transcripto

your coding agents keep a transcript of every session. it is the most valuable dataset you own, and you cannot scroll back far enough to read it. transcripto indexes it, keeps only the turns you actually typed, and grades them.

one command, no account, no signup, no cloud. it reads files that are already on your disk and never opens a socket.

uvx transcripto coach

which build you got. uvx transcripto --version should print 0.1.4. If --version is not a recognised flag at all you are on 0.1.1, which predates trace, Cursor support and this README. uvx --refresh transcripto forces a fresh resolve past uv's cache.

The repo always matches this README and needs nothing installed:

git clone https://github.com/Morkeeth/transcripto && cd transcripto
python3 transcripto.py coach

three harnesses, one instrument

transcripto coach                     # Claude Code, ~/.claude/projects
transcripto coach --harness codex     # Codex, ~/.codex
transcripto coach --harness cursor    # Cursor, ~/.cursor/projects/*/agent-transcripts

Authorship is not the same gate in all three, and the tool says so rather than pooling them. Claude Code stamps promptSource: typed, which is the measured-reliable signal: about 95% of raw type: user records are not the operator at all. Cursor has no such field. Its one honest equivalent is the <user_query> wrapper it puts around a submitted prompt, which injected and tool-result records do not carry. That is a weaker signal and it is labelled weaker.

trace — what actually happened after you asked

ask shows what you typed. find shows what a file went through. Neither answers the question that matters after the fact: you asked for X, did anything durable happen?

transcripto trace "the gate"

It walks each of your matching prompts forward inside its own session and lists the writes and edits that followed, stopping at your next prompt so one turn cannot claim the next turn's work. Green dot = something durable landed. Red = nothing was touched.

Honest limit: a write following a prompt in the same session is CO-OCCURRENCE, not proof the write was caused by that prompt or that it was correct. Same proxy coach uses, labelled the same way.

what you get back

this is a real run on one machine, 2026-08-31. The numbers are unedited; the two quoted prompts are synthetic stand-ins of the same shape, because your prompts never leave your machine and neither do mine:

  YOUR PROMPT HABITS, GRADED   (offline, your machine only)

  - your worst looped prompt, with its witness:
      "hmm ok lets just try again and see if it works this time, same thing as before but…"
      NO-DURABLE-RECORD: read-only Bash only, no file change  ·  corrections: 15  ·  assistant turns: 221

  + your best landed prompt, with its witness:
      "TERMINAL 3 — parser · ~/CODE/demo Read hack.md. Fix the counter, run the suite, commit if green…"
      COMMIT-WITNESSED: git commit  ·  corrections: 0

  SURVIVAL IS A PROXY: survival = a durable Write/Edit or an un-reverted git commit in-episode. A PROXY, not proof the work was correct or shipped.

  SURVIVES MOST  do more of these:
     65%  (71/110)  states-a-check-or-done-condition
     64%  (178/276)  detailed (>40 words)
     59%  (22/37)  no-object (pronoun/vague)
     59%  (451/763)  intent:CHANGE
     57%  (106/187)  cites-a-file-or-path

  SURVIVES LEAST  these tend to loop:
     31%  (4/13)  intent:REVERT
     40%  (451/1129)  intent:none
     42%  (320/765)  terse (<8 words)
     42%  (81/192)  intent:DESCRIBE
     45%  (5/11)  intent:TEST

  harness claude  ·  2957 transcript(s), 431,085 records  ·  3969 typed by you (0.92%)
  episodes: 2108 ranked, 992 survived (47%)
  commit 454  ·  write/edit 538  ·  reverted 2  ·  nothing durable 1114

  ─────────────────────────────────────────────────────────
   compare yours. numbers only, nothing from your prompts:

   transcripto coach · claude · 2108 episodes · 47% survived
   states-a-check-or-done-condition    65%  (71/110)
   intent:none                         40%  (451/1129)
   gap                                 1.6x
  ─────────────────────────────────────────────────────────

those are my numbers on that date, and they move every session i run, so treat them as a snapshot rather than a constant. yours will be different, which is the whole point. the last two lines are the ones that sting: it hands you back your own best and worst prompt, verbatim, with the receipt for why it scored each one.

