| name | otel-instrument |
| description | Instrument repositories in every language with an official OpenTelemetry implementation: C++, C#/.NET, Erlang/Elixir, Go, Java, JavaScript/TypeScript, Kotlin, PHP, Python, Ruby, Rust, and Swift. Use when a user asks to add or deepen traces, correlated logs, metrics, context propagation, OTLP/HTTP export, or an auditable observability-instrumentation inventory across a repository or monorepo. |
OpenTelemetry Instrumentation
Instrument supported repositories through a gated, convergent process. Prefer official OpenTelemetry or dependency-upstream integrations, preserve existing observability, and leave evidence for every changed, skipped, or blocked module.
Non-negotiable boundaries
- Support C++, C#/.NET, Erlang/Elixir, Go, Java, JavaScript/TypeScript, Kotlin, PHP, Python, Ruby, Rust, and Swift. Detect other stacks, report them as unsupported, and do not improvise edits.
- Use OTLP over HTTP only. Remote endpoints require HTTPS; HTTP is allowed only for loopback/local Collectors.
- Never disable TLS verification, capture credentials or payload bodies, or place business identifiers in metric labels.
- Treat every automatic integration's emitted attributes as untrusted until inspected; official instrumentation is not evidence of data safety. Block any boundary that can export secrets, raw dynamic paths, or unsanitized URL queries.
- Never use a package or fork based on model memory alone. Verify current official ownership, compatibility, and versions from primary sources. If internet access is unavailable, make no new dependency selection that cannot be verified locally.
- Accept only instrumentation maintained by OpenTelemetry or by the dependency's upstream project. If neither provides an integration or fork, skip it and explain the resulting coverage gap.
- Keep tenant tokens out of source, tracked configuration, plans, inventories, commands captured in files, and agent output. Never ask the user to paste, upload, reveal, or validate a tenant token in the agent conversation. Use a placeholder in generated guidance and direct the user to inject the real value through their local environment or secret manager without exposing it to the agent.
- Automatically edit existing tracked runtime/deployment configuration when the approved plan identifies it as the canonical startup surface. Never deploy, edit live resources, create secrets, invent secret references, or write credential values.
Required references
Read only the references whose condition applies:
- Before any analysis or modification, read references/execution.md, references/instrumentation.md, and references/contracts.md.
- After detecting supported stacks, read every matching language file: references/cpp.md, references/dotnet.md, references/erlang-elixir.md, references/go.md, references/java.md, references/javascript.md, references/kotlin.md, references/php.md, references/python.md, references/ruby.md, references/rust.md, and/or references/swift.md.
- Before generating runtime or deployment configuration, read references/backend.md and references/deployment.md.
Workflow
0. Check for a skill update
Before repository analysis, when the distribution base URL and current agent adapter are available from the installation request or current session, run the matching installer once with --skill otel-instrument --agent <agent-name> --upgrade. Do not add --yes: a single-skill upgrade is already explicit and non-interactive. On Windows, use the equivalent install.ps1 command with -Skill otel-instrument -Agent <agent-name> -Upgrade.
If the installed version matches the release index, continue without reporting an upgrade. If a newer clean version is installed, state that the skill was updated and continue using the already-loaded workflow; recommend starting a new agent session only when the host cannot reload changed skill files. If the installation has local modifications, do not force replacement: report the guard and continue with the current version. If the base URL or adapter is unavailable, do not guess or interrupt the instrumentation workflow to ask for it.
1. Establish repository safety
Inspect Git status before writing. Explain that an isolated folder (technically a Git worktree) keeps instrumentation separate without another clone, and offer to create one as the recommended option. Show the proposed branch and sibling path before approval. If the tool can switch safely, continue there; otherwise hand off with the exact reopen command and persist the approved state.
If the user stays in the current worktree, allow unrelated changes but stop when a planned edit overlaps them. Record the initial Git state so the final report distinguishes pre-existing work.
Completion criterion: the working location is approved, the initial Git state is recorded, and no planned file overlaps uncommitted user edits.
2. Collect the instrumentation profile
Ask one question at a time and always wait for the answer. On the first run, collect:
- Signals as a multi-select: traces, correlated logs, metrics.
- If traces are selected, depth: Light, Medium, or High.
- If traces are selected, the production sampling strategy/rate; keep it environment-configured.
- The OTLP/HTTP host. Reuse a host already supplied in the request; otherwise present the backend default for confirmation. Normalize a supplied base URL to a host and preserve explicit signal endpoints when provided.
- Whether to use the recommended isolated folder.
Reuse answers from .otel/plan.json only when stack, dependencies, module boundaries, and security-relevant inputs have not changed. Otherwise show the differences and ask again. Do not ask about unavailable branches: for example, do not ask trace depth when traces are not selected.
Completion criterion: every applicable input is explicit and serializable into the plan contract.
3. Build an exhaustive repository model
Read manifests, lockfiles, entry points, dependency wiring, deployment files, existing telemetry, request paths, jobs, consumers, external clients, and logging/metrics setup. Bind each deployable module to every tracked startup path and its canonical runtime configuration surface. Define modules as independently built or deployable units; keep internal packages beneath their owning module. Exclude generated/vendor code, copied third-party code, build output, fixtures, examples, benchmarks, migrations, and archived code unless the plan explicitly includes them.
Run discoverable formatting, build/type-check, and test commands before modification. Record existing failures. Do not modify a critical module that cannot build at baseline; it may remain in the plan as blocked.
