Lesson 7 · 17 min
Without measuring, "the session got expensive" and "the agent forgot" are just impressions. The kit reads the JSONL transcripts from Claude Code and Codex and gives you duration, compactions, tokens and cache per turn, the real cost of resuming, and a search across everything that has been said. In this lesson you measure a known case and then your biggest session.
Start the lesson
Goal: measure the duration, compactions, tokens and cache of a session, read the cost of resuming, and find an old decision with recall.
Four kit tools, Python standard library only, no network.
What was decided has to be in a file in the project. The transcript is for measuring and for remembering where to look.
The kit's test fixtures give fixed numbers: if your output matches, the tool is working right on your machine.
K=~/projetos/execucao-longa python3 $K/tools/medir-sessao.py $K/tests/fixtures/claude_mini.jsonl python3 $K/tools/medir-sessao.py $K/tests/fixtures/codex_mini.jsonl python3 $K/tools/medir-sessao.py $K/tests/fixtures/claude_mini.jsonl --json --por-turno
S=$(ls -S ~/.claude/projects/*/*.jsonl | head -1); echo "$S" python3 ~/projetos/execucao-longa/tools/medir-sessao.py "$S" python3 ~/projetos/execucao-longa/tools/custo-retomada.py 14 python3 ~/projetos/execucao-longa/tools/nota-retomada.py --relatorio
K=~/projetos/execucao-longa; R=/tmp/ensaio-recall rm -rf $R && mkdir -p $R/codex/2026/10/01 $R/claude/-home-voce-projetos-demo $R/arq cp $K/tests/fixtures/recall_codex.jsonl $R/codex/2026/10/01/rollout-a.jsonl cp $K/tests/fixtures/recall_claude.jsonl $R/claude/-home-voce-projetos-demo/s1.jsonl export RECALL_DB=$R/recall.db python3 $K/tools/recall.py indexar --codex $R/codex --claude $R/claude --arquivo $R/arq python3 $K/tools/recall.py indexar --codex $R/codex --claude $R/claude --arquivo $R/arq python3 $K/tools/recall.py PostgreSQL unset RECALL_DB
mkdir -p ~/.local/bin && ln -s ~/projetos/execucao-longa/tools/recall.py ~/.local/bin/recall recall indexar # first time: indexes ~/.claude/projects and ~/.codex/sessions recall "cleanupPeriodDays" -n 5 # search (options: --projeto X --fonte codex|claude --desde AAAA-MM-DD) cp ~/projetos/execucao-longa/tools/systemd/recall-indexar.* ~/.config/systemd/user/ sed -i "s#/home/nmaldaner#$HOME#g" ~/.config/systemd/user/recall-indexar.service systemctl --user daemon-reload && systemctl --user enable --now recall-indexar.timer # reindexes every hour
$ medir-sessao.py tests/fixtures/claude_mini.jsonl
…/tests/fixtures/claude_mini.jsonl (claude)
duração: 7200 s | compactações: 1 | turnos: 2
tokens de entrada: 2000 | saída: 55 | cache: 45.0%
cache lido: 900 | escrito: 1085 | saída de ferramenta: 37 bytes
modelos: claude-opus-5-5, claude-sonnet-5-5
$ medir-sessao.py tests/fixtures/codex_mini.jsonl
…/tests/fixtures/codex_mini.jsonl (codex)
duração: 3600 s | compactações: 1 | turnos: 2
tokens de entrada: 3000 | saída: 120 | cache: 60.0%
cache lido: 1800 | escrito: 0 | saída de ferramenta: 8 bytes
modelos: gpt-6-astra, gpt-6.1-sol
$ medir-sessao.py tests/fixtures/claude_mini.jsonl --json --por-turno
[{"i": 1, "timestamp": "2026-09-02T08:00:05.000Z", "input": 1000, "cached": 0, "ratio": 0.0, "compactacoes": 0},
{"i": 2, "timestamp": "2026-09-02T09:00:10.000Z", "input": 1000, "cached": 900, "ratio": 0.9, "compactacoes": 1}]
$ recall indexar … (duas vezes) ; recall PostgreSQL
2 arquivos processados; 8 trechos novos.
0 arquivos processados; 0 trechos novos.
2026-09-10T10:00:03.000Z codex alfa user Quero migrar o banco para PostgreSQL até sexta …/rollout-a.jsonl:4
$ nota-retomada.py --relatorio (sem medições ainda)
Notas de retomada (últimos 30 dias, 0 medições) — …/notas.jsonl
modo n acerto tokens (mediana)
compactada 0 — —
retomada 0 — —
continuada 0 — —
prime 0 — —Fixtures are small, synthetic transcripts from the kit's tests (model names included); the numbers are for checking the tool, not as a measure of usage. Paths shortened with …
| Measurement | Result | Date · sample |
|---|---|---|
| Resuming with expired cache × /prime | 284 thousand × 187 thousand eq on return; ~23 thousand × ~10 thousand per turn afterward | 14 days up to 10/17/2026 · 4,609 transcripts, 190 × 11 cases |
| Cache in long sessions | 93–99%, not dropping with compactions | 10/01/2026 · 3 sessions (46.7 h, 59.8 h, and 17.8 h) |
| Cost per turn without compacting | last tenth ~5× the first (810 × 157 thousand) | 10/01/2026 · 1 session of 17.8 h |
| Tool output | 916 MB of a 1.84 GB JSONL | 10/01/2026 · 1 session of 46.7 h |
Sources: docs/analise-geral-2026-10-17.md and docs/experimento-f7-retrospectivo-2026-10.md of the kit.
One line per run in progress.md, with the tool's numbers.
Between 93% and 99% is the normal measured in long sessions. Much lower than that in a long session, look for pauses longer than the cache or a model switch midway.
Compaction does not drop the cache: it is no reason to cut the session short.
In Claude, compact_boundary lines; in Codex, compacted lines. Three in the same session: the hook has already asked for a new session at the end of the task (lesson 3).
To see when, use --json --por-turno and look for where the compactacoes column goes up.
It is not a limit: the big file does not block anything; what weighs is the context. But if it dominates, the agent is reading a whole log instead of tail -50.
On the final checkpoint line: input tokens / cache read / output and compactions, as the pilot did.
It is what lets you compare the next run with this one.
| 2026-10-01 14:30 | medição do piloto | 7714c3c | sim (nível 3) | 332791 / 304000 / 4296 (cache 91,3%) | 0 |
First the known case, then your own data.
Measure both fixtures and run the recall rehearsal.
Done when: claude_mini → compactações: 1 | turnos: 2 and cache: 45.0%; codex_mini → cache: 60.0%; recall prints 2 arquivos processados; 8 trechos novos., then 0 arquivos processados and a line with PostgreSQL.
Measure the biggest transcript from your Claude Code (or one from Codex in ~/.codex/sessions) and write the line in the progress.md of your run, with the date.
Done when: the progress.md line has input tokens, cache read, output, and compactions equal to those medir-sessao.py printed for the same file.
Install the recall command, index, and look for a decision you remember making with the agent in the last week.
Done when: recall indexar → N arquivos processados with N greater than 0, and recall "[term]" --desde [date] returns the line with the file and the conversation line number.
Question 1 of 4
How does medir-sessao.py calculate the cache?skill
Question 2 of 4
Why is the JSONL transcript not a good source for what was decided?
Question 3 of 4
What does custo-retomada.py compare?skill
Question 4 of 4
Your session shows huge tool output. What does that indicate?
Pick an answer. The explanation shows up after you try.