Read Codex transcripts and flag re-asked requests

[ArkashJ]

Harvested 195 transcripts (155 Claude + 55 Codex) after twenty-three days in which the
weekly cron logged nothing but Not logged in · Please run /login. Three defects fell out
of the run itself.

Codex reduced to zero turns and exited 0

reduce_session read rec["message"]["role"]; Codex rollouts carry the turn under
payload. On a 1.4MB rollout the tool printed a well-formed reduction reporting
human turns: 0 and returned success — indistinguishable from a quiet session.

Normalising Codex records inside records() fixes every reader at once: reduction, JSON,
--only-corrections, --repeats. --list scans ~/.codex/sessions alongside
~/.claude/projects (--codex-root overrides), and fork snapshots collapse by session id
— Codex writes one file per fork, so 1721 files since 2026-08-11 are 55 real sessions.

An empty reduction now exits non-zero and says so on stderr. Silent zero is worse than
unsupported.

Re-ask detection

The correction heuristics are lexical, so the most expensive class is invisible: a user
re-asking for something the assistant already reported done.

The labelled case from this corpus — 42 PRs added to a project board, verified at 43 items
via gh project item-list with every field set, while all six default views filtered on an
iteration that was never set, so the board the user actually opens rendered empty. The
user asked twice more, escalating. That session reduces to likely corrections: 0.

A turn that restates an earlier turn and adds specificity or force is now flagged reask
and marked ↩. Containment, not Jaccard: escalation adds words rather than sharing them,
so an overlap score reads the repeat as a new request. Corpus rate 3.1% of turns, 58% of
them carrying no correction keyword.

This does not catch the board case, and I would rather say so than tune until it does.
After stopwords that turn is three content words — below any threshold that keeps the
false-positive rate sane. Catching it needs the assistant's own completion claim as an
input, not more lexical overlap.

find_repeats was mostly false positives

61 pairs over the corpus. ~13 were one- and two-word turns (try again, check again)
whose word sets overlap by accident; ~14 were the same template prompt reused in two
sessions (/mission read docs/launch/MISSION.md, the compaction preamble).

A six-content-word floor, a near-verbatim skip for long blocks, and treating a leading
slash command as an invocation rather than prose: 61 pairs → 39, template reuse 16 → 2.

Verification

• python3 -m unittest discover -s tests — 28 tests, all pass (8 new: Codex reduction,
  Codex shell/error recovery, Codex goal boilerplate, fork collapse, re-ask positive and
  negative, non-zero empty exit, repeat floor).
• uv build and uvx twine check dist/* — both pass at 1.1.0.
• Every number above is measured over the 195-transcript corpus, not estimated.

Version 1.1.0; CHANGELOG and README updated.

🤖 Generated with Claude Code

https://claude.ai/code/session_0189C5Lc8PeFkh88goiURRMN