Completion criterion: every supported deployable/build unit has an owner stack, entry points, important boundaries, existing telemetry state, validation commands, and an included/excluded/blocked disposition.
4. Research official integrations and design the plan
Research dependencies from their current manifests rather than from a maintained list. Prefer, in order: existing valid instrumentation, official automatic/framework instrumentation, official bridges/wrappers, manual OpenTelemetry APIs, then an official fork when source-compatible replacement is required. Verify publisher/repository ownership, release provenance, runtime requirements, and source compatibility from primary sources.
Select the newest version compatible with the repository's existing runtime and constraints; fall back through older official releases when necessary. Never upgrade the language runtime without separate approval. If no compatible official option exists, skip the dependency and document why.
Design hybrid coverage: use official library/framework instrumentation first, then add manual spans only at meaningful boundaries automatic instrumentation cannot see. Apply the chosen level and signal rules from references/instrumentation.md. List every proposed business identifier and whether it will be raw, transformed, or excluded; choosing Medium/High does not authorize identifiers silently.
For every HTTP client/server integration, inspect the exact attributes emitted by the selected version and map all dynamic path segments, query parameters, URL-carried credentials, object keys, and presigned URLs. Plan the sanitization mechanism and negative tests before approving that integration. If safe attributes cannot be guaranteed without changing the real request, skip the automatic integration for that boundary and use a safe manual span plus propagation.
Write .otel/plan.json. For each batch and module, include exact source/dependency/config edits, proposed official dependencies or forks with evidence URLs and versions, identifier decisions, tests, risks, and rollback. List every discovered runtime/deployment configuration surface, select the canonical surface for each startup path, and record whether tenant-header wiring can reuse an existing secret reference or remains blocked for human setup. In a large monorepo, plan the whole repository but execute in reviewable deployable-module batches.
Completion criterion: every discovered module and dependency opportunity is planned, skipped, excluded, or blocked with evidence and a validation path.
5. Obtain one plan approval
Present the plan in human-readable form. Make fork replacements, business identifiers, expected generated files, and known baseline failures conspicuous. One approval authorizes every explicitly listed edit and fork; it authorizes nothing absent from the plan. If the plan changes materially during implementation, update it and request approval for the delta.
Completion criterion: the exact plan revision to execute has recorded human approval.
6. Instrument convergently in batches
Preserve valid existing telemetry. Add only missing initialization, propagation, official integrations, selective business spans, log correlation, low-cardinality metrics, flush/shutdown handling, and environment-driven export configuration. When an approved canonical configuration surface exists, automatically edit it with the selected non-secret OTel protocol, signal endpoints, resource, sampling, and startup settings instead of only printing a handoff snippet. Reuse an exact existing secret reference when proven; otherwise leave credential wiring explicitly blocked and never guess one. Do not create a deployment system solely for telemetry, edit generated render output, or duplicate providers, exporters, middleware, spans, metrics, or configuration on reruns.
Complete one batch before starting another. Format source and update lockfiles with the repository's own tools. Never touch excluded areas merely to increase coverage.
Completion criterion: every edit in the approved batch is applied exactly once and every unplanned discovery is deferred to a plan delta.
7. Prove the batch locally
Run the same baseline commands plus targeted tests. Existing failures may remain; no new failure is acceptable. Verify initialization, context propagation, error/status recording, shutdown/flush, and the chosen signal/depth semantics. Render or parse every changed runtime/deployment configuration and verify that its effective startup command receives the intended settings. Run an OTLP/HTTP smoke test against a temporary local capture endpoint and assert the expected signal/module evidence without using a tenant credential. Inspect decoded exported attributes rather than only request paths or initialization logs. Exercise sensitive outbound and unknown inbound URLs with synthetic canaries, prove the canaries are absent from every exported signal, and separately prove the real destination received the unchanged request.
Attempt bounded repairs for instrumentation-caused failures. If still failing, preserve the reviewable changes, mark the batch failed, provide evidence and safe rollback instructions, and do not mark it complete.
Completion criterion: the batch has either passed all local gates or is explicitly failed with reproducible evidence; “compiled” alone is not success.
8. Publish the inventory and handoff
Update .otel/instrumentation.json and docs/observability-instrumentation.md using references/contracts.md. Account for every discovered module and every approved plan item. Include injected signals, depth, files, boundaries, dependencies, configuration, validation evidence, omissions, and coverage impact.
After all successful local batches, summarize the applied configuration from references/backend.md and references/deployment.md. Generate environment settings or an OpenTelemetry launch command only for startup paths without a tracked configuration surface or for unresolved secret wiring. Include http/protobuf, signal-specific /v1/traces, /v1/metrics, and /v1/logs endpoints for selected signals, and use only a non-secret placeholder when explaining the tenant Authorization header. Give exact human-only steps for placing the credential in the discovered local environment or secret facility, but never request it, read it back, print it, or run authenticated verification. Remote upload verification remains an optional human-run check based only on evidence the human chooses to summarize.
Completion criterion: the machine and human inventories agree, every module is accounted for, no secret value appears in tracked files or output, and the user has actionable deployment guidance.
Failure behavior
- Stop on overlapping user changes, unapproved material plan deltas, unverifiable dependency provenance, remote HTTP, disabled TLS validation, or attempted secret exposure.
- Continue other modules/signals when one official integration is unavailable; mark the gap instead of substituting an unofficial implementation.
- Never automatically roll back a dirty worktree. Preserve changes and provide precise rollback guidance.
- Never claim remote export success without human-provided evidence from a run outside the agent session.
