lifecycle: keep a no-change turn's tokens for every agent

[peyton-alt]

https://entire.io/gh/entireio/cli/trails/1536

Bug. When a turn changes no files, turn-end skips SaveStep. That path recorded only out-of-band usage (Antigravity), which caused two losses:

• Cursor: the stop payload's event.TokenUsage was dropped. Cursor's transcript has no usage fields, so the turn was missing from both the session total and the next checkpoint.
• Claude Code, Codex and other agents that keep usage in their transcripts: the turn never reached the session total (entire status). Their checkpoints were unaffected, because condensation recomputes them from the transcript slice.

A Factory AI Droid Worker turn that changed no files also returned before the Worker routing, so it left no task record on the parent, and its transcript and tokens never reached the parent checkpoint. This one predates the PR.

Fix.

• resolveTurnTokenUsage is now the single resolver for both turn-end paths. Precedence: hook-reported, then Codex's subagent ledger, then the transcript from the turn's start offset, then out-of-band.
• SaveStep's accounting is extracted as applyTurnTokenUsage. It covers the session total, the pending checkpoint usage, the subagent baseline rescoping, and dropping a child snapshot computed against a stale Codex ledger version. SaveStep and ApplyTurnTokenUsage both use it.
• A no-change turn's tokens are applied inside the same MutateSessionState that moves the session from ACTIVE to IDLE, and only while it is ACTIVE. A repeated, retried or concurrent Stop for the same turn finds it IDLE and adds nothing. Without this, a retried Claude Code Stop, which runs after the pre-prompt state is cleaned up, re-added every earlier turn.
• Droid Workers are resolved before the no-changes gate. A read-only Worker turn becomes a completed, zero-file task record on the parent, which is itself a condensation trigger. Each Worker turn adds its own usage, from the turn's start offset, to that record. Only the Stop that ends the Worker's own active turn adds it. Condensation drops a completed record once it is stored, so a Worker turn already in one checkpoint is never counted by the next, and none of it lands on the Worker's own session total.
• A Stop that finds no active turn records nothing. That covers a repeated Stop, and also a turn whose TurnStart never fired, which has no start offset to scope the usage to. Every built-in agent fires a TurnStart; Droid installs its prompt hook on SessionStart too, so exec mode is covered.

No checkpoint double-counts the turn. For transcript agents, condensation still takes the main counters from the transcript slice, and the pending checkpoint usage is only a fallback.

Not in this PR. SaveStep (the file-changing path) has the same repeated-Stop exposure, which predates this PR: a retried Stop adds the tokens again and also records an extra step. The phase gate does not carry over, because SaveStep also legitimately runs on a non-ACTIVE session, for example the first Stop after Entire is enabled mid-session, which initializes state as IDLE. Fixing it needs a per-turn step marker, so it is a follow-up.

Tests.

• Integration tests for Cursor, Claude Code and Codex drive the real hooks: a file-changing turn, a no-change turn, a commit, then one more turn and commit. They assert each turn lands once in the session total and once in a checkpoint. All three fail on main.
• Repeated and concurrent Stop tests for Cursor (hook tokens) and Claude Code (transcript tokens) on no-change turns assert that the session total and the pending checkpoint usage hold the turn exactly once. Without the phase gate they read 200 for 100 (Cursor) and 280 for 140 (Claude Code).
• A Droid integration test covers a read-only Worker that consumes tokens, with its Stop fired twice, then a commit, a second Worker turn, and a second commit. It asserts the parent's completed task record, the Worker transcript and task.json usage in the first checkpoint, that the Worker's own session records none, and that the second checkpoint's task holds only the second turn. It fails before this change.
• Strategy unit tests cover subagent rescoping and a stale ledger version on the no-change path.

🤖 Generated with Claude Code

[Copilot]

🟢 Approval recommended

The fix matches existing token-source precedence and is covered by focused end-to-end regression testing.

0 open findings

What changed in this PR

Fixes lost Cursor token usage when a turn changes no files by preserving hook-reported usage in session and checkpoint accounting.

Changes:

• Uses hook-reported usage before out-of-band fallback on checkpoint-less turns.
• Adds integration coverage for session totals and checkpoint scoping across three turns.

File	Description
cmd/​entire/​cli/​lifecycle.go	Preserves token usage when SaveStep is skipped.
cmd/​entire/​cli/​integration_test/​cursor_token_condensation_test.go	Verifies no-change turn accounting and checkpoint isolation.

🧠 Review effort: Balanced

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.