You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
v0.4.4: Scaffolded tasks can now fix the code they test: upgrade before verifying your own module
--validate warns about scaffold placeholders left in a task. A file
still carrying TODO in its requirements, criteria or interface passed every
other check, and a run would have sent the placeholders to the agents as the
specification.
The refinement loop proposes fixture rows for the criteria it adds.
Refinement reads converged code and adds criteria about what that code could
still get wrong, which is often an input the fixtures do not contain, and a
criterion no row reaches yields a test that passes whatever the code does. refine_acceptance_criteria now runs the fixture proposer over the refined
bar when it finishes, lists the added criteria first, and logs how many have
no data behind them. Like propose_fixtures itself, it writes a report for
review and never a fixture.
Fixed
A scaffolded task could never fix the code it was testing.--scaffold
wrote interface.module as the file's path relative to the project, such as billing, but a run installs the code into outputs/agent_src/code/<task>/
and runs pytest from the project root. So the tests imported the original
file, the coding agent patched the copy, and no fix ever reached a test. The
module is now outputs.agent_src.code.<task>.<file>, the path the bundled
tasks already used. Found by an audit and reproduced end to end: a patched
copy returning PATCHED against tests that saw ORIGINAL. A regression test
now runs that same check. Affects every task scaffolded with 0.4.3 or earlier;
change the module: line of one by hand to fix it.
The MCP qikly_scaffold tool and --scaffold named the same task
differently, BILLING against BILLING_VERIFY. The suffix is now added in
one place, so they agree.
Scaffold's next step now runs --validate --tasks <task>. Unscoped, it
also checked every bundled example, which buried the one result that mattered.
The TODO warning matches only scaffold's own placeholders, so a
requirement that mentions TODO comments is no longer flagged.
Scaffolding a module on a different drive from the project failed on
Windows.os.path.relpath has no answer across drives and raises, so --scaffold and the MCP qikly_scaffold tool stopped with "path is on
mount 'D:', start on mount 'C:'". A module on another drive is now named
by its file, with its seed path written absolute. Found because GitHub's
Windows runners are laid out that way, the checkout on D: and temporary
folders on C:, and reproduced locally on a mapped drive.
Changed
The VS Code install button and the Claude Code line use uvx. The
button wrote a bare qikly-mcp, which pip on Windows often installs off PATH, so the first click frequently produced a server that would not
start. It now writes uvx --from qikly[mcp] qikly-mcp, the command the MCP
Registry listing already uses, which needs uv and no separate pip install.
In Claude Code the one line is claude mcp add qikly -- uvx --from "qikly[mcp]" qikly-mcp.
The pip route stays documented in docs/mcp.md.
Fewer places saying the same thing. The README's MCP section is a short
pointer to docs/mcp.md rather than a second copy of it, the troubleshooting
page links the explanation of why runs stall at the unit stage instead of
repeating it, and design_3_mechanism.md defines the task file's three parts
once. The README also gains a free first step near the top, qikly --explain CALC_TAX, which needs no API key.
A place to say how trying it went. The README and the quick start link a
Discussions thread for reports, whether a run worked, stalled or never got
past install. The MCP docs now say what has been verified: the server is
tested in VS Code and in Claude Code, and the setup for other hosts follows
the standard but has not been checked end to end.
--scaffold writes one task file, not two. By default it writes <NAME>_VERIFY.yaml, the task that tests the code you already have, and --fresh writes the task for a new implementation of the same interface
instead. It used to write both and ask you to delete one. It now ends with
numbered next steps that name the input file to add and put --validate
before the run: the old Then: qikly --tasks X failed straight away, because
scaffold never creates input data. Its comments use the decisions versus
consequences rule, and the landing page links the quick start on your own
data beside the install commands. The MCP qikly_scaffold tool returns the
same default task and names the file to save it as, so step 1 of the quick
start also works from inside VS Code.
Quick start is its own page, docs/QUICK_START_ON_YOUR_OWN_DATA.md.
The README kept its three-command demo, but everything needed to try qikly
on your own module sat after about 2,600 words of explanation. The new page
starts with the demo and five steps from --scaffold to a first run,
including the one scaffold leaves to you, adding input data, and then keeps
the command table and everything the old USING_YOUR_OWN_DATA.md had. That
page is removed rather than left behind as a pointer.
The split between requirements and acceptance_criteria now has a rule
you can apply. The docs said spec-level statements go in one and checkable
edge cases in the other. That sends a measurement convention or an exemption
into the withheld half, where it produces the stuck loop the very next
paragraph warned about. The rule is now decisions versus consequences, with
one test for any line: could two competent developers, given only the
requirements, legitimately disagree about it? The authoring guide, the
mechanism write-up, the starter task and the --scaffold output all say so.
The starter task had followed the old rule: its criterion that a quantity of
0 is rejected was a decision its requirements never stated, and the
requirement is now there.
"Doesn't a failing test give the criteria away?" is answered where the
objection is raised, in the README and on the landing page, rather than
only in the case study.
The stale-name guard matches whole words, so a test named after the
package, such as test_qikly_run, no longer fails CI. It still catches a
path that names the old project.