on that machine, on that date, prompts that wrote down what done looks like survived 63% of the time (66 of 104). prompts with no stated intent survived 39% (411 of 1043). i had spent a year blaming the model.

one day later, 2026-08-29, the same command on the same machine read 63% (67 of 107) and 40% (424 of 1072) over 2,874 transcripts. the percentages held and the denominators moved, which is what a snapshot is supposed to do.

the proxy caveat, which travels with every number

an episode "survived" if a Write or Edit landed, or a git commit ran and nothing reverted it inside the same transcript.

that is a durable keystroke, not a durable outcome. a commit is not proof the code was right. a revert in a later session is invisible to it. a prompt whose payoff was a decision rather than an edit reads as dead. it is a coaching signal, not a verdict. if it ever prints something that flatters you, distrust it.

the caveat is printed in the output itself, every run, on purpose.

correction rate

in the repo since 2026-09-02, not yet on PyPI. python3 transcripto.py coach prints it.

one more line under the coach footer:

  correction rate: 6% measured (227 of 4061 typed turns) · v1 catches ~1 in 6, so the real rate is ~26-37%

correction rate = typed turns that correct the agent ÷ typed turns. the denominator is the same authorship gate as every other number here (typed or queued, never isMeta, isSidechain or a tool result), so a tool result that happens to say "wrong" cannot move it. a typed turn counts as a correction when a marker fires in its head: the first 80 words after URLs and file paths are stripped, case-insensitive, whole-word: a leading or bare no, again, wrong, not that, I meant, revert, undo, stop, instead, plus a few measured additions. whole-word, so "against" is not "again" and "now" is not "no". the v0 "nudge" rule (a short turn naming the agent's last file) is gone: measured over 100 flagged rows it fired alone five times and was wrong five times. TRANSCRIPTO_CORRECTION=v0 brings the old classifier back for comparison.

it is one pure function, is_correction(text), and it is a floor, now measured rather than asserted: on a 200-row labelled sample the v1 classifier has precision 0.81 and recall 0.16 (docs/CORRECTION-PRECISION-2026-09-03.md), which is why the printed line carries the "~1 in 6" correction and a range. read it as a trend on your own history. test_correction.sh pins the rule and the gate on fixtures-correction/.

export-run

in the repo since 2026-09-02, not yet on PyPI.

transcripto export-run latest                 # the newest session on this machine
transcripto export-run 3f9c1a2b               # a session id, or a prefix of one
transcripto export-run path/to/session.jsonl  # a transcript file
transcripto export-run latest --harness codex # --root / --harness as for coach

one run's numbers as JSON, read straight from the transcript file (no index needed). this is the contract other tools read (Agent Grinder's card, ZUP's board); the keys are frozen under schema, and a new key is an addition, never a rename.

key	meaning
schema	transcripto.export-run/1
session_id	the harness's session id (Claude Code: the file name; Codex: session_meta.id; Cursor: the file name)
project	the run's cwd
harness	claude · codex · cursor
transcript	absolute path of the file read
started · ended	first and last record timestamp, UTC, …Z; null if the file carries none
duration_s	ended − started, whole seconds
records	every record in the file, before any gate
typed_turns	records that pass the authorship gate: promptSource typed or queued, never meta, sidechain or tool result. the same count coach prints as "typed by you"
corrections	typed turns is_correction() flags (see above)
correction_rate	corrections / typed_turns, 3 decimals; null when nothing was typed
tool_calls	every tool_use block the agent emitted
files_touched	sorted set of file_path (or notebook_path) from Edit / Write / Read / MultiEdit / NotebookEdit calls
commits_in_window	commits stamped inside [started, ended] in the project's git reflog; null when project is not inside a git repo
commits	those commits as {sha, ts, subject}, oldest first; null when not a repo
proxy	the caveat, in the JSON so it travels with the numbers

