> --auto". Either way, the ledger is the persistence layer: if a run stops mid-way, resume with /freddie-orchestrate 10 - Projects/<P>/Plans/Orchestration_<Topic>.md. agy ignores the model: field, so choose the model with /model or --model (see AGENTS.md §3).
Ledger template
# Orchestration: <Topic>
**Status**: Running | Checkpoint | Blocked | Done · **Phase**: <phase> · **Mode**: checkpoint | auto
**Repo**: `<path>` @ `<branch>` · **Started**: YYYY-MM-DD
## Objective
## Success Criteria
- [ ] SC1: …
## Artifacts
- Research: [[Research_<Topic>]] · Plan: [[Plan_<Topic>]] · Spec: [[Spec_<Topic>]] · Review: [[Review_<Topic>]]
## Run Log
| # | Time | Phase | Model | Result | Next |
| :- | :--- | :--- | :--- | :--- | :--- |
| 1 | HH:MM | Research | sonnet | ✅ gate passed | Plan |
| 2 | HH:MM | Implement | sonnet | ❌ rate limited | retry on another Tier 2 model |
## Decisions & Assumptions
## Completion AuditDoes it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

