Observe Run

Run a command, then read the telemetry that run just emitted, and grade a stated behavioral claim against it.

A test asserts through a keyhole. A trace of the same execution exposes the whole thing: fan-out that went from 2 calls to 14, a retry that fired 5 times, an error swallowed into a 200, an N+1 query, a user ID landing as unbounded metric cardinality — none of which turns a test red. This skill closes that gap by giving the dev loop a fourth telemetry role this repo did not have: telemetry as verification evidence, read seconds after the run that produced it, rather than telemetry read as coverage, as yesterday's production exposure, or as a post-deploy confirmation.

This SKILL.md is a thin index. Detailed rules live in rules/*.md and load on demand.

Inputs

The inputs are a command plus an expectation set — nothing else.

Skill("observe-run")
  command:      <the command that runs the code under test>      # required
  expectations: <one or more behavioral assertions>              # required; by-construction ones are refused
  rung:         1 | 2                                            # optional; default resolved by rules/rungs.md
  run_id:       <string>                                         # optional; defaults to a generated dev.run.id
  caller:       <invoking skill or agent>                        # logging only

Return value is exactly one verify-behavior receipt, whose final line is [receipt] verdict: <confirms|contradicts|ambiguous|null> — or nothing at all when the self-skip gate below fires.

This skill has zero dependency on any aw artifact. It takes no worktree for granted, reads no plan artifact, and checks no acceptance-criteria ledger. It runs anywhere a command can run.

The self-skip gate

Before doing any work, resolve the repo's committed Observability Profile (<repo>/memory/observability-profile/INDEX.md, written by measurable setup — see skills/quality/measurable/rules/setup-profile.md) and check it names a dev run target (the field measurable setup now asks for — see Integration: measurable).

This mirrors the quiet-exit shape of the pr-reviewer measurability lens (agents/shared/rules/measurability-review.md): a clean state produces nothing, not a line saying so.

The expectation set is checked too: an empty expectation set has nothing to assert, so this skill returns the same quiet self-skip rather than a vacuous confirms.

Workflow

Step-by-step

1. Self-skip gate (above) — resolve the profile, resolve the dev run target, check the expectation set is non-empty.
2. Reject by-construction expectations. For each expectation, apply rules/assertion-provenance.md's discriminator. Refuse any expectation satisfiable by reading the source alone, and name the behavioral alternative — never run it, never grade it.
3. Read dev-run lessons. After the gate has passed (so a genuine skip still spends nothing) and before rung selection, read the merged repo:: + global lessons in the loop::observe-run-lessons bucket and apply the matches as considerations that bias the rung, the run-identity mechanism, and the reader choice. A missing LoreKit connection makes this a silent no-op. See rules/lessons.md.
4. Select a rung. Default to rung 1 (in-memory/file exporter) for a tight inner loop; escalate to rung 2 (the proxy plus dash0 spans query) only for cross-process fan-out, multi-service claims, or a baseline comparison. See rules/rungs.md.
5. Stamp run identity. Resolve or generate dev.run.id; stamp it alongside deployment.environment.name and vcs.ref.head.name. See rules/run-identity.md.
6. Run the command, then read the telemetry via the selected rung's reader. See rules/reader-adapters.md.
7. Grade the verdict. Map the observed stream state (or the rung-1 span list) to one of confirms / contradicts / ambiguous / null, per rules/receipt-mapping.md, and return the receipt.
8. Write a dev-run lesson — on friction only. If the run hit friction (a blind reader, a timing wait, a forced rung fallback, an SDK-side stamping need), write one lesson to the bucket at the scope the classification table picks. This is a side effect after the receipt is returned — it never changes, delays, or blocks the verdict, and it drops any candidate carrying a secret. See rules/lessons.md.

What this skill reuses

Integration: measurable implement

measurable's implement mode calls this skill as a prove-it step after writing instrumentation, turning a static file:line claim into an executed one. Skipped with one report line when this skill (or its prerequisite dev run target) is unavailable — advisory, consistent with measurable's own Core Principle 6.

