t3-tools
Skill1 starsChanged 3 days agoMITView on GitHub
Use when the work involves T3 Code itself: linking or watching a pull request, delegating to another model or provider, waiting on a background command or test run, showing a chart, diagram or HTML visual in the thread, previewing a page or simulator, scheduling, or requesting a secret. Gives the exact T3 tool names and how to load or call them in this harness, so nothing is searched or guessed. Also use to explain visually, for an inline diagram, or when the user says show me.
- Commits and pushes
In this repository22 other files it gives its agents.
- analyze-logsSkill
- architectSkill
- build-audit-logsSkill
- cloudflareSkill
- code-reviewSkill
- design-engineeringSkill
- diagnosing-bugsSkill
- full-sendSkill
- full-stopSkill
- ghSkill
- grillingSkill
- impeccableSkill
- improve-animationsSkill
- interrogateSkill
- no-use-effectSkill
- prSkill
- retroSkill
- triageSkill
- unslopSkill
- vitestSkill
- wait-whatSkill
- writing-for-agentsSkill
The file
We split this skill into the rules it gives an agent
6 rules. About 0 are common, 2 are distinctive, and 4 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈2
Distinctive · estimate
This author's own choices. Worth a look.
4
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
- Distinctive: line 40
Every sub-agent model comes from the user's T3 favourites, unless a skill names one that is on that list.
- Very specific: line 47
Check the model with orchestrator_capabilities, passing its name as search.
- Very specific: line 52
If a model fails or is deprecated, take the next model in the same tier and say so.
- Distinctive: line 86
Write a PNG file only when the user asks for a file to post elsewhere.
- Very specific: line 90
Wait on a background command or test run with t3_background_wait, passing the background task ids.
- Very specific: line 97
Call link_pull_request with the PR URL right after gh pr create, before anything else.
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 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
T3 tools
Reach for the tool named here; do not search or guess.
Route per provider
- Claude Code: T3 tools are deferred. Load a family with one call:
ToolSearchwithselect:mcp__t3-code__<name>,.... Then callmcp__t3-code__<name>. - OpenCode: T3 tools are not top-level. Find one with
search({query: "<name>"}). Then call the returned path insideexecute:await tools["t3-code-<threadId>"].<name>({...}). Do not usetools.shell,read,webfetch,mcp__t3_code__..., or the dot form. - Codex: T3 tools are direct. Call
t3-code.<name>. No load step.
Families
| Family | Claude select: value | Tools |
| Delegate | mcp__t3-code__orchestrator_capabilities,mcp__t3-code__delegate_task,mcp__t3-code__task_status,mcp__t3-code__task_cancel | orchestrator_capabilities, delegate_task, task_status, task_cancel, create_threads |
| Threads | mcp__t3-code__t3_thread_search,mcp__t3-code__t3_thread_read,mcp__t3-code__t3_thread_list,mcp__t3-code__t3_thread_send,mcp__t3-code__t3_thread_wait,mcp__t3-code__t3_thread_launch | t3_thread_search, t3_thread_read, t3_thread_list, t3_thread_send, t3_thread_wait, t3_thread_launch, t3_thread_configure, t3_thread_update, t3_thread_organize, t3_thread_transfer, t3_thread_interrupt |
| PRs | mcp__t3-code__link_pull_request,mcp__t3-code__watch_pull_request,mcp__t3-code__unwatch_pull_request,mcp__t3-code__list_thread_pull_requests | link_pull_request(s), unlink_pull_request(s), list_thread_pull_requests, watch_pull_request, unwatch_pull_request |
| Background | mcp__t3-code__t3_background_wait | t3_background_wait |
| Visuals | mcp__t3-code__html_preview,mcp__t3-code__html_render | html_preview, html_render |
| Browser | mcp__t3-code__preview_open,mcp__t3-code__preview_navigate,mcp__t3-code__preview_snapshot,mcp__t3-code__preview_act,mcp__t3-code__preview_wait_for,mcp__t3-code__preview_status,mcp__t3-code__preview_evaluate | preview_*, t3_preview_list, t3_preview_close (browser work only) |
| Device | mcp__t3-code__device_list,mcp__t3-code__device_open,mcp__t3-code__device_screenshot,mcp__t3-code__device_close | device_* (taps and text go through the agent-device command that device_open returns) |
| Schedule | mcp__t3-code__schedule_task,mcp__t3-code__list_scheduled_tasks,mcp__t3-code__delete_scheduled_task,mcp__t3-code__run_scheduled_task_now | schedule_task and the matching list, delete, and run-now tools |
| Secrets | mcp__t3-code__request_secret | request_secret |
For cloud routines or one-time scheduled agent runs, use the schedule skill
instead. Those are not T3 tasks.
