Count checkpoint "steps" by prompts, not file-modifying turns

[dipree]

Problem

The "steps" count shown for each session in the UI was often wrong — and frequently 0. It was sourced from the CLI's StepCount, which only counted turns that modified files and reset to 0 after every condensation. Roughly 1 in 5 recent checkpoints displayed "0 steps".

Three issues:

1. Bug — written before a step was recorded. Checkpoints created mid-turn or by a plain git commit (commit-only sessions) condense before any step is recorded, so the count was 0.
2. Bug — entire attach ignored it. The attach path never set the count at all, so every attached checkpoint wrote 0.
3. Wrong thing counted. Even when non-zero, it counted file-modifying turns, not prompts — so conversational/planning turns didn't count.

E.g.

But session has clearly two prompts before the checkpoint https://entire.io/gh/entireio/cli/session/eaa96109-502f-4be2-9357-cf4fde90deef#timeline-46304ab84156

Solution

One concept everywhere: a checkpoint's "steps" = the number of user prompts in that checkpoint's window (prompts since the previous checkpoint, with a deferred reset). Derived from the existing SessionTurnCount (counts every turn, and is safe for exec-mode agents that report turns via hooks):

steps = SessionTurnCount − PromptWindowBase   (floored at 1)

• Deferred reset: writing a checkpoint doesn't zero the window — it restarts only when the next prompt arrives. So two checkpoints with no prompt between them report the same count, not 0. The reset is gated on the turn count actually advancing, so a repeated/stale hook can't re-anchor early.
• Floored at 1, so we never display "0 steps" again.

Scenarios

Sequence	Result
3 prompts → checkpoint A, then 2 prompts → checkpoint B	A = 3, B = 2
3 prompts → checkpoint A, then checkpoint B with no prompt between	A = 3, B = 3
checkpoint with no prompts counted yet (mid-turn)	1

entire attach

Attach is the same counting concept applied post-hoc. The hooks never fired, so instead of live turn events it counts the user prompts parsed from the transcript (meta.TurnCount), floored at 1. Attach produces a single checkpoint covering the whole session, so its window = the whole transcript = the total prompt count. (Earlier this counted len(prompts), which only ever held the first prompt — so every attached session reported 1 regardless of length.)

Related

The combined-attribution gate previously keyed off checkpoints_count == 0. Since that's now floored to ≥1, the gate moves to a new honest SaveStepCount metadata field ("did a real SaveStep run?").

Forward-only — existing rows keep their stored values (no backfill); the API/frontend pass the value through unchanged.

🤖 Generated with Claude Code

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

Comment @cursor review or bugbot run to trigger another review on this PR

^{Reviewed by Cursor Bugbot for commit 615839a. Configure here.}

[cursor]

Stale TurnEnd clears window reset

Medium Severity

After a checkpoint, PromptWindowResetPending is cleared whenever a lifecycle event has TurnEnd or TurnCount > 0, even if SessionTurnCount did not increase. A repeated Cursor Stop hook with the same cumulative loop count can re-anchor PromptWindowBase early, so a later back-to-back checkpoint shows 1 step instead of matching the prior count.

 

^{Reviewed by Cursor Bugbot for commit 615839a. Configure here.}

[Copilot]

Pull request overview

This pull request fixes inaccurate (often zero) session “steps” in checkpoint metadata by redefining the displayed count to represent user prompts (derived from SessionTurnCount and a new prompt-window base), with a deferred window reset so back-to-back checkpoints without an intervening prompt keep the same step count. It also introduces SaveStepCount to preserve the previous “did SaveStep actually run?” signal for combined-attribution gating.

Changes:

• Compute displayed steps as SessionTurnCount - PromptWindowBase (floored to 1) and defer window resets until the next counted turn.
• Add SaveStepCount to committed session metadata and switch combined-attribution gating to use it instead of checkpoints_count == 0.
• Ensure entire attach writes a non-zero displayed step count; add/adjust tests for the prompt-window behavior.

Reviewed changes

Copilot reviewed 10 out of 10 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
cmd/entire/cli/strategy/manual_commit_test.go	Updates test setup to reflect new CheckpointsCount semantics (prompt-window-based).
cmd/entire/cli/strategy/manual_commit_hooks.go	Changes combined-attribution gating from CheckpointsCount to new SaveStepCount.
cmd/entire/cli/strategy/manual_commit_condensation.go	Introduces checkpointStepCount, writes SaveStepCount, and sets deferred reset flag after condensation.
cmd/entire/cli/strategy/manual_commit_condensation_test.go	Adds unit test coverage for checkpointStepCount window math and flooring behavior.
cmd/entire/cli/session/state.go	Adds PromptWindowBase and PromptWindowResetPending to session state to support deferred window resets.
cmd/entire/cli/lifecycle.go	Implements deferred prompt-window re-anchoring when a turn is first counted after checkpoint write.
cmd/entire/cli/lifecycle_test.go	Adds tests for deferred reset behavior and exec-mode cumulative TurnCount handling.
cmd/entire/cli/checkpoint/committed.go	Persists new SaveStepCount field into committed session metadata.
cmd/entire/cli/checkpoint/checkpoint.go	Adds SaveStepCount to options/metadata types (but one existing comment now mismatches semantics).
cmd/entire/cli/attach.go	Ensures attached checkpoints write a non-zero displayed steps count via attachStepCount.