[EHotwagner]

Refs #4801 (source-only Delivery; parent Outcome #4640 keeps installed adoption, Host/receiver ingestion and display)

Summary

Claude Code native usage now reaches the existing local telemetry path. It uses the same semantics as Codex and needs no new wire schema.

• Strict transcript reader (RuntimeUsage.ClaudeTranscript, src/FS.GG.Coord.Core/Telemetry.fs{,i}). The CLI and the passive scanner share it. It refuses invalid UTF-8, comments, trailing commas and duplicate properties at any depth. Lines are capped at 4 MiB and files at 128 MiB. It keeps only allowlisted identities and counters. It tolerates other record types such as user, attachment, system and mode, and ignores <synthetic>-model local notices.
• RuntimeUsage.collectClaudeTranscript emits one row per distinct API response, keyed by message.id, with requestId, model, effort, version and timestamp bound.
  • Per-content-block repeats are deduplicated. Only completed records (stop_reason set) count. Repeated completed usage that disagrees is refused.
  • Input is uncached + cache-read + cache-creation. reasoning is output_tokens_details.thinking_tokens and is set only when present.
  • turn_id is the promptId of the nearest preceding user record at completion. Turn totals accumulate within (thread, prompt), and thread totals stay empty.
  • Sidechain records get a derived, stable version-8 thread identity per agentId, so they never join main-thread totals.
• Fail-closed, never zero. These all refuse with a named reason: partial tail, malformed line, oversized line or file, missing or null counters, thinking > output, missing version, no preceding prompt, mixed sessions, and an incomplete streaming response in the selection. An incomplete response also makes later turn totals in its turn unknown.
• CLI. fsgg-coord telemetry usage collect claude takes --session-file with the Codex --turn-id/--all-responses/--since/--until contract. The existing --snapshot still works; the two options are mutually exclusive. Source provenance is claude-transcript-jsonl:sha256:<digest>.
• Passive capture.
  • A fsgg.telemetry.passive-selection/1 file may name exactly one of codexHome or claudeHome. Claude sources must be regular files below <claudeHome>/projects, opened with the existing no-follow opener.
  • Rows are provider-usage-observation/1 with provenance claude-local-transcript/1 and stable identities in the claude-transcript/1 namespace. Coverage stays partial/unknown and attribution canonical-only.
  • The cursor keeps the prompt context, so resumed passes attribute the same way a full read does. Codex cursor bytes are unchanged.
  • The digest leaves out per-block timestamps, so identical repeats are not corrections, while disagreeing ones become numbered revisions.
  • TelemetryStore maps each admitted provenance to its own identity namespace.
• Guidance. The pnext-item lifecycle-ledger, work-roadmap lifecycle-log and runtime-observation now name the working transcript route in place of the unwired status-line/Stop-hook recipe. Both roots are byte-identical and the driver manifest is regenerated. The Cli README is updated too.
• Atlas. New lesson github/claude-transcript-native-usage in docs/atlas and both seeds.

Measured (one host, Claude Code 2.1.299; shape only, no transcript content committed)

• Main-thread transcript: 151 assistant records across 78 message ids. Every record is completed, and usage is identical within each id.
• Subagent transcript: observed at <session>/subagents/agent-<agentId>.jsonl with a .meta.json. The writer streams: per-block records carry stop_reason: null and provisional output_tokens before one completed record, and 11 of 17 ids disagreed. That is why dedupe uses completed records only.
• Turn attribution: sequential promptId agreed with parentUuid-chain resolution for all 209 records. In subagent transcripts promptId follows the session's current prompt.
• Local parity check (not committed): on copies of these real transcripts, CLI rows and passive records gave 86/86 (main) with 0 counter/turn/thread mismatches. The subagent gave 92 CLI rows inside the --until window against 103 passive records, with 0 mismatches.

Verification (local, at this head)

• dotnet test tests/FS.GG.Coord.Core.Tests -c Release: 1227 passed
• dotnet test tests/FS.GG.Coord.Cli.Tests -c Release: 750 passed (focused Claude passive tests rerun after the rotation control was added: 3 passed)
• dotnet test tests/FS.GG.Telemetry.Tests -c Release: 118 passed
• Causal mutation checks: removing the disagreement refusal, defaulting a missing counter to 0, or counting streaming records each fails the focused Claude tests (1, 1 and 6 failures).
• python3 tests/telemetry-parity/run.py: pass (51 positive, 47 rejection)
• python3 scripts/check-fsc02-telemetry-closure.py: no issues
• bash scripts/check-skill-quality: OK (29 skills). scripts/skill-view-parity.sh: AGREE. scripts/generate-driver-manifest --check: current.
• python3 tests/atlas/check.py --baseline 1ddc8c33…: 355 records, 674 local citations verified
• python3 scripts/check-signature-doc-siting.py: OK
• The skill-quality workflow step was run locally: skill-fsharp contracts/run.py, readers/run.py, adapter/run.sh, preflight/run.sh, cli/run.py and tests/skill-quality/run.sh all pass. ReadersProbe now references Core, because the passive reader shares Core's transcript decoder; its lock file is updated.

Unknown / deferred

• Whether current status-line payloads still carry prompt_id/timestamp was not measured; --snapshot is unchanged.
• The subagent transcript path is one observation; the readers take explicit files and assume no path convention.
• Older transcripts without promptId, isSidechain or agentId fail closed.
• A receiver must run a Host built from this source to admit claude-local-transcript/1. Installed adoption, Host ingestion and display stay with [V2-EFF-01.2–.6] Qualify efficiency telemetry, dashboard and selected receiver adoption #4640, and Project 3 enrollment is a separate Programme/root effect.
• --since/--until compare timestamps ordinally, as Codex does, so …:30.000Z sorts before …:30Z.

Coherent113 source composition

The PR edits packaged kit guidance and the CLI, so it composes source 0.113.0 like #4793: FsggCoherentSetVersion, registry/dependencies.yml, regenerated version projections (architecture map, compatibility, publishing-and-deployment inventory), CLI PackageReleaseNotes, registry/skills changelogs, and the skills.yml/driver-manifest digests. Source only: published112 and installed pins are unchanged, and no candidate or publication is authorized. After merge, main's kit-published-coherence / published arm stays red until coherent113 is published, the same state as after the 0.112 source bump.

🤖 Generated with Claude Code