Choosing a model
Every sub-agent model comes from the user's T3 favourites, unless a skill names one that is on that list. This is the user's standing request to set a model on sub-agents, so it counts as asked. Do not fall back to the parent's model because a harness rule says to set a model only on request.
- Pick the tier the skill asks for from the table below. When a skill says "cheapest" use Fast, and when it says "strongest" use Strong.
- Check the model with
orchestrator_capabilities, passing its name assearch. The default result is a short list, so a missing name there proves nothing. - Set it with
delegate_tasktarget:providerInstanceId,model, andoptionswith the option shown in the table. Native sub-agent tools run on the parent's model, so use them only when that is the right pick. - If a model fails or is deprecated, take the next model in the same tier and
say so. If the tier is used up, read
favoritesin~/.t3/userdata/settings.jsonand pick the closest match. Never go outside that list. Ignoreclient-settings.json, which is older.
Tiers
| Tier | Use for | Models, in order (providerInstanceId / model / option) |
| Fast | Cheapest and fastest. Scouts, builders, mergers, code-review | codex / gpt-6-luna / reasoningEffort: high; claudeAgent / claude-haiku-5-5 / effort: medium |
| Balanced | Medium reasoning, still fast. A job a Fast model got wrong, or that needs more thought than Fast gives | opencode / opencode-go/mimo-v2.6-flash / no option |
| Strong | Design and adversarial review | claudeAgent / claude-opus-5-5 / effort: medium; opencode / opencode-go/muse-spark-1.3-contributor / variant: high; opencode / opencode-go/deepseek-v4.1-flash / variant: high |
| Cheap strong | Strong work that runs many times, such as review rounds | opencode / opencode-go/deepseek-v4.1-flash / variant: high; opencode / opencode-go/muse-spark-1.3-contributor / variant: high |
Opus, Muse Spark, and DeepSeek are three different model families. When a skill asks for models from different families, take them from Strong.
Parameters that fail when wrong
| Tool | Rule |
html_preview | html is the page text, not a path. |
html_render | title and height are required. |
delegate_task | mode is "async" or "wait". Use a distinct clientRequestId. |
preview_resize | preset is a device name such as iphone-se. |
run_scheduled_task_now | Pass taskId, not scheduledTaskId. |
t3_thread_read | afterPosition is 0 or greater. |
t3_queue | Use action (list, read, edit, cancel, reorder, promote_to_steer). |
Visuals
Show a chart, diagram, or page in the thread: call html_preview with html,
then html_render with title and height, before the final reply. Keep the
visual inline in the reply. Never open an HTML file, and never start a server.
Write a PNG file only when the user asks for a file to post elsewhere.
Background waits
Wait on a background command or test run with t3_background_wait, passing the
background task ids. On OpenCode, wait with t3_thread_wait (or
t3_background_wait if the search returns it). Never wait by polling with Monitor, sleep, or
setTimeout. execute has no setTimeout or sleep.
PRs
Call link_pull_request with the PR URL right after gh pr create, before
anything else. Call watch_pull_request only when this thread owns the PR and is
babysitting it. Call unwatch_pull_request before any handoff or question.
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.

