ai-loops / platforms
CLAUDE.md0 starsChanged 7 days agoMITView on GitHub
Read with skills/work/SKILL.md. Facts: specs/PLATFORMS.md §2-4. Verify: loops version resolves (the root shim builds from the checkout; needs Go), then /loops:init. Other language packs: the toolchain-env hook prefixes their build and test commands with the run-scoped env (loops lang show <id> → cache_env).
In this repository35 other files it gives its agents.
- AGENTS.mdAGENTS.md
- CLAUDE.mdCLAUDE.md
- llms.txtllms.txt
- clarifySkill
- consistencySkill
- debuggingSkill
- documentationSkill
- featureSkill
- hexagonalSkill
- initSkill
- jobs-and-state-machinesSkill
- lang-clojureSkill
- lang-cppSkill
- lang-csharpSkill
- lang-elixirSkill
- lang-erlangSkill
- lang-goSkill
- lang-javaSkill
- lang-kotlinSkill
- lang-phpSkill
- lang-pythonSkill
- lang-rubySkill
- lang-rustSkill
- lang-scalaSkill
- lang-swiftSkill
- lang-typescriptSkill
- memorySkill
- parallel-workSkill
- performanceSkill
- qualitySkill
- reportSkill
- review-integrationSkill
- statusSkill
- testingSkill
- workSkill
The file
We split this claude.md into the rules it gives an agent
4 rules. About 0 are common, 1 is distinctive, and 3 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈1
Distinctive · estimate
This author's own choices. Worth a look.
3
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 3
Read with skills/work/SKILL.md.
- Very specific: line 12
Put every Agent call of a wave in ONE message.
- Distinctive: line 31
read-guard and model-guard (PreToolUse deny), post-edit (the pack's formatter and linter, e.g.
- Very specific: line 43
Verify: loops version resolves (the root shim builds from the checkout.
≈ “Common” is an estimate: at least 1 in 5 of the 27,641 CLAUDE.md 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
Platform: Claude Code (tier A: resumable subagents)
Read with skills/work/SKILL.md. Facts: specs/PLATFORMS.md §2-4.
Spawn
Agenttool withsubagent_type: "loops:<role>"(for exampleloops:domain-engineer) andrun_in_background: true.- Pass
model(haiku|sonnet|opus) only to escalate, under--models=cheap, or for a fallback. - Fallback (WORKFLOW §3a). If a spawn fails because the role's model isn't available (an error naming the model, entitlement or capacity), re-spawn the same role with the next-best
model. Reasoning roles go opus → sonnet →inherit; others go one tier down. Record amodel-fallbacknote, and tell the user when the planner ran below opus. - The prompt is the header (Run dir, Specs, Templates, Role, Agent id, Wave) plus the agent's plan block.
Parallel wave
Put every Agent call of a wave in ONE message. Research (3 agents) is one message too.
Wait
Completion notifications arrive on their own. Do not poll. Do not enter plan mode while agents run.
Loop-backs (resume)
SendMessage to the same agent (use the agentId from the Agent tool result) with the QA or review file path. The agent keeps its context. If it was stopped, spawn anew with its report and a ≤10-line history.
Labelling lanes
Right after each spawn, in the same turn as the spawn result:
loops run label --agent <agentId from the Agent tool result> --name <Agent id> --role <role> --wave <n>
Ask the user
AskUserQuestion (Approve / Request changes / Stop; Push branch?; how to proceed after limits).
Stop agents
TaskStop on the agent id once its lane is VERIFIED and it will not be resumed, and at run end.
Hooks and guards (all native)
- read-guard and model-guard (PreToolUse deny), post-edit (the pack's formatter and linter, e.g. gofmt + go vet), trace and trace-agent (SubagentStart/Stop lanes), stop-gate (blocks finishing with changed source files of any pack;
LOOPS_GATE=offis the user's escape hatch), session-start context. - Role tool restrictions come from agent frontmatter.
Live view
- Status line (
/loops:statusinstalls it) and the VS Code/terminal live pane. loops watchin a side terminal andloops report(HTML) also work.
Install and verify
/plugin marketplace add arxdsilva/ai-loops
/plugin install loops@ai-loopsVerify: loops version resolves (the root shim builds from the checkout; needs Go), then /loops:init.
Hardware and budgets
The hardware-researcher runs as a normal subagent (loops hw --write). Token budgets (--budget) are measured from agent transcripts on this platform.
Update
claude plugin marketplace update ai-loops && claude plugin update loops@ai-loops (or run loops update --platform claude).
Applied automatically at work start: <plugin root>/bin/loops update --apply --platform claude (PLATFORMS §9). Reload with /reload-plugins or a new session.
Run-scoped toolchain env
Go commands on this platform are rewritten automatically (the hook rewrites the command; a lone go build|test|vet|list is auto-approved). Either way they use the run's own Go cache, which loops run end deletes.
Other language packs: the toolchain-env hook prefixes their build and test commands with the run-scoped env (loops lang show <id> → cache_env).
Does 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.

