Rp Telemetry

Capture per-run telemetry that tells us what to improve in the skills and the underlying Wix tooling. This skill stays active alongside the migration skills for the entire run; the migration skills themselves are unchanged and know nothing about telemetry.

The one rule that governs everything you record

Observation, not diagnosis. Every field records what was observed — what happened, what was expected, what actually occurred, where, and how often. Never record a root cause, a fix, a recommendation, or a workaround-as-solution. There is deliberately no field for them; the schema rejects unknown fields. Root cause is derived later, at review time, by a reviewing agent with full context — not asserted by you in the moment.

Three discipline rules apply to every free-text field (what_happened, expected, actual):

• Types, not instances. Refer to entity types and classes, never specific client data. entity_type: "product" — never a product’s name, SKU, price, or body text.
• Secret-safe. No credentials, tokens, URLs, config values, or file contents — the recorder also runs a mechanical scrub as a last line of defense, but do not rely on it.
• Short and structural. A sentence or two, max 400 chars. Prefer stating expected-vs-actual over narrating. The coded fields carry the structure.
• No remediation narration. “The canonical lib was fixed and re-synced” is a fix story, not an observation — even when true (dev-mode backports). Record what was observed to work (“adding fieldsets=FULL returned the field; the lookup succeeded”) and leave the repair to the execution log.

The recorder

All telemetry is persisted through the bundled recorder — never hand-write or edit run-telemetry.json or telemetry/ files. Run it from this resource directory (see CONVENTIONS.md), pointing at the active migration project:

node scripts/rp-telemetry.js <command> [...] --project <abs path to migrations/<project>>

The recorder owns everything mechanical: schema validation (it rejects invalid enums and unknown fields — fix and retry, never guess around it), timestamps, per-class event folding, the run/attempt/session resume model, the privacy scrub, and the well-formedness gate at finalize. It prints one JSON result per call; {"ok":false,...} lists exactly what to fix.

Run lifecycle

1. start as the first telemetry act of the run, with whatever dimensions are already known:

   bashCopy

   node scripts/rp-telemetry.js start '{"source_platform":"wordpress",
     "source_site_url":"https://client-site.example","source_acquisition":"public_storefront",
     "delivery_mode":"management","destination_strategy":"new_site",
     "runtime_env":{"agent_runtime":"claude-code","model":"<model id>"}}' --project <dir>

   start on a project with an unfinalized run resumes it (same run, same attempt, one more session) and closes any open wait interval. start after a finalized run opens the next attempt. Never try to manage run identity yourself.

   source_acquisition is an open set of class tokens, but reuse an established one (admin_api, public_storefront, public_content, file_export) rather than minting a synonym — cross-run folding depends on stable tokens.

   Skills provenance is auto-stamped — you don’t pass it. At start the recorder stamps two identifiers onto the run so every record can be traced back to the skills that produced it:

   • skills_version — the bundle’s semver release label (from VERSION). Coarse and hand-bumped by design; it is the group-by/order-by key that binds issues to a release line (“all issues on 1.2.x”, “regressed since 1.1.0”).
   • — the exact source commit the bundle was built from: the precise within a release. Many commits ship under one hand-bumped , so a run reporting is otherwise unattributable to the change that produced its issue. The recorder resolves it, in order of authority, from the stamped into at publish time (the only source that works in a partner runtime, and the only correct one once the bundle is vendored into ), else a dev-mode of the source checkout, else .

Metering — splitting active_ms into where the time actually went

active_ms alone cannot locate a bottleneck: it is wall-clock between two stage boundaries, fusing model reasoning + subprocess execution + remote API latency + defect-repair time into one number. meter splits it, and cost is derived from the token counts.

# a generated script reporting its own measured work
node scripts/rp-telemetry.js meter --api-ms 250000 --api-calls 73 --api-retries 2 --script-ms 1200 --project <dir>
# the agent runtime reporting its own usage for the stage it just finished
node scripts/rp-telemetry.js meter --model-ms 42000 --input-tokens 8000 --output-tokens 1500 --cache-read-tokens 120000 --project <di

Derive the API half mechanically — do not hand-count it. After a generated import runs:

node scripts/meter-from-audit.js --project <migration dir> --stage import   # --dry-run to preview

It reads logs/import-audit.ndjson and emits measured api_ms / api_calls / api_retries. It exists because a bulk write logs one audit row per item, each carrying the batch’s total latency — summing rows multiplies one call’s latency by its item count, which on a real run reported 1,050,992 ms of API time inside a stage that only existed for 346,505 ms. The helper collapses rows per distinct (runId, endpoint, batch, latencyMs), so a batch counts once while genuinely separate calls each count.

Fields: durations model_ms / api_ms / script_ms; counts input_tokens, output_tokens, cache_read_tokens, cache_write_tokens, api_calls, api_retries. Defaults to the open stage; --stage <stage> targets another. Repeated calls accumulate, so each script invocation and each stage re-entry reports independently.

