The lab journal

Claude Code files a transcript under the path of the directory the session ran in, and the tooling looked only under the current directory and its checkout root. From a worktree, the recount and the digest would never have found the sessions that ran in the main checkout. They now look under every checkout that git worktree list names.

Conductor needed one more step. A Conductor workspace is a worktree of Conductor's own clone, so from there git worktree list never names the main checkout. Transcripts are now found by session id in any Claude Code project folder, since ids are unique. Tested in a workspace chat, Conductor ran the prompt hook but showed no SessionStart note. The guide now says to run pnpm lab:journal status rather than read a missing note as a failed hook.

Before the writer was extended to media, an audit held its brief against the two published pages. The brief asked for one chart, tokens per session, and said the agent could not see or create images, so it would only place media ids a person handed it. That line was out of date. The agent has every tool, and Read shows it images.

A second gap sat under the first. pnpm cms:upload sent a file to CMS_UPLOAD_SERVER, else NEXT_PUBLIC_SERVER_URL, which in that workspace was localhost, while the sas-cms MCP writes to production. An uploaded id would name a local media document that the production draft could not resolve.

The figure plan adds four questions to the digest's request for each journal entry: screenshot, diagram, code listing and chart. The first wording asked whether the entry described a mechanism a diagram could show. Jev answered 0.84 to 0.97 for all 17 entries. Rewritten as a score from 0 to 3, 15 of the 17 scored 2.9 or higher. A plan that lists everything is not a plan.

TypeSafe's own limitations page for Jev 1.13 names the cause: it reads a question literally, and literally anything could be drawn. The fix asks what the entry is mainly about, and writes the near miss into the false side of the question. With that wording the diagram answers spread from 0.16 to 0.88, and five entries pass 0.5. Finding it took three full reruns of the digest at about 55,000 input tokens each.

Miles asked for an agent that stays with a feature across days and sessions. Claude recommended against a subagent for that part, and Miles accepted. A subagent runs in its own context, sees only what it is handed and returns one message, so it cannot watch a session, and keeping it informed would mean resending the context each time.

So the work splits three ways. A skill tells the main session what to log and when, because only that session has the context. Hooks capture prompts and token counts with no model involved. One subagent runs at the end, the lab-project-writer, where a fresh context helps: it reads the journal, not five days of conversation.

The journal, sessions.jsonl and meta.json are committed. Prompts go, redacted with the same redactFreeText the Ask assistant uses, to a folder under ~/.claude/lab-journals that every workspace and worktree on the machine shares.

Rejected: everything in the repository, which travels with the branch and publishes raw prompts, and a gitignored folder in the checkout, which a second Conductor workspace would not see. The cost is that prompt history does not follow the work to another machine. It was Claude's call, made without asking because it is the safe default and reversible, and flagged to Miles.

Miles's call: the writer takes its screenshots through the Claude in Chrome extension, in the Chrome window he is already signed into. For admin shots he signs in first and the agent moves around the admin. A reference site uses the same browser.

Rejected: a Playwright capture script with its own admin login, which needs a stored password or a saved login state on every machine, and fixture data unless it points at production. The extension shares the browser's login state, and its screenshot tool can save the image to disk, so the file goes straight to pnpm cms:upload. The cost is that the writer needs a session started with claude --chrome and a visible Chrome window.

Miles asked that every piece of text the writer produces follow the editorial guide's voice, tone and style, with TypeSafe wherever a judgment fits. He noted the guide was current on voice and stale on composition: its page structure describes service pages built with an older block set.

Three layers, each catching what the one before cannot. The rules move to one file, docs/editorial/voice.md, and a test keeps the code's phrase lists identical to it. A Payload plugin refuses, by path, any save from an API key that holds an em dash or a banned phrase. A check reads the rest, with code for what is exact and Jev for what is a judgment. Only the gate can promise anything. The other two report.

Rejected: rewriting the guide, because the ask was copy that follows it. A Claude review of each draft was rejected too, since it costs what the draft cost and judges with the writer's own bias. People editing in the admin are not gated, because a quotation or a client's own words may hold a phrase the doc bans.

