All notable changes to this project will be documented in this file.
The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.
1.2.0 - 2026-09-15
spec-kit 0.11.2 added /speckit.converge, which assesses the tree against the spec, plan, and constitution and appends unbuilt work as new tasks. Converge finds what is not built; verify-tasks finds what is falsely marked built. This release wires the two together.
after_convergehook. Tasks that converge appends get implemented and marked[X]too, and those marks need verifying. The extension now registers an optional hook after converge alongside the existing one after implement, so verify-tasks is the final gate once converge reports Converged.PARTIAL (code)/PARTIAL (record)qualifier. Every🔍 PARTIALrow now says which loop to enter. A code gap (symbol defined nowhere, dead, stubbed, behavior absent, file missing with no rename) means the task is not done. A record gap (file renamed, symbol defined and wired elsewhere, only Layer 2 negative, or a false or stale claim in the task text) means the code is done and the task's own line is wrong. Unclear cases are taggedcode. The scorecard shows the two counts separately and the verdict line carries the tag.- Walkthrough action D (demote). For a
NOT_FOUNDorPARTIAL (code)item, flip the task's checkbox from[X]to[ ]after explicit confirmation, so/speckit.convergeand/speckit.implementpick it up. One character changes: no renumbering, reordering, deleting, task-text edits, or Convergence header changes. The report immutability rule is unchanged. - Source-ref rule in Layer 5. Convergence tasks carry
per <source-ref> (<gap-type>). Task parsing now captures it, and Layer 5 reads that exact FR, SC, acceptance scenario, plan decision, or constitution principle instead of searchingspec.mdfor the concept. - Scope statement and converge handoff. The advisory now says what verify-tasks does not do (
[ ]tasks, spec coverage, unrequested code, constitution compliance) and points at converge; the command frontmatter offers a converge handoff next to the implement one. - Convergence-phase fixture tasks.
tests/fixtures/phantom-tasks/gains a## Phase 2: Convergencesection with one planted code gap (T012) and one planted record gap (T011);tests/expected-verdicts.mdcovers both.
- Installs no longer copy the whole repository. The repo shipped a
.specifyignore, a file spec-kit never reads, so every archive install copied this repo's own.specify/project (constitution, templates, scripts),specs/,tests/,.github/prompts and.vscode/into the user's.specify/extensions/verify-tasks/. Nothing consumed them, but the "prevent recursive installation" intent never worked. It is now.extensionignore, the name spec-kit honours, and an archive install copies onlyextension.yml,commands/,README.md,CHANGELOG.md, andLICENSE. Verified against spec-kit 1.0.6.
- Hook blocks adopt the spec-kit 1.0.6 boilerplate. An unparseable
.specify/extensions.ymlis reported rather than skipped silently; a hook without anenabledfield is enabled; dots in hook command names become hyphens in the invocation; a mandatory hook must actually be invoked, not just printed. - Constitution 1.1.0. Article VII now permits a user-confirmed walkthrough disposition to edit the flagged task's own line in
tasks.md(fix a record gap, or demote) and nothing else. Rationale is recorded in the constitution's amendment history. - README: converge pairing section with a comparison table, both hooks documented with the real
extensions.ymlshape, fresh-session advice extended to converge, walkthrough table gains D.
1.1.0 - 2026-09-13
- Command body references to sibling commands use the agent-neutral
__SPECKIT_COMMAND_*__token instead of a hard-coded/speckit.…spelling, so the rendered advisory is correct for every agent (dot, hyphen, or prefixed separators). - Repository links in this changelog and the README pointed at the wrong GitHub organisation.
- #4 reported that spec-kit 0.4.3 rejected the two-part alias
speckit.verify-tasksat install time. spec-kit 0.5.1 (2026-04-08) restored alias compatibility for community extensions, and the alias is kept:/speckit.verify-tasksand the canonical/speckit.verify-tasks.runare both installed. Installation of this release was verified against spec-kit 1.0.6, andrequires.speckit_versionnow excludes the affected versions.
- Step 3a test gate. When the completed task list names a canonical test command (
make check,pytest,cargo test,npm test, or a runner named in the repository's agent guidance) and the environment can run it, the command runs it once in the background and records the exit status and summary lines in the report header. A task whose only claim is "the suite is green" is then✅ VERIFIED (by execution)instead ofWEAK. A run is never fabricated or improvised; when the check cannot run, the report says so. - Checkpoint line. One progress line per
tasks.mdphase (or per ~10 tasks) during the verification cascade, so a long run is not silent. - Report immutability rule. The walkthrough may only append the
## Walkthrough Logsection; it must never re-score a row in the original Flagged or Verified tables.
- Layer 4 no longer blanket-skips SQL. Schema objects (tables, columns, types, views, indexes) still need no caller, but SQL functions, procedures and triggers are callable code and are checked for callers like any application symbol. When a symbol's legitimate caller lives in a sibling component or another language, Layer 4 searches the whole repository rather than the nearest project root, and includes the caller's language extensions in the grep. A spec or contract saying "component Y calls X" is a checkable assertion, not evidence.
- Minimum spec-kit version raised to 0.12.17, the release that resolves
__SPECKIT_COMMAND_*__tokens in extension skills.
1.0.0 - 2026-03-12
/speckit.verify-tasksslash command (commands/speckit.verify-tasks.md) — five-layer phantom completion detector- Layer 1: File existence verification via
test -f/find - Layer 2: Git diff cross-reference supporting four scopes (
branch,uncommitted,plan-anchored,all) - Layer 3: Content pattern matching via
grepfor declared symbols - Layer 4: Dead-code / wiring detection via
git grepacross the repository - Layer 5: Semantic assessment with invariant — semantic evidence alone never yields
VERIFIED - Five verdict levels:
✅ VERIFIED,🔍 PARTIAL,⚠️ WEAK,❌ NOT_FOUND,⏭️ SKIPPED - Structured
verify-tasks-report.mdoutput: summary scorecard, flagged items with per-layer detail, verified items table, machine-parseable verdict lines - Interactive walkthrough (Step 11) for flagged items with Investigate / Fix / Skip options
- Fresh-session advisory banner in both the command prompt and the generated report
- Historical branch verification notes for deleted-file and plan-anchored edge cases
- Graceful degradation when
gitis unavailable (layers 2 and 4 skipped) - Shallow-clone warning
- Full error condition table (missing
tasks.md, no[X]tasks, malformed entries, unwritable report path, etc.) - Project constitution (
.specify/memory/constitution.md) with eight governing principles - Test fixtures:
tests/fixtures/phantom-tasks/— 10 tasks, 5 genuine + 5 planted phantoms (missing file, empty class, dead code, wrong function, behavioral gap)tests/fixtures/genuine-tasks/— 10 tasks, all genuinely implemented with verified cross-referencestests/fixtures/edge-cases/— behavioral-only tasks, malformed entries, glob patterns, zero[X]tasks scenariotests/fixtures/scalability/— 50-task synthetic fixture for session overflow testing (42 source files)
tests/expected-verdicts.md— expected evidence level with rationale for every task in every fixture.markdownlint.jsonconfiguration