v0.16.0

[0.16.0] - 2026-07-11

Breaking Changes

• JIRA and Designer are now opt-in integrations. The core install is 100% local by construction — the tool groups that talk to an external service (JIRA → your Atlassian instance, Designer → claude.ai) are no longer exposed by default. Enable them with specship install --with-jira and/or --with-designer; a later plain re-install preserves the opt-in. If you used these tools before, re-run install once with the flags (a blocked call tells you exactly this). Designer is labeled experimental — it drives claude.ai through a debug Chrome session and can break without notice. JIRA tools are never auto-allowed, so Claude Code prompts per call.
• The dashboard Chat page is gone. It duplicated what the graph, specs, and search pages (and Claude Code itself) already do, so it was removed along with its /api/chat endpoints. The reviewer workflow it hinted at — read the plan, approve or reject with a comment, follow the revision — lives on the Workflows + Runs page.

New Features

• SpecShip compacts its output for smaller models. When a session runs on Haiku or Sonnet — detected automatically on every default install (the per-prompt hook reads the session transcript, so mid-session /model switches track too; the status line and SPECSHIP_MODEL are additional channels) — code-graph tool responses compress their prose scaffolding — boilerplate notices become terse one-liners and, on Haiku, long dependency lists cap with an explicit "+N more" — while source code, paths, and line numbers stay byte-for-byte intact. Compacted responses say so and name the off switch (SPECSHIP_COMPACT=0); on frontier models nothing changes.
• The docs site's reference pages are now generated from source. The CLI command list, MCP tool list, environment variables, and supported language/framework matrix on specship.cc are derived from the code at build time; the site build and the test suite both fail if a committed reference block drifts from source — documented-but-fictional commands can no longer ship.
• SpecShip now steers the agent toward the graph. specship install adds a lightweight per-prompt nudge that reminds Claude Code to answer structure/flow questions with specship_explore before reaching for file reads — the pattern our benchmarks show saves the most time. It only speaks in projects with a SpecShip index, and SPECSHIP_NO_STEERING=1 turns it off entirely. specship uninstall removes it.
• Rejecting a workflow run no longer destroys the work. Reject now parks the run as rejected with its worktree and artifacts intact, and resuming drives the gate's on_reject revise prompt with your rejection comment before pausing again for re-review — a real feedback loop instead of disposal. Deleting a worktree is now only ever done by the new explicit specship workflow purge <runId> (also available on the run page); cancel keeps the worktree too.
• Workflow verify steps stopped false-failing for environmental reasons. Test steps in the bundled workflows now install dependencies and build the worktree before running tests, and workflow shell steps run on the same Node runtime SpecShip itself uses — so a verify failure means the tests failed, not that the environment was missing.
• "Verified" now means proven. A spec can only be promoted to verified when a test linked to it as evidence has passed. Declare evidence with a verifies: block in the spec (- <test-file>:<test-symbol>) or an @verifies REQ-X comment on the test; both are indexed as durable test links. Specs without evidence stay implemented and are visibly flagged, and a failing linked test demotes only the specs it evidences — a green suite no longer blanket-promotes everything.

Fixes

• Release bundles always package the current dashboard. The bundle build now builds the desktop UI itself (installing its dependencies from its own lockfile when missing) instead of copying whatever earlier build happened to be on disk — a standalone or offline bundle build could previously ship a stale dashboard, or none, without failing. A bundle that ends up without its dashboard now fails the build loudly; SPECSHIP_SKIP_WEB_BUILD=1 remains the explicit opt-out for non-shipping builds.
• Published benchmark numbers can no longer drift from measurement. The README's performance numbers are now rendered from a dated, version-stamped results file written by the A/B benchmark harness, and the test suite fails if the README block and the measurements diverge — or if a hand-typed percentage claim appears anywhere outside the governed block.
• Doc comments on exported symbols are now indexed. A /** ... */ or // comment directly above export function / export class was silently dropped by the extractor, which also disabled the @implements / @verifies comment-link backstop for exported symbols.
• Asserted spec links now survive a full reindex. specship_link_assert writes the link into the spec file's implementations: block (idempotently, creating the block when needed), so links are git-versioned, reviewable in PRs, and rebuilt from the file on every index — instead of living only in the local database and silently vanishing on specship index.
• No more confident $0.00 for unpriced sessions. When a session's model has no pricing row (e.g. a brand-new Claude model), the dashboard now marks it "unpriced" in the cost tile, sessions list, and session detail instead of showing $0.00 — and cost totals disclose how many unpriced sessions they exclude. Adding a pricing row heals everything retroactively.
• Transcript ingest shows its parse coverage. The dashboard now reports how many transcript lines the latest ingest pass parsed vs. skipped, so a Claude Code format change surfaces as a visible stat instead of silently missing analytics.