Session audit counts the interventions this user actually makes

[EdbertChan]

Summary

The session audit measures user signals and work performed so teams can decide when reflection or automation is needed.

Previously, typed intervention commands were invisible to the repeated-intervention signal, so two direct requests could pass without a must-automate flag.

Now, the audit counts intervention commands from structured user-turn fields and treats transcripts written during auditing as unchecked.

Review Claim

Approve counting repeated intervention commands and withholding partial subagent totals while their transcripts are still changing.

Review Lane

behavior

Review Unit

engine-runtime

Safety Invariant

The auditor only reads transcripts and prints a report; it writes nothing outside its own output file. Existing counts and flags keep their current meaning, and the new count is additive. No regex is added for meaning: the slash command is read as a field, not matched out of prose.

Assumption: this supplied safety invariant is treated as confirmed for this headless PR-body draft.

Slice Rationale

This slice changes one auditor and its focused tests. A phrase dictionary for meaning belongs in a separate change because it would introduce a different classification claim.

Non-goals

• No change to existing frustration or thrash kinds.
• No change to pricing.
• No phrase dictionary or prose matching for meaning.

Architecture

Before

graph TD
    A["user transcript"] --> B["tone and intervention-kind analysis"]
    B --> C["must-automate flag"]
    D["subagent files"] --> E["totals printed as complete"]

Loading

After

graph TD
    A["user transcript command-name fields"] --> B["intervention command count"]
    B --> C["must-automate flag"]
    D["subagent file modification times"] --> E["unchecked state when active"]
    E --> F["no combined total"]

Loading

Test Plan

Test Plan

• python3 engine/skills/make-pr/scripts/preflight.py --base origin/main — ok preflight passed
• python3 -m unittest discover -s engine/skills/reflect/scripts/tests -v — Ran 204 tests in 40.857s; OK
• python3 scripts/check_no_new_comments.py --base origin/main — ok no new comments
• bash scripts/scrub-handoff-artifacts.sh — scrub-handoff-artifacts-ok

Revert Plan

Revert Plan

• Safe to revert? Yes.
• Revert command: git revert <merge-commit-sha>.
• Post-revert steps: Rerun the focused unittest command and preflight checks.
• Data migration? No.

---

Note

Low Risk
Read-only transcript analysis with additive flags and conservative omission of combined totals when subagent files change mid-audit; no auth, billing, or persistence beyond existing audit JSON output.

Overview
The reflect token audit now treats explicit user slash commands as intervention signals, and it refuses to publish partial subagent totals while those transcripts are still being written.

Intervention commands: Claude sessions are scanned for human user turns containing structured <command-name> markers for /automate-me and /thrash. Those counts are added to the frustration payload and folded into intervention-must-automate—two or more such commands trigger yes the same way repeated tone-based intervention kinds do. Flag rationales and console summaries now mention intervention_commands=…. Subagent transcript paths are excluded from this scan.

Subagent stability: At audit start, audit_claude records a timestamp; if any subagent agent-*.jsonl is modified at or after that moment, subagent aggregation is marked unchecked, combined_total is omitted (null), subagent-thrash is unchecked, and output explains which files were still active.

Unit tests cover the 0/1/2 intervention-command thresholds and the mtime-based unchecked subagent path.

^{Reviewed by Cursor Bugbot for commit e6e92ad. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Bugbot couldn't run - usage limit reached

Bugbot is counted against Cursor usage for this user or team, and this run hit a usage or spend limit.

A user or team admin can review and increase usage limits in the Cursor dashboard.

(requestId: serverGenReqId_317d37e6-a160-4c71-8006-58af33aa24c4)