Small model
For short steps with one clear answer: sorting, renaming and reading test output.
- Sort breaking changes by dependency
- Apply a codemod and read the test output
Big problems. Clear steps.
Break the work into steps, give each step the skills and tools it needs, then check the result before moving on.
1Find
Look for relevant guidance and reusable code
2Choose
Check that it fits the task and your permissions
3Prepare
Give the step its selected files and instructions
4Work
Use an appropriate model, tool or existing code
5Check
Test the result against your requirements
Follow one task, step by step.
Pick a step to see what it reads, the files it pulls and the check it must pass.
Upgrade an app to the next major framework version
Step {{stepNum}} of 4FeedsProMax
Understand the input
Map every breaking change in the new version onto the code it touches.
What this step reads
package.jsonand your lockfile- Release notes and breaking changes, from the Agent Feeds collection Plan platform and tool upgrades
- Your test and build commands
Files it pulls
- breaking-change-mapSkill, MIT, 5.8 KBsha256 9e2d41c0…
- lockfile-diffCode module, Apache-2.0, 2.9 KBsha256 c07b5a13…
What it can do
Read the repository, run read-only commands and write one file.
What you get
upgrade-plan.md, with each breaking change, the files it touches and its change set.
Check it must pass
- Every breaking change in the release notes appears in the plan
- Each change set fits in one step and names the tests that cover it
- An independent check accepts the plan before step 2 starts
Choose an approach
Put the change sets in a safe order and give each one a model of the right size.
What this step reads
upgrade-plan.mdfrom step 1- Benchmark, cost and latency sources, from Choose a coding model and Compare model cost and latency
- The call limit and time limit you set
Files it pulls
- model-routing-settingsHarness settings, MIT, 2.1 KBsha256 71c3e9a4…
- change-set-plannerSubagent, Apache-2.0, 3.6 KBsha256 d24f0b87…
What it can do
Read the plan and the sources, then write one file.
What you get
run-order.md, with each change set in order, its model size and its call limit.
Check it must pass
- Every change set in the plan has a place in the order
- Each change set names a model size and a call limit
- An independent check accepts the order before step 3 starts
Build or reuse
Apply one change set at a time with its codemod, and run your tests after each.
What this step reads
run-order.mdand the change set for this step- The codemod notes for your framework version
- Your test command
Files it pulls
- framework-migrationSkill, MIT, 7.8 KBsha256 9c2e5b14…
- upgrade-codemodsCode module, Apache-2.0, 3.1 KBsha256 5a07d3e1…
- test-after-changeHook, MIT, 1.2 KBsha256 e81b6c29…
What it can do
Edit the files the change set names and run your tests, with the permissions you choose.
What you get
One commit for the change set, with its test results.
Check it must pass
- The codemod applies cleanly to the files it names
- Every test passes after the change
- An independent check accepts the change set before the next one starts
Check the result
Run the full suite on the merged change sets and queue the rest for the night.
What this step reads
- The commits and test results from step 3
- The open change sets in
run-order.md - The call limit and time limit for the night
Files it pulls
- coverage-reportHook, MIT, 1.9 KBsha256 3d6a2f58…
- morning-report-templateInstruction file, MIT, 1.1 KBsha256 b90c4e17…
What it can do
Run your full test suite, queue the open change sets and write the morning report.
What you get
A morning report: merged, ready for review, waiting for you.
Check it must pass
- The full suite passes on every merged change set
- Each queued change set starts from a task checkpoint
- The morning report lists every change set and its state
Match each step to the right model.
Route small steps to small models and keep the large model for the hard part. Every call runs on your own model access.
Mid-size model
For scoped edits across a few files, with tests to guide each change.
- Apply one change set and update its tests
- Write the summary for each change set
Large model
For the hard part: planning, unfamiliar code and changes that span many modules.
- Plan the upgrade path across the codebase
- Rework a module the codemods leave for review
Where does your agent spend its effort?
Tick each one your setup already does. Each answer shows where Baltor helps.
Put knowledge, code and experience to work.
Each step draws on four kinds of knowledge. Files from the library carry the first two. Earlier runs and your own feedback add the other two.
Browse the libraryYour next step draws on all four
From the library
Context Intelligence
Methods, questions, constraints, examples and output contracts that guide the work.
Code Intelligence
Reusable functions, tools, files and workflows with declared inputs, outputs and effects.
From earlier runs and from you
Runtime History and Solution Intelligence
Outcomes, repairs and measurements from earlier runs, ready to inform a new task.
User Feedback Intelligence
Corrections, priorities and instructions from a person, scoped to the work.
Get useful material. Keep control of your work.
Read Access and dataYour model access stays yours
Your agent calls your models through the provider connection you control. Your project and model keys stay on your machine.
Exact versions with checksums
Each download names its source, licence and exact version, with a SHA-256 checksum that proves its bytes.
Your agent decides what runs
A download is a file in your folder. Your agent runs it with the permissions you choose, so look over unfamiliar code first.
Give your next big task clear steps.
Available today. With an account, you create a Baltor key for each tool, search the library from your agent, download exact versions and see your usage.
Feeds is free through December 31, 2026 (Eastern). Max is in Public preview. Savings depend on the task, the model and the setup.