jobsearch 2.4.3

• The outcome skill's reason log example now passes the required --confidence;
  copied as written it exited 1 with NO_CONFIDENCE. A docs test asserts every
  reason log example carries it.
• Two tests assumed the workspace is the checkout and failed with
  JOB_SEARCH_HOME set elsewhere: one passed a workspace-relative --file, the
  other imported a template by an absolute path that only resolves when the Typst
  root is the code root. The product was unaffected.

jobsearch 2.4.2

Fixed

• Every delegated tool rejects an unknown flag. In 2.4.1 only the tools that parsed their own flags strictly did. jobsearch run markets --x exited 0 with a normal result, and 20 of the 23 tools dropped or misreported an unknown flag. Each tool now declares its flags in the tool table, and the tools entry rejects anything else before the tool runs: {"error","code":"BAD_ARGS","hint"} on stderr, empty stdout, exit 2. A value that starts with -- is passed as --flag=value (#17).
• Usage errors exit 2. Tools that reported BAD_ARGS or BAD_CMD and exited 1 (application, select-template and others) now exit 2, as the contract documents (#17).
• The outcome skill lost observations. It ran reason log --note, a flag the tool never had and silently ignored. It now passes --finding. A test checks that every jobsearch run <tool> --flag in the plugin docs is a declared flag.

jobsearch 2.4.1

Fixed

• Project spacing matches 2.3.0 again. Splitting the project block in 2.4.0 also cut about 10pt between a project's title, description and highlights, so identical content came out about 95pt shorter. That made an executive CV that passed final-page density in 2.3.0 fail in 2.4.0. The 2.3.0 spacing is restored to within 1pt, and a paired test renders the same content with the 2.3.0 code and the current code (#15).
• Non-final page slack is 17%, the measured height of the smallest unit that cannot split with that spacing: section heading, entry title and a three-line paragraph. It was 15% in 2.4.0 and 12% before (#15).
• An unsupported flag on a delegated tool is a JSON error. jobsearch run <tool> with a flag the tool does not accept printed a Bun TypeError and stack trace. It now writes {"error", "code": "BAD_ARGS", "hint"} to stderr and exits 2, for every tool in both a checkout and an installed plugin (#16).

jobsearch 2.4.0

Findings from the owner-confirmed end-to-end run (9 cells, Typst 0.13.1).

Fixed

• Page 1 no longer ends early before a project. A project was one unbreakable block (title, description and every highlight), so it jumped to page 2 and left 15-19% of page 1 empty. That failed layout:density on 6 of 9 cells, across all layouts. Only the heading and title are now kept with the description; highlights break normally. A non-final page may end up to 15% early, the size of the smallest unit that cannot split.
• layout:density says which page and why. "underfilled final page" was printed for page 1 too. A page that ends early now reports that a block moved to the next page, with a layout hint instead of advice to change the content.
• ATS section order reads headings, not prose. The word "experience" in the summary was taken for the Experience heading and failed a correctly ordered CV. Headings must now be a line of their own. The CV is checked against the order its source declares, so a custom sectionOrder and the project-first layout no longer fail.
• GCP and AWS match their long names. A profile with "Google Cloud Platform" left a posting's "GCP" as a gap, and review then rejected the abbreviation.

Added

• EMPLOYER_UNCONFIRMED review error. A project may appear as a bullet under a role only when its engagementId names that role. Client cases with no confirmed employer were rendered under the current employer. Project evidence now states "Employer: ..." or "Employer: not confirmed".
• interests in the profile, printed as the CV's last line; render --interests omit leaves them out.
• profile:check warnings for project domains the matcher does not know (such a project is never selected) and for an engagementId that matches no role.

Changed

• workPreferences.maxOfficeDaysPerWeek is optional, so work authorization can be recorded first. An explicit on-site office-day count then goes to review.

jobsearch 2.3.0

Added

• render --fit. When the chosen CV layout leaves the last page nearly empty or breaks the market's page budget, render compiles the other layouts on the same content and reports which fit, densest first, with page counts. It never switches layout; the agent shows the list and re-renders only with the layout the user picks. The density hint and the cv skill point to it (#7).

jobsearch 2.2.0

Added

• links shipping gate. Every clickable link in a CV or letter must be a well-formed http(s) URL that the profile owns. A stale or invented link now fails instead of shipping unseen (#8).
• jobsearch tools-install --tool typst installs a pinned Typst (0.13.1) into the workspace (#9).
• hosts-doctor reports which MCP server each host will start (#10).
• render --links chooses the profile links on the CV and reports what was included and omitted (#8).

Fixed

• Declared honest gaps no longer license affirmative claims in the cover letter (#13).
• Danish postings are recognised as Danish; an undetermined posting language requires an explicit choice (#14).
• Market detection matches place names as whole words ("bern" inside "Kubernetes" no longer means Switzerland).
• Section headings stay with their content on every Typst version (#7); section order is honoured (#5); every layout renders differently (#4).
• Attributed employer requirements are not candidate claims (#6); the plan never asks to prove a gap (#12).
• Missing toolchain gets install hints (#3); render and gate agree on paths (#11); Hermes entries matched by value (#1); hermetic host tests (#2).

Full details: CHANGELOG.md.