davila7/claude-code-templates / cli-tool/components/skills/development/pudu-task-telemetry
pudu-task-telemetry
VerifiedMeasure local AI task latency, token usage, errors and verified outcomes using Pudu AI hardware evidence and installed Ollama models. Use when comparing local task runs, choosing a local model for a bounded subtask, or recording reproducible task telemetry.
| File | Kind | Size |
|---|---|---|
| SKILL.md | skill md | 4.9 KB |
| references/examples.md | markdown | 3.2 KB |
| references/model-selection.md | markdown | 3.0 KB |
| references/setup.md | markdown | 3.0 KB |
| references/task-schema.json | config | 5.7 KB |
| references/telemetry-contract.md | markdown | 4.4 KB |
| scripts/lib/common.mjs | script | 1.8 KB |
| scripts/lib/model-selection.mjs | script | 3.5 KB |
| scripts/lib/ollama-adapter.mjs | script | 10.5 KB |
| scripts/lib/pudu-adapter.mjs | script | 4.7 KB |
| scripts/lib/report.mjs | script | 4.6 KB |
| scripts/lib/task-store.mjs | script | 8.4 KB |
| scripts/lib/telemetry.mjs | script | 3.0 KB |
| scripts/pudu-task.mjs | script | 12.7 KB |
references/examples.md
View on GitHub# Inputs and workflow
Create `task.txt` describing the bounded task. Create `request.json`, for example:
```json
{
"schemaVersion": 1,
"messages": [{"role": "user", "content": "Reply with exactly the uppercase word READY and nothing else."}],
"options": {"temperature": 0, "num_ctx": 4096, "num_predict": 32},
"limits": {"timeoutMs": 120000, "maxResponseBytes": 4194304},
"contextManifest": [],
"verification": {"type": "exact-text", "expected": "READY"},
"loadPolicy": "uncontrolled"
}
```
This is a smoke check, not a coding-quality benchmark. Exact-text verification
trims outer whitespace; it never executes output. Omit `verification` for work
that needs the host's review. Only role/content text messages and the documented
generation options are accepted; images, tools and arbitrary endpoints are not.
For code tasks, put the required snippets in `messages` explicitly. Optional
manifest entries contain relative `path`, file `sha256`, `startLine`, `endLine`.
They are hashed for equivalence, not automatically read from disk. Prompts,
expected answers and manifest paths are not persisted in telemetry.
The preflight context bound uses UTF-8 input bytes plus a template reserve,
not a measured token count. If it exceeds `num_ctx - num_predict`, explicitly
reduce the input or increase the configured window within model capacity. The
runner never silently truncates input. Model-specific tokenization/templates
can still differ; inspect runtime behavior when a task requires exact context
occupancy. Unknown metrics remain unavailable.
## Commands
Use the script path relative to the installed skill. Replace placeholders with
values returned by the commands:
```bash
node .claude/skills/pudu-task-telemetry/scripts/pudu-task.mjs start --task-file task.txt --json
node .claude/skills/pudu-task-telemetry/scripts/pudu-task.mjs run-local --config local-config.json --task-id TASK_UUID --request-file request.json --model INSTALLED_MODEL --output result.txt --json
node .claude/skills/pudu-task-telemetry/scripts/pudu-task.mjs finish --task-id TASK_UUID --status completed --json
node .claude/skills/pudu-task-telemetry/scripts/pudu-task.mjs report --task-id TASK_UUID --format markdown
node .claude/skills/pudu-task-telemetry/scripts/pudu-task.mjs compare --task-ids TASK_UUID_A,TASK_UUID_B --json
```
Relative inputs and output resolve against `--repo` (default current directory).
`result.txt` must not exist. To retry, choose a new output filename and pass
`--retry-of` with a prior attempt ID whose request hash is identical.
## Host-reported verification
After reviewing the result or running authorized project checks, create a file:
```json
{
"schemaVersion": 1,
"status": "passed",
"source": "host_reported",
"rubricId": "code-summary-v1",
"checks": [{"id": "inputs-described", "passed": true}, {"id": "outputs-described", "passed": true}]
}
```
Pass it to `finish --verification-file verification.json`. This sample contains
illustrative results only; replace them with checks actually performed. The
status must match all check results; no checks means `unverified`. Identifiers
are hashed in persisted telemetry. External input cannot claim `source: runner`.
A host-reported pass does not qualify a model for automatic selection.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| slack-gif-creatoranthropics/skillsKnowledge and utilities for creating animated GIFs optimized for Slack. Provides constraints, validation tools, and animation concepts. Use when users request animated GIFs for Slack like "make me a GIF of X doing Y for Slack." | Stars: 197k | Updated | Claude CodeCodex | Verified | Install |
| ponytailDietrichGebert/ponytailLazy senior dev mode for any coding task: the smallest change that fully solves it, and a reply a busy human understands in one read. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-debtDietrichGebert/ponytailHarvest every shortcut: comment into one debt ledger, so deferrals get tracked instead of forgotten. One-shot report. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-debtDietrichGebert/ponytailList every `shortcut:` comment (and older `ponytail:` ones) as a debt ledger. One-shot report, changes nothing. Use for "ponytail debt", "what did ponytail defer", "list the shortcuts", /ponytail-debt. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |