telemetry-report
Skill19 starsChanged 7 days agoMITView on GitHub
Build the local agent-telemetry dashboard for this project from the session logs coding agents already keep: tokens per day, task, skill and session, and before/after medians around context changes. On request, publish a counters-only monthly report to the cfd-telemetry branch so a central dashboard can read it. Use when asked about agent usage, token consumption, or whether a context change made sessions cheaper.
In this repository10 other files it gives its agents.
- CLAUDE.mdCLAUDE.md
- cfd-upgradeSkill
- issue-newSkill
- issue-startSkill
- new-decisionSkill
- review-prSkill
- run-siiSkill
- session-closeSkill
- session-startSkill
- validate-contextSkill
The file
We split this skill into the rules it gives an agent
5 rules. About 0 are common, 1 is distinctive, and 4 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.
4
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 10
Run the bundled scripts. Do not read scripts/telemetry.py, scripts/central.py or scripts/dashboard.html into context: running them costs …
- Very specific: line 20
If this skill was installed somewhere else (a plugin, a user-level skills folder), run scripts/telemetry.py from the folder that holds this …
- Very specific: line 25
Relay the summary, then the gaps: Every note line is something the script could not read or had to infer.
- Very specific: line 29
Keep the output per-developer: If the script warns that .cfd/ is not git-ignored, add .cfd/ to .gitignore.
- Distinctive: line 97
Adherence first: Read "context at first request" and "tokens before first edit" before any total.
≈ “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
Run the bundled scripts. Do not read scripts/telemetry.py,
scripts/central.py or scripts/dashboard.html into context: running
them costs nothing, reading them costs ~21,000 tokens.
Build the dashboard:
python3 .agents/skills/telemetry-report/scripts/telemetry.py --openIf this skill was installed somewhere else (a plugin, a user-level skills folder), run
scripts/telemetry.pyfrom the folder that holds this file, and pass the project directory as its argument. It writes.cfd/telemetry/index.htmland prints a summary.- Relay the summary, then the gaps. Every
noteline is something the script could not read or had to infer. Report those as plainly as the totals: a format change in an agent shows up there first. - Keep the output per-developer. If the script warns that
.cfd/is not git-ignored, add.cfd/to.gitignore. Publish only when the user asks (
/telemetry-report publish):python3 .agents/skills/telemetry-report/scripts/telemetry.py --publishThis commits
reports/<git user.email>/<YYYY-MM>.jsonto the localcfd-telemetrybranch through git plumbing; the checkout is not touched. Show the usergit show --stat cfd-telemetry, state what leaves the machine (below), and wait for an explicit yes before re-running with--publish --push. Never force-push that branch.Never publish from a public repository. The branch would show per-person usage and git identities to everyone. The script refuses when the forge reports the repo as public; when it cannot tell (it prints a warning), ask the user and do not push until they confirm the repo is private. Do not pass
--allow-publicon your own.Team view, only when the user asks (
/telemetry-report team):python3 .agents/skills/telemetry-report/scripts/central.py <repo> [<repo> ...]Each
<repo>is a local clone or a git URL whosecfd-telemetrybranch holds published reports. It writes.cfd/team/index.html: the same dashboard with contributor and repository filters, and a footer naming who has not reported in--stale-days(14). The page names people: keep it out of git and do not paste per-person figures into issues or PRs. Contributors appear alphabetically; never order or compare them by consumption (agent-telemetry-from-local-logs, rule 7).
What the data source is, per agent
The skill installs the same way for every agent (.agents/skills/,
with .claude/skills as a symlink). What differs is where each agent
keeps its numbers:
| Agent | Source | To enable |
| Claude Code | ~/.claude/projects/ transcripts | Nothing |
| Codex | ~/.codex/sessions/ rollouts | Nothing |
| Copilot CLI | ~/.copilot/session-store.db, its usage table only | Nothing. Versions that record no usage (the summary says so) need a CLI update; their earlier sessions stay uncounted. Never open the store or session-state/ yourself: they hold prompts and responses. |
| Copilot (VS Code) | OpenTelemetry file export | Ask the developer to set three VS Code user settings (never change them yourself): github.copilot.chat.otel.enabled to true, github.copilot.chat.otel.exporterType to file, github.copilot.chat.otel.outfile to an absolute path outside any repo; reload the window; run one short agent-mode session in a git repo with a GitHub remote, one turn editing a file; report the outfile path back. The script finds that export and says whether one exists, but does not read it yet: this starts the history. |
| Gemini CLI | Telemetry outfile in .gemini/settings.json | No adapter yet |
Any agent can run this skill and report on every agent's sessions:
the script reads logs, it does not ask the running agent for anything.
Copilot CLI and Copilot Chat are separate agents (copilot-cli,
copilot); COPILOT_OTEL_FILE_EXPORTER_PATH is shared by both, so the
script never attributes an export to Copilot Chat through it.
What a published report contains
Counters per session: tokens by type, request and turn counts, active
minutes, context size, the day (never the time of day), branch names,
skill names, model names. The developer is the git identity, so a task
shared by several people adds up per person and per agent. Never
prompts, responses, file contents or paths. --anonymous replaces the
identity with a random id; forge admins can still see who pushed.
Schema: references/report-schema.md.
Rules for reading it
- Adherence first. Read "context at first request" and "tokens before first edit" before any total. Totals say how much was spent; those two say whether the context is doing its job.
- Before/after rows with fewer than 3 sessions on a side are anecdotes. Say so; do not draw a conclusion from them.
- Run it at least monthly. Agents delete their own logs (Claude Code after 30 days); the script keeps a local history, but only of what it saw.
- Never rank people. Per-person numbers explain a task's cost; they do not measure the person.
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.

