Context.
Files, commands, dependencies, and conventions.
How it works
Plan. Act. Observe. Reflect.
Picasso is a coding-agent harness built around the loop engineers already use. It studies the repo, chooses a path, acts through tools, reads the result, and keeps going when the next step is clear.
Composition.
Repair the expired-session test and keep the auth suite green.
Changes
EDITsrc/auth/session.tsenforce one-hour max age.
CREATEsrc/auth/session.test.tscover expired and current tokens.
Verification
Run `pnpm test src/auth/session.test.ts` and report any remaining failures.
⏎approveeeditescreject
∠maintools4ctx28k / 200kcost$0.00modeplanwaiting
First, it understands the work.
Picasso reads the repo, the request, relevant files, tests, scripts, configs, and useful memory. For larger tasks, Plan mode turns that context into a reviewable composition before edits begin.
agent loopplan / act / observe / reflect
01Planread repo, draft plan
02Actedit files, run tools
03Observetests, logs, diffs
04Reflectcontinue or stop
completion criteria
Memory.
Facts, style, and decisions retrieved when relevant.
Review.
A written plan before serious changes.
Then it works through tools.
Picasso acts through declared tools. It edits files, runs shell commands, reads output, searches the repo, calls MCP servers, opens subagents, and records what happened.
File tools.
Scoped reads, writes, diffs, and patches.
Shell tools.
Tests, builds, linters, formatters, diagnostics, and scripts.
MCP.
External systems connect through native client and server support.
Hooks.
Project policy can run around plans, tools, sessions, and goals.
It learns from the result.
After each stroke, Picasso observes. It reads errors, test output, command results, and changed files. Then it reflects and chooses the next step.
terminalpicasso / repo
›
picasso plan "repair failing auth tests"readsrc/auth/session.ts84 lines
editsrc/auth/hash.ts+18 / -9
testpnpm test auth7 passed
mode safetools 3tokens 12.4Kbudget visible
Failure aware.
A failed command becomes input.
Test guided.
Checks tell the agent what is actually true.
Stateful.
The canvas keeps the path intact.
Parallel work, when it helps.
Subagents are named brushes with isolated context. One can inspect tests while another studies a module. Results return to the main session for a single decision.
Named roles.
Explorer, worker, reviewer, or project-specific brushes.
Isolated context.
Each subtask stays focused.
Controlled merge.
The main canvas integrates the result.
Sponsors support managed usage. They do not enter the loop.
Sponsors never appear in code, model output, tool results, plans, routing, memory, or subagent selection. They support access through labeled surfaces outside the agent's work.
trust boundaryoutside the agent loop
generated codeclean
model outputclean
tool resultsclean
plan documentsclean
routingnot influenced
memoryretrieval only
Outside output.
No sponsor text in generated code or model responses.
Outside tools.
Tool results remain factual.
Outside routing.
Sponsors do not steer models, memory, tools, or brushes.