jobhunt-agent v1.0.0: updates I can trust, and a brief I can answer

jobhunt-agent runs every morning on my laptop, searches job boards, scores new ads against my CV and drafts applications for the best ones. The setup post covers how to install it.

Five releases in, the agent was careful about what it wrote: v0.3.0 fenced in the unattended run, v0.4.0 made the scores measurable and v0.5.0 checks every letter against its sources. Two things were still loose. "update the engine" pulled whatever was on main, so an update was only as safe as the last commit. And the only way to talk to the routine was to edit a file and wait for the next morning. v1.0.0, "versioned & interactive", fixes both. 1.0 is the promise that updates are safe.

Updates come from releases only

Until now the engine followed the main branch. main only gets release merges, but "the newest commit" is still a moving target. A new setting picks what an update means:

engine_channel: stable # newest vX.Y.Z tag, the default
# engine_channel: next  # follows the next branch, for testing

On stable, "update the engine" checks out the newest release tag, so engine/ sits on v1.0.0 and not on a branch. The daily brief says "engine update available (release v1.0.0)" only when a newer tag exists, not after every commit. An engine that is ahead of the newest release, like mine after testing on next, is never moved back. The tests for this run against a local bare repo, so they don't need GitHub.

The tag itself is now checked before it becomes a release. Pushing vX.Y.Z starts a workflow that compares the tag with the VERSION file, makes sure the CHANGELOG has a section for it, reruns the full test suite and then creates the GitHub Release with that section as its notes. A tag that fails any check makes no release. The v1.0.0 release on GitHub came out of that workflow.

Every output says what made it

When a score looks wrong, the first question is which engine and which model produced it. Every brief now starts with:

engine 1.0.0 (1d7384b) · model claude-opus-5-5

Each job.md has an - **Engine:** line, and every new ledger row stores the engine version and the model, not only the rows that ledger.py set touched in v0.5.0. The CV and letter files stay free of it: their document metadata names only me, because an applicant tracking system reads it too.

A history of my workspace

My workspace has my CV, my letters, the tracker and the ledger, and every run changes some of them. If a run went wrong, I had backups for engine updates but nothing for an ordinary day. setup_check.py --init now makes the workspace a local git repo, with a .gitignore that leaves out engine/, backups/ and Office temp files. Each run ends with one commit:

daily run 2026-11-02, engine v1.0.0

So "what did the last run change?" is a git show, and undoing it is a revert. The history never gets a remote and is never pushed: it holds my personal data, and it stays on my laptop. A workspace that already is a git repo is left as it is, and without a git identity the commits go in as "JobHunt routine".

Commands for one job

The daily run handles the jobs it finds. For a job I find myself, I used to start a full run or explain the steps in chat. Four commands in Claude Code now do it with the same tools and checks as the daily run:

tailor <url> goes through the same quarantine as the daily run: the ad reader fetches the ad, so the session that holds my CV still never sees its text. why and status change nothing. Claude Code already has its own /status, so this one also answers to /job-status.

A brief I can answer

The commands need me at the laptop. Most mornings I only read the brief, so the brief now asks me things. It starts with a Decisions checklist:

## Decisions
 
- [ ] Tailor Acme: Werkstudent Online-Marketing (score 64) `tailor https://www.arbeitsagentur.de/jobsuche/jobdetail/10000-1234567890-S`
- [ ] Mark as sent: 2026-11-01_Example-GmbH_Werkstudent-Marketing `sent 2026-11-01_Example-GmbH_Werkstudent-Marketing`
- [ ] Skip the draft 2026-10-23_Muster-AG_Praktikum-PR (9 days old) `skip 2026-10-23_Muster-AG_Praktikum-PR`

I tick a box in the file, and the next run acts on it with the same tools and lists it under "Done from your decisions". A box I don't tick stays on the list for 7 days, then drops off. Weekend runs only search, so a ticked "tailor" waits for Monday.

The tricky part was a run that stops halfway. If a run marks a job as sent, adds the tracker row and then stops before it writes a new brief, the box is still ticked in yesterday's brief, and the next run would add a second row. decisions.py now logs each box it acted on in data/decisions_done.txt, and a logged box comes back as done, not as a new decision.

Smaller changes

• The title filter is fairer to student jobs. manager, staff, leiter and lead dropped every title with them, so "Werkstudent Produktmanagement" never got a score. They now drop only junior and unlabelled titles, where they still match inside a word ("Teamleiter"). Every other exclude word matches at the start of a word only: art drops "Art Direction" but no longer "Startup Marketing".
• setup_check.py warns about a workspace without history, for example one inside another git repo, where the daily commits can't work.

The eval, before the tag

The release checklist runs the eval before every tag. On the 33 labelled ads, v1.0.0 agrees with v0.5.0 within the tolerance on all six metrics: in scope 30/33 (was 29), role type 20/21, hard mismatch 20/21, score inside the range 19/21, tailor decision 30/30, German level 31/33. It was recorded with Claude Opus 5.5 and cost $3.87. The title filter change gives the same answer on all 33 ad titles for both sample profiles, so the scores weren't expected to move, and they didn't.

Upgrading

Open the workspace folder in the Code tab and say:

update the engine

This is the update that changes your workspace folder itself: setup_check.py --init makes it a git repository. Nothing leaves your computer. The migration adds engine_channel: stable to config.yaml and copies the four command skills and the updated tailor into .claude/. upgrade.py compare should say SAME DOCUMENTS, and it lists the titles in your ledger that the new title filter treats differently.

Then one Run now, and three checks:

1. The brief's first line is engine 1.0.0 (…) · model ….
2. The brief has a Decisions section. It can be empty on a quiet day.
3. git log -1 in the workspace shows daily run <today>, engine v1.0.0. If the commit failed, a global commit.gpgsign=true is a known cause.

python engine/tools/upgrade.py rollback puts the old engine back. The workspace history stays, and v0.5.0 ignores it. The upgrade guide has every step, and the changelog lists every behaviour change.

What's next

Nothing is scheduled yet. On the list: a read-only pipeline dashboard, a weekly model-graded check that every claim in a sent letter can be traced to a source, and a look at moving the engine into a Claude Code plugin, where the plugin version would replace the git tags. From here on, the stable channel only moves when a tag passes the checks above.