Illustrative mission, phase by phase
One change. Four repositories. One mission.
An illustrative run across four repositories, from the goal to the record it leaves.
One mission can coordinate AI coding agents across multiple repositories. Each specialist works in its own worktree, with the browser and terminal attached to the same goal.
Mission manifest
rl-311 · per-tenant rate limiting
Objective
Introduce per-tenant rate limiting across the public API surface.
──no breaking change for existing clients
──limits configurable per plan tier
- registered checkouts
- 4
- work mode
- Orchestrate
- access
- Ask for approval
- agents
- 4 specialists + reviewer
Illustrative product walkthrough, not measured customer evidence6 steps below
Step 1 · Discover
Read the workspace before touching it.
The lead agent reads first: which checkouts the goal reaches, and what earlier missions learned. This mission runs at Ask for approval, so every edit or command waits for you.
- access
- Ask for approval · edits stop for you
- impact map
- observed before planning
- found
- registered checkouts · relevant memory
enforcement point
plan tiers · usage
usage dashboard
redis counters
build + deploy matrix
client retry behavior
Step 2 · Plan
One plan across four repositories.
One plan covers every checkout and pins the interface all four depend on. The new Redis cluster waits for your decision. To approve the whole plan before work starts, choose Plan mode.
- workstreams
- 4 · one per repository
- contract
- rate-limit headers · pinned v1
- approval
- Ask for approval · you decide
Pinned interface · v1
RateLimit-Limit · RateLimit-Remaining · RateLimit-Reset · 429 + Retry-After on exceed · limits.yaml per plan tier
approval requested · new redis cluster · est. $84/mo
waiting on youexample approval · scope pinned · work continues
approvedStep 3 · Delegate
One specialist per repository.
In Orchestrate, each specialist gets its checkout, its slice of the goal and the pinned contract, in its own git worktree. Specialists can run on Codex, Claude Code, Cursor Agent or Skyflo's own agent.
- agents
- 4 specialists + 1 reviewer
- reviewer
- read-only · cannot edit
- isolation
- own worktree per specialist · shared contract
Lead agent
Holds the goal and the pinned contract
api-gateway
Codex
middleware + headers
model = "frontier"
billing-service
Claude Code
tier limits + events
model = "frontier"
platform-infra
Skyflo
redis · alerts · IAM
model = "fast"
web-console
Cursor Agent
usage UI
model = "fast"
Step 4 · Implement
Four worktrees, kept apart while the work runs.
Each specialist's edits stay in its own Git worktree, out of yours and out of each other's, and the lead owns the seam between them. Bringing them together is still a change you review: worktrees keep edits apart, they do not settle a contract the repositories disagree on.
- changes
- 4 bounded worktrees
- integration
- the lead owns the seam
- execution
- your Mac · registered repositories
Step 5 · Review
A reviewer with no write access.
Orchestrate added a reviewer here. It can read every worktree, ask you a question and report what it finds. It cannot change a line.
- reviewer
- independent · mutation disabled
- tools
- file.read · file.search · user.ask · agent.update · agent.report
- when
- Orchestrate · by strategy and risk
file.read
reviewer can inspect the implementation
file.search
reviewer can trace relevant references
user.ask
reviewer can ask you a question
agent.update · agent.report
progress and findings return to the mission
mutationAllowed: false
reviewer cannot change the files
Step 6 · Remember
Three records, each with its source.
When the work finishes, Skyflo drafts what the mission taught, each record linked to its source. Records backed by the work's own evidence are kept; anything weaker waits for you.
- records
- scoped · cited · removable
- recall
- retrieved when relevant
- compounds
- proven twice · becomes a Skill
Public API rate limits enforced at api-gateway via token bucket in Redis. Limits per plan tier live in platform-infra/limits.yaml.
429 responses must include Retry-After. SDK v2 clients retry automatically; v1 clients surface the error.
platform-infra owns the Redis cluster. billing-service owns plan-tier limit definitions.
❯recall“how do I raise limits for an enterprise tenant?”
→ retrieved from learned mission memory in a later session.
example finished·six steps shown·one objective, four repositories, one review report
Other missions that span systems.
Cross-repository migrations, coordinated upgrades, security responses, and domain splits all need one goal held across several systems so the change lands consistently.
Move checkout workloads to ARM without interrupting traffic
6 systems · one mission
Introduce tenant-level encryption across the platform
4 systems · one mission
Upgrade the framework major version in every affected service
8 systems · one mission
Patch a critical CVE everywhere it is actually reachable
8 systems · one mission
Split the monolith's billing domain into its own service
5 systems · one mission
Instrument p95 tracing across the request path
7 systems · one mission
How missions behave.
Scope, isolation, and review. For what a mission keeps once it closes, see how Skyflo learns between missions.
What is a mission?
A mission is one goal with its conversation, agents, approvals, and record. It persists on your Mac: when the work stops it goes idle rather than ending, so follow-ups go in the same place instead of a new chat.
Can one mission span repositories?
Yes. A workspace is an ordered set of registered Git checkouts, up to eight, and one mission may span them. A mission in Skyflo Cloud can span one to five.
Do I have to approve a plan first?
Only if you choose Plan mode. This example waits for approval because it runs in Orchestrate, which starts at Ask for approval. In Auto, the agent edits and runs straight away, with Full access for a new Auto mission unless you pick another access mode.
Can other coding agents work across repositories?
Yes. Several coding agents and orchestration tools can work across repositories. Skyflo's difference is the mission record that keeps the plan, the specialists' work, approvals, review and source-linked memory together, with Codex, Claude Code, Cursor Agent, and Skyflo's own agent able to work on the same mission in Orchestrate.
Does Skyflo open pull requests?
A mission you start in Skyflo Cloud from a connected GitHub repository opens a pull request into the base branch you chose when it finishes. Skyflo Cloud is on paid plans. Missions on your Mac do not open pull requests on their own; their changes stay in your checkout or the specialists' worktrees until you push.
How is the work reviewed?
In Orchestrate missions, Skyflo adds a separate reviewer by strategy and risk, and Skyflo's own agent can start one. The reviewer can read files, search, ask you a question, and report findings, but cannot edit files or run mutating operations. Not every mission is reviewed.
Run one of these on your own repositories.
Skyflo Desktop is free with an agent you already use or your own Moonshot or Z.ai key. Add the checkouts a change touches and give the mission its goal.
Missions run on your Mac, or in Skyflo Cloud on a paid plan.