Rules that keep the numbers trustworthy:

• Measure, never estimate. Every field is a number something actually observed — a script’s own elapsed time, an audit log’s summed latencyMs, the runtime’s reported usage. If you do not have a measurement, omit the field; the gap is reported as unattributed_ms and flagged, which is far more useful than a guess.
• Do not meter the same interval twice. Meters sum, so re-reporting a stage’s API time after a resume double-counts it. The recorder flags stage_over_attributed:<stage> when attributed time exceeds elapsed time, because unattributed_ms clamps at zero and would otherwise hide it.
• Cost is derived, never recorded. Pass model_pricing_snapshot in the start dims ({ "<model>": { "input_per_mtok": 15, "output_per_mtok": 75, "cache_read_per_mtok": 1.5 } }) and the rollup computes cost.estimated_cost_usd. A stored dollar figure silently goes wrong when list prices change; tokens plus a dated snapshot stay recomputable. Without a snapshot, cost is null with cost_basis: "no_pricing_snapshot" — never a fabricated number.
• contained_recovery is derived, not self-reported. Any error or pipeline_defect in a stage marks it, because a stage that spent its time debugging is the one an agent is least likely to remember to flag. timing.stages_with_recovery lists them, so a clean run and a thrash are distinguishable instead of both reading as “this is what the stage costs”.

The rollup’s timing.agentic_ms vs timing.deterministic_ms (and agentic_share) is the number that shows whether moving work into deterministic code is paying off. agentic_share is null when nothing was metered rather than 0, which would falsely read as a fully deterministic run.

Collecting operator_acceptance

At the finish handoff, ask the operator one plain question: does the migrated result look right to them — accepted, rework_needed, or rejected? Record their answer in finalize. If the run never reaches finish or they do not answer, it stays unknown. This is the only field separating “completed and good” from “completed and unusable” — ask it, but never pressure or interpret; their verdict as given, coarse by design.

Event taxonomy — when to record what

Events are folded per problem class by the recorder (same type + stage + entity type + API surfaces + app + error code + subtype fold into one event with a count), so record every occurrence class once and pass count when you observed many at once — e.g. a generated script reporting 4,000 identical write failures is one record call with "count": 4000. Distinct problems are distinct subtypes or error codes, not bigger counts.

Every event needs: event_type, stage (defaults to the open stage), skill (the active rp-* resource, e.g. "rp-mapper"), severity (blocking | degraded | cosmetic | info), and what_happened. Add entity_type, wix_api_surface (e.g. stores/v3), source_api_surface (endpoint class like wp/v2/posts — never a URL), wix_app_id, and error_code whenever they apply — they are the cross-run group-by keys.

1. halt_needs_user — the run stopped at a defined needs-user state. Prefer the mechanical form — wait start --halt <subtype> … emits it for you (lifecycle step 3); record it yourself only for a halt with no wait interval. subtype: missing_input (missing/invalid required input or credential) | manual_only (genuinely manual step, no API) | systemic_failure (systemic failure or data-loss risk).

2. manual_action_required — the execution plan flagged something as “can’t be done, needs manual action” (e.g. a storage-plan upgrade), whether or not the run halted on it. subtype: plan_or_billing | dashboard_only | external_dependency | other.

3. error — an API or script error, including recovered ones. Carry error_code (required — it is the discriminator; no subtype), retry_count, and recovered. Record one event per resolved retry chain, after the outcome is known — the final recovered, the total retry_count — never one call per attempt (per-attempt calls fold into one class, and the failing first attempt’s fields would bury the recovery). When recovery required a change (not just a retry), actual must record what change made it succeed, as an observation: “retried with the description as plain text instead of HTML; the write succeeded” — never “the fix is X”. If that change deviated from the documented path — including a change to a planned command or value, like simplifying an input the plan specified — also emit a paired skill_coverage_gap (undocumented_workaround).

4. fidelity_loss — the migration technically proceeded but lost something. subtype: dropped_field | unverified_enum | no_target (source entity with no clean Wix target) | coerced_value. For a loss triggered by specific records, inline the offending record’s sanitized shape in observed_shapes — field names and types and which field was dropped/coerced, never values.

5. api_gap — a missing or insufficient Wix API capability. subtype: missing_api | missing_capability (API exists but cannot express the operation) | (capability exists but is not publicly exposed) | . Carry (required) and — that signature is matched centrally across runs.

evidence_refs — rare, not routine

Author every event to be self-sufficient: coded fields + bounded free text + observed_shapes should let a reviewer triage it without opening any artifact. Add evidence_refs ({"artifact":"execution-log.md","locator":"## Import"}) only for the exceptional event whose signal genuinely cannot carry the full observation. Use project-root-relative paths and prefer section headings as locators (line ranges break when files regenerate). Never reference secret-bearing config files.

What never to do

• Never hand-write, edit, or re-read run-telemetry.json or telemetry/ files — the recorder appends; you only call it.
• Never record client data values, names, URLs beyond the recorded source origin, or secrets — in any field, including shapes and locators.
• Never record a fix, root cause, or recommendation — observations only.
• Never estimate durations — timing comes from stage/wait boundary calls.
• Never skip a rejected call: fix the listed fields and retry. Rejections are counted against telemetry health either way.
• Never create or update improvement/feature-request backlog files during a run — those are maintained centrally from many runs’ telemetry, not per run.
• Never delete or trim project artifacts for telemetry reasons — capture is a non-destructive view; size is not a reason to drop anything.