commits_in_window reads .git/logs/HEAD directly, not git log, because this file does not shell out (see privacy). the reflog is the record of what that working tree did: a commit made there in the window is in it, a commit pulled in from elsewhere is not. git expires the reflog after 90 days by default, so a run older than that can read 0 here while git log would still show its commits. commit (amend) counts; a rebase's pick lines do not. a .git file (a worktree) is followed to its gitdir.

why your own gate matters here

at fleet scale roughly 95% of the type: user records in a transcript are not you. they are tool results, injected skill bodies, sub-agent prompts, and messages from other terminals, all wearing your role. transcripto gates on promptSource (typed/queued, no meta, no sidechain) so it grades what you typed.

you can watch the gate do work: in the run above, 3678 of 381,804 records survived it. that is 0.96%.

the same gate is what makes cost produce a number a spend tracker cannot:

cost per human decision  last 30 days  2026-07-28 → 2026-08-27

  API-equivalent spend      $8,892.49
  your decisions            2934   turns you actually typed (promptSource typed/queued)
  ────────────────────────────────────────────────────
  cost per human decision   $3.03

  46.4k agent messages · 16 per decision · 11.5B tokens · 149 sessions
  57.7k raw `type: user` records in the same window. dividing by those instead
  would read $0.15, 19.7x too cheap.

it prints both, so the gate's effect is something you can check rather than something i am asserting. these are API-equivalent dollars at list rates, because a transcript has no cost field, only token counts. on a subscription you did not pay this.

honest limits

read these before you quote a number at anyone.

• survival is a proxy, described above. durable keystroke, not durable outcome.
• one operator's corpus. every figure in this README comes from one machine. it is an existence proof that the measurement runs, not a finding about how people prompt. run it on yours and you get yours.
• three harnesses today: Claude Code, Codex, Cursor. nothing else is supported. aider and the rest are not read. and the three are not equal: Claude Code has a measured-reliable authorship field, Cursor has only the <user_query> wrapper, which is weaker and is labelled weaker wherever it is used.
• the habit labels are heuristics. "states-a-check-or-done-condition" is a pattern match over your text, not comprehension. it will misfile some prompts.
• correlation, not instruction. detailed prompts surviving more often does not prove that padding a prompt causes survival.

privacy

it runs locally and never touches the network. there is no socket, no urllib, no requests, no subprocess, no telemetry, no analytics, and no account. that is a claim, so here is the grep that settles it against the single file it ships as:

$ grep -nE '^[[:space:]]*(import|from) ' transcripto.py
8:import sys, os, json, glob, re, sqlite3, argparse
9:from datetime import datetime, timezone
261:    import time
1160:            import datetime
1170:    from the separator), so the result is checked on disk and dropped if it is

five lines, four of which are imports and all four are stdlib. time and datetime sit inside functions, which is why the pattern allows for indentation — anchor it at ^import and you would miss two, so do not take my word for the anchor either. line 1170 is the pattern catching a docstring that happens to begin with the word from; it is prose, not an import, and it is left in rather than tuned out, because a grep you tuned until it agreed with you proves nothing.

what the list does NOT contain is the actual claim: no socket, no urllib, no requests, no http.client, no subprocess. that one is checkable too, and the right answer is no output at all:

$ grep -nE '\b(socket|urllib|requests|http\.client|subprocess)\b' transcripto.py
$

your transcripts stay in ~/.claude, ~/.codex and ~/.cursor. the index it builds stays in ~/.trace.

the rest of it

coach and cost read your transcript files directly and need nothing set up. the other six read a local index, so run this once first:

transcripto index      # a few minutes on a large corpus, incremental after that

on a 2,874-file corpus that was 164 seconds, measured 2026-08-29. if you skip it, the six say so and exit 2.

