fix(claudecode): extract model from transcript to fix Unknown attribution (#1804)

[gtrrz-victor]

https://entire.io/gh/entireio/cli/trails/899

Summary

Fixes the primary CLI-side cause of #1804 — Claude Code sessions attributed as "Unknown" because their checkpoints carry an empty model.

Root cause: Claude Code was the only major agent without a ModelExtractor. It sourced the model solely from the SessionStart hook payload (stored in the model-hint file, recovered on later turns). When SessionStart never delivered a model for a session — hooks installed mid-session, a resumed/--continue session, compaction spawning a fresh segment, or a cleared hint — state.ModelName stayed "" permanently with no fallback. Every checkpoint then carried an empty model, and the entire.io frontend renders an empty model as the "Unknown" agent.

Pi, Copilot, and Factory Droid already recover the model from the transcript at condensation time via sessionStateBackfillModel (manual_commit_condensation.go). This makes Claude Code do the same. The CC transcript always records the model — on the system/init line and on every assistant message.model — so recovery is reliable.

What changed

• ExtractModel for ClaudeCodeAgent (agent/claudecode/model.go): prefer the most recent assistant message.model (clean, hook-consistent identifier, reflects mid-session model switches), falling back to the system/init line's model for very short transcripts. Malformed lines (e.g. an incompletely-flushed transcript) are skipped rather than fatal.
• Registered the agent.ModelExtractor interface assertion, so the existing generic backfill (AsModelExtractor) now fires for Claude Code with no other wiring changes.

Why this shape

Matches the existing Pi/Copilot/Droid pattern exactly (KISS/DRY) and realizes the reporter's suggestion #2 ("propagate the session-level model onto checkpoint records"). Low risk: system.init.model is line 1 of the transcript, present even in a partial/unflushed file.

Not in scope (deliberately)

• The 3s flush-sentinel timeout (reporter's suggestion use github runners #1) is a separate issue and not the cause of the empty model — the sentinel governs transcript-flush completeness, and CC never sourced the model from the transcript before this change. finalize re-reads the full transcript later regardless. Worth revisiting for token-count accuracy on large transcripts, but decoupled from attribution.
• Server/frontend hardening: empty model → "Unknown" agent (same class as Factory Droid checkpoints show 'Unknown' agent in entire.io even though checkpoint metadata contains the correct agent/model #859) would be better fixed by falling back to the checkpoint's agent field when model is empty. Out of scope for this CLI PR.

Tests

• agent/claudecode/model_test.go — most-recent-assistant precedence, system/init fallback, empty transcript, no-model, malformed-line resilience.
• strategy/manual_commit_condensation_test.go — end-to-end guard that sessionStateBackfillModel recovers the model for Claude Code (the reporter's exact scenario).

Full unit suite green (8175 tests); golangci-lint v2.11.3 clean on both changed packages.

Closes #1804

🤖 Generated with Claude Code

---

Note

Low Risk
Localized attribution/backfill logic with no auth or data-path changes; behavior matches existing agents’ transcript model extraction.

Overview
Fixes Claude Code checkpoints showing Unknown model when SessionStart never populated session state (mid-session hooks, resumed sessions, cleared hints).

Claude Code now implements ModelExtractor via ExtractModel, scanning the JSONL transcript: the latest assistant message.model wins (including mid-session switches), with system/init model as fallback when there are no assistant lines yet. Bad lines are skipped instead of failing the scan. A compile-time agent.ModelExtractor assertion wires this into the existing sessionStateBackfillModel path at condensation—same pattern as Pi, Copilot, and Factory Droid.

Tests cover precedence, fallbacks, empty/malformed transcripts, and an end-to-end sessionStateBackfillModel case for Claude Code.

^{Reviewed by Cursor Bugbot for commit 0b5dc22. Configure here.}

[Copilot]

Pull request overview

This PR fixes Claude Code sessions being attributed as “Unknown” by ensuring the CLI can backfill a missing model during condensation: Claude Code now implements agent.ModelExtractor, allowing existing strategy logic (sessionStateBackfillModel) to recover the model from the JSONL transcript when hooks didn’t provide it.

Changes:

• Added ExtractModel for ClaudeCodeAgent, preferring the most recent assistant message.model and falling back to the system init model.
• Registered Claude Code as an agent.ModelExtractor via compile-time interface assertion so generic model backfill runs without extra wiring.
• Added unit + strategy-level tests to guard the backfill behavior and edge cases (malformed lines, empty/no-model transcripts).

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 2 comments.

File	Description
cmd/entire/cli/strategy/manual_commit_condensation_test.go	Adds an end-to-end test ensuring condensation-time model backfill works for Claude Code (regression coverage for #1804).
cmd/entire/cli/agent/claudecode/model.go	Introduces transcript scanning logic to extract Claude Code’s model from JSONL.
cmd/entire/cli/agent/claudecode/model_test.go	Adds focused unit tests for Claude Code model extraction behavior and resilience.
cmd/entire/cli/agent/claudecode/lifecycle.go	Registers ClaudeCodeAgent as implementing agent.ModelExtractor to activate generic backfill.