Capture Claude Code native usage through the existing telemetry path #4801

Owner: Programme /root is accountable for intake, enrollment and acceptance. The .github telemetry owners own the source change. No dispatch follows from filing.

Done when: A Claude Code session's native usage reaches the existing local telemetry path with the same semantics Codex has, and no new wire schema is needed. Specifically: (1) a strict, bounded reader for Claude Code transcript JSONL (~/.claude/projects/<project>/<session>.jsonl) emits one row per distinct API response. Rows are keyed by message.id with requestId, model, effort and timestamp, and are deduplicated across the repeated per-content-block records. Uncached, cache-read and cache-creation input and output are counted, and output_tokens_details.thinking_tokens is reported as reasoning only when present. Main-thread and sidechain/subagent usage stay separate. Malformed, partial, rotated or oversized input fails closed or withholds totals, and is never reported as zero. (2) fsgg-coord telemetry usage collect claude accepts that transcript as well as the existing --snapshot, matching the Codex --session-file/--turn-id/--all-responses/--since/--until contract. (3) Passive capture accepts an explicitly selected Claude transcript through the existing provider-usage-observation/1 path, with its own provenance, stable passive identities and canonical-only partial attribution, as with codex-local-rollout/1. (4) The pnext-item lifecycle-ledger and runtime-observation guidance name the working Claude route instead of the unwired status-line/Stop-hook recipe. Focused tests use real-shape fixtures, include causal negative controls, and pass in the required CI.

Current deliverable: Source-only Delivery: the transcript reader, CLI/passive-capture wiring, tests and guidance in FS-GG/.github. Installed adoption, Host/receiver ingestion and browser display stay with parent #4640. Candidate: PR #4802 (routine/claude-native-usage-capture), head 1552a12ffbfb3dbda1c29ec35e7098bde31ae4a8, open and unmerged. Independent source review has not been recorded. Merging this source delivers no installed, Host or receiver acceptance.

Next action: The accountable source owner implements the reader beside SkillTelemetryReaders/Telemetry.collectClaude in one PR. Protected enrollment in the Project 3 cohort is a separate pending Programme/root effect before any board-selected dispatch.

Blocked on: Nothing blocks source preparation. Unknown: the subagent transcript location/shape (no Claude subagent transcript exists on the observing host), and whether the current status-line payload still carries prompt_id/timestamp as collectClaude requires. Both must be measured, not assumed.

Acceptance evidence: Gap observed 2026-10-22 on main 5a80f17. Codex has session collection (collectCodex), passive rollout capture (codex-local-rollout/1) and dispatched-child usage via App Server. Claude Code has only collectClaude, which needs a status-line snapshot file. No status line, Stop or SubagentStop hook is configured in .claude/settings.json or user settings, and scripts/v2-progress/collect-local.py reads only ~/.codex/sessions. Live Claude transcript shape observed: 69 distinct message.id values across 125 assistant records, with identical usage repeated within each id. .fsgg/agents.yml requires requireEquivalentClaudeAndCodexBehavior. Related parity fix: #4800; Atlas github/claude-codex-runtime-parity. Record the PR/head, CI and review here after they exist.

Candidate evidence, recorded 2026-10-22 for PR #4802 at head 1552a12f:

• Local test runs at this head: dotnet test tests/FS.GG.Coord.Core.Tests -c Release 1227 passed; tests/FS.GG.Coord.Cli.Tests 750 passed; tests/FS.GG.Telemetry.Tests 118 passed.
• Other local checks: tests/telemetry-parity/run.py passed; scripts/check-fsc02-telemetry-closure.py reported no issues; scripts/check-signature-doc-siting.py OK; skill-fsharp contracts, readers, adapter, preflight and cli all passed; scripts/check-skill-quality OK; skill-view-parity AGREE; generate-driver-manifest --check current; tests/atlas/check.py checked 355 records.
• Causal mutations, each caught by the focused Claude tests: removing the disagreeing-usage refusal, defaulting a missing counter to 0, and counting streaming records.
• Real-shape parity (local only, nothing committed): on copies of one main and one subagent transcript, CLI rows and passive records matched with 0 counter, turn or thread mismatches.
• Measured on Claude Code 2.1.299 (shape only): the subagent transcript lives at <session>/subagents/agent-<agentId>.jsonl. Its streaming writer emits stop_reason: null records with provisional counters, and 11 of 17 ids disagreed until completion, so dedupe uses completed records only. promptId is on every user record.
• Required checks passed. Non-required kit-published-coherence / pr-arm is red because .claude/skills/pnext-item is kit source and the Kit <Version>, which is the coherent set 0.112.0, is not above the published 0.112.0. A coherent-set/kit successor and kit release are needed under their own authority before or with merge. CI at this head: 67 pass, 1 skipping, and only pr-arm failing; coord-engine / engine passed.
• Still Unknown: whether status-line prompt_id/timestamp exist; --snapshot is unchanged.
• Atlas: github/claude-transcript-native-usage.