transcripto index      build / refresh (incremental)
transcripto watch      live, new sessions get picked up as your agents work
transcripto ask        YOUR OWN messages about a topic, newest first + a rollup
transcripto search     full-text across everything (you + agents + tool logs)
transcripto find       every session that wrote / edited / read a file
transcripto trace      what durably happened after each prompt you typed (0.1.2+)
transcripto sessions   recent sessions + the first prompt YOU typed in each
transcripto stats      what you actually work on
transcripto cost       what ONE of your decisions costs
transcripto coach      which of YOUR prompt habits survive (a proxy)
transcripto export-run one run's numbers as JSON (typed turns, correction rate, commits)

ask is the one that kills "wait, did i lose something?". it answers "what was i thinking about X across ALL my sessions", in your own words only.

$ transcripto find USER-JOURNEY.md          # run 2026-08-29
USER-JOURNEY.md  4 touches across sessions (3 were writes/edits)

2026-08-20  WROTE  ~/CODE/demo/USER-JOURNEY.md                              abd9e871
2026-08-21  WROTE  ~/CODE/demo/docs/onboarding-notes.md                       0f845ede
2026-08-27  read   ~/CODE/demo-api/docs/USER-JOURNEY.md                       cddfde29
2026-08-27  WROTE  ~/CODE/demo-api/docs/USER-JOURNEY.md                       cddfde29

the file you lost, found across every session you ever ran, with the session id that touched it. find needs transcripto index first.

Codex

transcripto coach --harness codex

reads ~/.codex (sessions + archived_sessions), normalises it into the same rows, and applies the identical survival proxy. it also ingests history.jsonl purely as a control on the gate: it reports how many of its input lines also show up as typed rollout turns, so you can see the gate agreeing with a second source.

install

uvx transcripto coach

no install, nothing to set up. or put it on your PATH:

pipx install transcripto

or run the single file with no packaging at all:

git clone https://github.com/Morkeeth/transcripto
cd transcripto
python3 transcripto.py coach

no dependencies, stdlib only, one file. the packaging adds nothing at runtime, it just gives the file a name on your PATH.

tests

./test_coach.sh        15 assertions
./test_codex.sh        14 assertions
./test_cost.sh         12 assertions
./test_small_n.sh       7 assertions
./test_label_bands.sh  13 assertions
./test_correction.sh   32 assertions   (correction rate + export-run, 2026-09-02)
./test_cursor_partial.sh 7 assertions
./test_version.sh       1 assertion    (VERSION matches in transcripto.py and pyproject.toml)

91 assertions, all green, re-run 2026-09-02.

offline, no keys, on fixtures that inherit the real transcript shape including all four ways a non-human record disguises itself as type: user.

the load-bearing one in test_coach.sh is REVERTED IS NOT SURVIVED: a commit that got reset --hard in the same session left no durable record. flip that one line and the suite goes red, which is the point. a generous proxy is a broken one.

test_label_bands.sh is the other one, and it exists because 0.1.1 shipped the defect it pins. SURVIVES MOST took the top five habits and SURVIVES LEAST took the bottom five, which overlap whenever you have fewer than ten rankable habits — so a new user, who necessarily has few, read the same habit at the same percentage under both "do more of these" and "these tend to loop". on a 3-habit corpus 0.1.1 reprinted all three, all at 65% (22/34). the suite is red on the published 0.1.1 file and green on this one, and its wide band asserts the fix leaves a large corpus byte-identical.

why though

your agent history is proof. every "yeah it's done" has a real trace sitting behind it. transcripto is the index that makes it checkable.

it is the fuel layer. on top of it you check what your agents claim against what the trace shows, which is mountain of helicon. the pitch was never "search your history". it is prove your agent did what it said, from your own local traces.

local, MIT, no telemetry. star it if it finds you something you'd lost ™