A Story beats block printed a heading only when the page set one, by design. Miles's call reversed that design note and the agent's recommendation, which was to write a page heading on each block that needed one, taken from the record by hand. In his words:

incorrect it should fallback to heading only overrides can "override" it. story beat headings would need to follow proper heading hierarchy with correct heading type set. use /typesafe:typesafe-ai to classify this automatically in the future.

The rejected option put the same words in two places and left 64 blocks on the two Lab Pages to fill in by hand each time a page is composed. The chosen rule has a cost of its own. The Prose heading usually prints the first beat's heading already, so something has to decide when a beat heading only restates the heading above it, and at what level a beat sits. That judgment went to Jev.

Four Claude Code hooks are registered in .claude/settings.json, SessionStart, UserPromptSubmit, Stop and SessionEnd, and each runs scripts/lab-journal/hook.ts. Between them they capture each prompt, redacted, and each session's token counts, with no model involved.

Measured on this machine, one run each: a trivial TypeScript file starts in 52 ms under Node's native type stripping and 164 ms under tsx, and the real prompt hook took 163 ms end to end. tsx was kept. It matches every other script in the repository, so the hook reuses Ask's redaction instead of a copy. The prompt and Stop hooks run async, so those 163 ms never delay a reply. Only SessionStart is waited for, once per session.

By 16:47 UTC the first version existed: scripts/lab-journal with its library, hook, command line and digest, ten unit tests, the lab-journal skill, the lab-project-writer agent and the four hooks. Every hook event was exercised by hand with real payload shapes, and start, log, status and the digest ran for real against the session that built them.

The milestone also listed what was not verified. The hooks had not fired from Claude Code itself, which needed a new session. The writer had never run. The digest thresholds were untuned defaults.

The digest, scripts/lab-journal/digest.ts, reads the raw record once and asks Jev about each item on its own. A prompt gets three questions: what it was doing, how much of the story it tells, and whether it must not be published. A journal entry gets its story section, and in the same request whether it could carry a screenshot, a diagram, a code listing or a chart. For each agent message of 280 characters or more, four questions ask whether it reads as a decision, a problem, a measurement or a lesson, so an unlogged moment can be found without rereading transcripts.

Raw probabilities are cached, the thresholds live in one constant, and an item already judged is never sent again. The writer starts from what comes out: the prompts worth quoting and the ones held back, each entry's story section, a figure plan, and candidate moments to check against the journal.

pnpm lab:journal:verify holds a Lab Project draft against the journal. Code goes first, because code is exact: it lists every em dash and every number in the draft that the record does not hold. Then Jev takes each sentence with the three journal entries that share the most words with it and answers supports, contradicts or says nothing, following TypeSafe's citation check cookbook. A second question over the same state asks whether the sentence states a checkable fact at all, so a line of connective prose is not reported as unsupported.

The idea came from Miles asking whether TypeSafe could make the writer's media work cheaper and better. Jev 1.13 takes text only, so it cannot look at a screenshot, and that check stays with Claude. What it can do is plan the figures before the writer starts and hold the finished draft to the record.

docs/editorial/voice.md now holds the house voice: how it should feel, sentences, punctuation, the language to favor and to avoid, the contrast frames, the flattened claims, and a table of what is checked where. The editorial guide's voice section points at it, and a note at the guide's top marks its composition sections as the service-page system that predates Sections, Story beats and Lab Pages.

src/features/editorial/voice.ts holds the same rules as code. Its test reads voice.md and fails if the two lists differ, so the two cannot drift.

The check, pnpm editorial:voice, runs on any Lab Project, any document by collection and id, a file or a string, and pnpm lab:journal:verify ends with the same Voice section. Code lints what is exact: banned phrases, contrast frames, semicolons, paragraph length. Jev scores each paragraph on the guide's own terms: specific or generic, composed or inflated, human or formulaic, and whether it closes on a punchline. For headings, Jev names each one's grammatical form and code flags a run of one form.

Jev cannot read taste. The check finds the faults the guide names, not whether a line is good.

At render, src/blocks/shared/story-headings.ts finds the Section's first Prose heading and its level. Each beat prints its record heading one level under it, is hidden when it repeats the opener exactly, and gives way to the block's own heading when the page overrides it. Nothing at render calls a model.