Integration: verify-behavior Tier 3

verify-behavior/rules/ladder.md's Tier 3 table gains a third approach delegating here, alongside "run the covering test" and "synthesize a minimal repro." Skipped with one report line when this skill is not installed.

Integration: fix-bug

For a telemetry-sourced bug, fix-bug's Phase 2.5 reproduction step checks repro fidelity, and the division of labour is the important half: fix-bug reads the originating production span's shape out of its own Evidence Record and passes it in as literal expectations; this skill then grades the local run against them.

observe-run never fetches the production span and cannot. Both rungs read only the telemetry the run they just executed emitted, scoped to that run's dev.run.id in the dev dataset (rules/run-identity.md) — so an expectation phrased as "matches the production span" asks this skill to assert on data its reader is structurally unable to see, and it would return null every time.

Core Principles

1. Behavioral, never by-construction. An assertion satisfiable by reading the diff alone is refused, not graded.
2. Cheapest rung first. Rung 1 (in-memory/file exporter) is the default; rung 2 (the proxy) is for the deliberate cross-process pass.
3. OTLP is the contract, not Dash0. A missing Dash0 CLI costs rung 1 nothing, and rung 2 has portable fallback readers.
4. Self-skip is genuine. No profile, no dev run target, or an empty expectation set means no tokens spent and no report line — never "ran and reported skipped."
5. Zero aw coupling. No autonomous-workflow workspace artifact, no plan document, no worktree assumption. This skill runs anywhere a command runs.
6. Reuse the receipt vocabulary verbatim. confirms / contradicts / ambiguous / null — never a fifth token.
7. Lessons accrete, config stays committed. The dev-run lessons loop (rules/lessons.md) is additive over the Observability Profile, never a relocation of it: the profile stays committed and repo-scoped so it is diffable and teammate-visible; lessons live in LoreKit (global + repo::) so they auto-update and merge at read. A missing LoreKit connection costs the loop, and the run, nothing.

Anti-patterns

• Asserting "a span named X exists" instead of a behavioral claim about what the run did.
• Escalating straight to rung 2 for a claim a rung-1 in-memory exporter could already decide.
• Reading a dash0.cli.otlp_proxy.forwarded delivery-receipt event as proof of span content — content comes only from dash0 spans query.
• Treating a degraded proxy state (an error event, *.failed > 0, a deadline-hit shutdown) as contradicts — it is ambiguous, never proof of absence.
• Stamping dev.run.id as a metric dimension.
• Assuming a dev dataset already exists instead of recommending one be created.
• Moving a profile field (the package map, the dev-run command, a dataset name) into the LoreKit lessons loop — that loses PR-diffability and teammate-visibility. The loop is additive beside the committed profile, never a relocation of it (rules/lessons.md).
• Writing a lesson for the verdict itself. confirms/contradicts/ambiguous/null is the receipt's job; a lesson is only about running and reading better next time.

Definition of Done

• The self-skip gate resolved before any other work, with no tokens spent and no report line on a genuine skip.
• Every expectation passed the assertion-provenance discriminator; any by-construction claim was refused, not graded.
• The cheapest rung that could decide the claim was used.
• Run identity (dataset + the three resource attributes) was stamped, each on the signals rules/run-identity.md scopes it to — dev.run.id on spans and logs only, never metrics. At rung 2 that attribute is a requirement, not a best effort: a run that cannot stamp it falls back to rung 1 rather than proceeding without it.
• The receipt's verdict is one of the four canonical tokens, mapped per rules/receipt-mapping.md.
• Dev-run lessons were read (merged repo:: + global) before rung selection, and a lesson was written only on genuine friction, at the scope its classification picks, with any secret-bearing candidate dropped. A missing LoreKit connection left both steps silent and did not block the receipt.