Skip to content

notque/vexjoy-agent / skills/meta/do

Classify user requests and route to the correct skill, pipeline, and stack. Primary entry point for all delegated work.

references/routing-telemetry.md

View on GitHub
# /do routing telemetry

Hooks record routing automatically. The router records one case by hand: an observed route failure (below).

| Capture | Hook | Event |
|---|---|---|
| Routing decision (`{agent}:{skill}`) + right-sizing feedback | `routing-decision-recorder` | PostToolUse:Agent |
| Outcome — validate pending | `routing-outcome-recorder` | SubagentStop |
| Outcome — finalize (boost/decay) | `routing-outcome-finalizer` | UserPromptSubmit |
| Outcome — errored and stale (over 6 hours) fallback | `routing-outcome-stop-fallback` | Stop |
| Outcome — resolve every pending entry | `routing-outcome-stop-fallback` | SessionEnd |
| Review findings | `review-capture` | PostToolUse:Agent |

These feed the route health report:

```bash
python3 ~/.claude/scripts/learning-db.py route-health
```

**Outcome fidelity.** The next-turn finalizer scores acceptance: on the next user turn, deterministically and at zero LLM cost, three ways: failure on errors or rejection (decay); success on explicit acceptance (boost); neutral otherwise. Silence counts as neutral, never as acceptance. `Stop` fires after every turn, before the next prompt exists, so there the fallback scores only two kinds of entry: an errored dispatch as failure at once, and a clean one as neutral after 6 hours with no user turn. It leaves every other entry pending for the finalizer. On `SessionEnd` no next prompt can come, so the same hook resolves every entry still pending: errored as failure, clean as neutral. `.claude/settings.json` registers the hook under both `Stop` and `SessionEnd`.

**Report route failures.** Report only a failure you are confident about:

```bash
REASON_FILE=$(mktemp); printf '%s' "<cause>" > "$REASON_FILE"
python3 ~/.claude/scripts/learning-db.py route-failure AGENT:SKILL --reason-file "$REASON_FILE" --routing-relevant yes --session $SESSION --marker $DISPATCH_ID
rm -f "$REASON_FILE"
```

Triggers: re-route, lazy re-dispatch, validator misroute, harness reject. A right route with bad execution takes `--routing-relevant no`. Skip ambiguous cases. One decay per dispatch key. The temp file keeps the reason text off the shell command line.

---

## Re-deriving routing figures

SKILL.md quotes no telemetry figures since the specialist agents retired (plans/009). Before quoting one, measure it again: a figure is a DATED observation, not an invariant, and no test pins it, because pinning a live DB read would make CI depend on runtime data.

Source: `evidence_route_decisions` in `~/.claude/learning/learning.db` (two other learning.db files exist and lack this table). Read-only.

**The agent column after retirement.** `agent` reads `main` (work run in the /do thread) or `general-purpose` (the normal worker) on new rows, and the schema is unchanged. A `general-purpose` share is no longer a fallback rate: it is about 100% of dispatches by design. Rows before the retirement still name specialist agents, so split any trend at the retirement date. Count real fallbacks from the marker's `fallback=` token instead (the recorder's fallback reason; after retirement `build-dispatch.py` emits it only for a coerced agent name, as the `invalid` reason with the name appended):

```bash
DB=~/.claude/learning/learning.db
sqlite3 "$DB" "SELECT agent, COUNT(*) FROM evidence_route_decisions GROUP BY agent ORDER BY 2 DESC;"
python3 ~/.claude/scripts/learning-db.py route-health   # fallback reasons and route health
```

State the metric as the query computes it. "Named domain skill, fallbacks excluded" is reproducible; "correct skill" is not.

**Combination depth — the definition decides the number.** Depth counts filled slots with fallbacks excluded: skill named and other than `workflow` (the fallback skill; `process` before #993), pipeline present, stack present. The agent slot no longer counts: every dispatch names `general-purpose`, so counting it would score every route one deeper for no routing work. Counting any value present instead of excluding fallbacks returned a healthier number for identical data (2.48 vs 2.01 on the 2026-08-15 rows, when the agent slot still counted), which is why a depth figure quoted without its definition is worthless.

```bash
sqlite3 "$DB" "SELECT ROUND(AVG(depth),2), ROUND(100.0*SUM(depth<=1)/COUNT(*),1) FROM (
  SELECT (skill IS NOT NULL AND skill NOT IN ('','-','workflow'))
       + (pipeline IS NOT NULL AND pipeline NOT IN ('','-'))
       + (stack IS NOT NULL AND stack NOT IN ('','-','[]')) AS depth
  FROM evidence_route_decisions);"
```

Report depth against a REACHABLE maximum (3 slots). On the 2026-08-15 rows `pipeline` and `matched_components` were NULL on all 301, so those rows cannot exceed 2. Confirm the ceiling before quoting the mean.
writing-skillsobra/superpowersUse when creating new skills, editing existing skills, or verifying skills work before deploymentStars: 322kUpdated Claude CodeCodexVerifiedInstall
xlsxanthropics/skillsUse this skill any time a spreadsheet file is the primary input or output. This means any task where the user wants to: open, read, edit, or fix an existing .xlsx, .xlsm, .xltx, .csv, or .tsv file (e.g., adding columns, computing formulas, formatting, charting, cleaning messy data); create a new spreadsheet from scratch or from other data sources; or convert between tabular file formats. Trigger especially when the user references a spreadsheet file by name or path — even casually (like "the xlsx in my downloads") — and wants something done to it or produced from it. Also trigger for cleaning or restructuring messy tabular data files (malformed rows, misplaced headers, junk data) into proper spreadsheets. The deliverable must be a spreadsheet file. Do NOT trigger when the primary deliverable is a Word document, HTML report, standalone Python script, database pipeline, or Google Sheets API integration, even if tabular data is involved.Stars: 198kUpdated Claude CodeCodexVerifiedInstall
shipping-and-launchaddyosmani/agent-skillsPrepares production launches. Use when preparing to deploy to production, or when asking what needs to be in place before shipping. Use when you need a pre-launch checklist, when setting up monitoring, when planning a staged rollout, or when you need a rollback strategy.Stars: 119kUpdated Claude CodeCodexVerifiedInstall
childhoodthedotmack/claude-memBabysit a shipped release after merge, publication, or deployment. Verify published packages, production behavior, installed plugins, and late CI or user reports until the observation window is clean; use babysit for PR readiness before release.Stars: 118kUpdated Claude CodeCodexVerifiedInstall