What code cannot read goes to Jev when the page is saved. A plugin on the collections that offer the block, Lab Pages and Work Pages, asks two questions about each beat that code has not settled: does its heading only restate the opener, and does it open a subsection or a short passage. A subsection sits one level under the opener, a passage two. The answers are stored on the block with a hash of what was judged, and the renderer trusts one only while its hash matches. Autosave is exempt, and a failure leaves the beat to the code rule, so a save never fails for it.

The stored answers needed one new json column on the Story beats tables of both collections and their version tables, so the change waited on a migration, Miles's call.

PR 18 went to main at 23:08 UTC on 2026-09-21, after Miles started a new Conductor workspace and found none of the work in it, since a workspace starts from main. His call was to merge then and keep working after. It went in as a merge commit, not a squash, so the branch and main share history. GitHub's merge call answered with a server error twice and then an empty body, and the merge had gone through anyway. The pull request's state was the thing to check.

PR 19 followed at 01:56 UTC with the voice layers, the strict pass and the record beat headings, again as a merge commit. The schema check found the one additive column, heading_auto. A migration with exactly that change was already in the tree when pnpm migrate:create ran, the command added nothing, and pnpm check:migrations and the drift check both passed against it.

Miles set the direction and made the calls the journal marks as his: build it, use TypeSafe, never journal main, screenshots from his own Chrome, merge PR 18 early, headings from the record, one Lab Project. Claude designed and built the tooling, wrote the docs, logged the journal entries, and made the calls the journal marks as Claude's, such as keeping prompts out of the repository. Jev judged the record and the drafts. No model ran in the hooks.

Seven Claude Code sessions built it, between 16:33 UTC on 2026-09-21 and 01:56 UTC the next morning. Two models shared them. The model claude-fable-5-1 carried the three long sessions: 2,668 input, 318,339 output, 1,191,464 cache write and 47,836,354 cache read tokens. The other, claude-opus-5, ran four short sessions and part of a long one, for 152 input, 41,453 output, 319,049 cache write and 5,330,506 cache read. Cache reads are most of both totals, and they do not compare with output.

The write-up runs on claude-opus-5-5 in further sessions on 2026-09-27, counted in the same file and not in these sums. Jev's runs are logged apart, in jev.jsonl. The journal puts the first digest at 17 requests and 13,556 input tokens, each of three figure-plan reruns at about 55,000 input tokens, and the planted draft check at 5 requests and 6,052. Four of those rows were restored by hand after a careless git checkout dropped them, with approximate times. The voice calibration and the heading judgments ran outside that file, and their costs are in their own entries.

pnpm editorial:voice ran on 2026-09-21 against copy Miles had approved, Lab Project 3 with 70 paragraphs and Lab Project 2 with 48, and a planted four paragraph agency draft written to break every rule in voice.md. Code found no em dash and no banned phrase in any approved copy field, and refused the planted draft at every paragraph.

At the first thresholds Jev listed approved paragraphs too. The approved passages it read as inflated were the confident assertions the doc asks for, at 0.60 to 0.78, against 0.87 to 0.94 for the planted hype. Approved copy read formulaic at 0.68 at most, the planted draft from 0.79. So the thresholds moved into the gaps: inflated 0.8, formulaic 0.75, generic 0.5.

After the move Lab Project 3 lists 7 generic passages and nothing else, Lab Project 2 lists nothing, and the planted draft still lists 3 of 4 as generic and 4 of 4 on the other two. The runs cost 119 requests and 76,854 input tokens for Lab Project 3, 71 and 47,517 for Lab Project 2, and 8 and 4,987 for the planted draft.

The heading rule and the save-time judgment were run over Lab Page 2, From Webflow to Payload, on 2026-09-22: 24 Sections and 48 beats. Of those, 33 beat headings print and 15 are hidden. Code settled 13, all exact repeats of the Section's Prose heading and all the first beat of their Section. Jev was asked about the other 35, one request each.

It differed from the code rule on two, both paraphrases code cannot see, and both right: "Every case study in the same template" under "One template for every case study", and "A Content Hub and a website" under "Two layers: a Content Hub and a website". Every other beat read 0.05 to 0.29 against a threshold of 0.6. On level, Jev called 22 beats a subsection and 13 a passage: the passages are the short single-idea beats, the subsections the multi-paragraph steps. One page is not an accuracy figure.

Five lessons, each from a moment logged above, and what the build left open.