Toil Mining Skill for Codex

Toil Mining — rank repeated friction

Mine explicitly supplied session, shell, RTK, or CASS history without modifying

the sources. The result is evidence for a caller; this skill does not file work,

schedule automation, or mutate a tracker.

Constraints

• Read-only over the supplied sources, because the skill gathers evidence and

must never become a mutation lane.

• No work creation: it never files a tracker item, schedules automation, or

names an owner, because those are caller decisions the evidence informs.

• Measured, not remembered: every count, cost, and error rate is read from the

supplied history, so recency and salience cannot masquerade as frequency.

• Observations stay separate from recommendations, so a caller can re-weigh the

ranking without inheriting an unstated conclusion.

Procedure

• Record the input sources, time window, filters, and query.
• Normalize repeated human actions while excluding documented machine echoes

and generated repetitions. For caller-supplied Codex JSONL, use the

deterministic helper below rather than an ad hoc transcript scan.

• Cluster equivalent actions and preserve representative evidence references.
• Score each cluster from measured frequency and observed pain such as elapsed

time, failure count, interruption, or token cost.

• Emit a ranked report and stop.

Each candidate must contain a measured count, source references, confidence in

the clustering, pain evidence, and the smallest plausible automation shape.

Separate observations from recommendations.

Repetition threshold: measured, not remembered

A cluster qualifies as toil only above a measured floor: at least three

occurrences in the supplied window, each resolvable to a source reference. Two

occurrences are a coincidence; a vivid memory of "doing this constantly" with

one resolvable instance is an anecdote. The named failure mode is

salience mining — ranking by how annoying the last occurrence felt rather than

by count, which surfaces yesterday's irritation over the quiet weekly drain.

If the supplied history cannot establish the count, report the candidate as

below-threshold with its actual measured count; never round an impression up

to a frequency.

Weighted priority: frequency x cost x error-proneness

Rank clusters by the product of three measured factors, not by any single one:

• frequency — occurrences per window, from the cluster count;
• cost — median elapsed time or token cost per occurrence, from the

evidence, not from recall;

• error-proneness — fraction of occurrences showing a failure, retry, or

correction in the source.

Score each factor from cited evidence and show the three inputs next to every

composite score so the caller can re-weigh them. A factor the history cannot

support is reported as unmeasured — scored at the floor, never guessed at the

midpoint. The named failure mode is frequency-only ranking: a daily two-second

nuisance outranking a weekly half-hour error-prone ritual because only one

axis was measured. The product form exists precisely so that a high-frequency,

near-zero-cost, never-fails cluster ranks where it belongs: low.

Deterministic recent-human extraction (Codex JSONL)

The helper accepts only explicit session paths and requires an explicit,

timezone-qualified window:

python3 skills/toil-mining/scripts/recent_human.py --since 2026-07-12T00:00:00Z \
  --until 2026-07-16T00:00:00Z /path/to/session-a.jsonl /path/to/session-b.jsonl \
  > /tmp/recent-human.json

It extracts event_msg / user_message records with source_path, one-based

line, normalized UTC timestamp, and request text. Codex attachment and IDE

wrappers are normalized by keeping the text after # My request for Codex:.

Restored or forked copies are deduplicated by client_id, with the earliest

occurrence retained.

The extractor treats a nonempty client_id as the high-confidence UI-origin

boundary. Records without it are reported as missing_client_id, not guessed to

be human. It also excludes these narrow generated envelope families and reports

their counts: internal context tags (codex_internal_context, environment,

permissions, skill/app/plugin instructions), fresh-context cross-family refuter

prompts, and agent Message Type: envelopes. This is deliberately conservative:

a directly typed message from a client that omits client_id remains unchecked.

The JSON result includes input/parsed/candidate/emitted counts, exclusions by

reason, checked facts, and not-checked facts. Malformed records are counted, not

silently discarded. The helper reads only the supplied JSONL and writes only to

stdout; it does not read attachment contents, discover more sessions, cluster

meaning, score toil, file work, or schedule automation. It is

retrieval/report-only.

Output

Write .agents/scratch/toil-mining/YYYY-MM-DD-candidates.md only when the caller

asks for a local artifact; otherwise return the report inline. Include checked and

not-checked sources. Do not include owners, priorities, claims, queues, or a next

action.

Quality

• Every ranked candidate shows its measured frequency, cost, and

error-proneness inputs beside the composite score.

• Sources are cited and resolvable; a factor the history cannot support is

reported at the floor, never guessed at the midpoint.

• Observations and recommendations stay separate, and the report names both its

checked and not-checked sources.

References

• Automation shape routing
• CASS