Repository navigation
v3.14.0
Summary
A report release. One report view now feeds a one-screen terminal summary,
Markdown, a self-contained HTML report and its A4 PDF, and two outputs for AI
agents: the AI report (--format ai, JSON Lines) and the AI work order
(--format ai-md), a Markdown document an agent can verify, fix and prove
from. A new hollow channel lists code that reports work it does not do as
review candidates, separately from the score. Every change is listed in the
Changelog. The surfaces are shown in the README under
What You Get; the work order is described in
AI_WORK_ORDER.md.
What moves when you upgrade
- The terminal shows a summary. The default screen holds the score, the
top 5 findings, the top 3 hollow candidates and the next steps. The 3.13
tables are behind--details;--quietprints one line. Scripts that read
stdout should use--jsonor--format ai; JSON only gains keys and exit
codes are unchanged. - Hollow review candidates are new and unscored.
success_without_effect,
input_never_used,failure_reported_as_success,
finally_return_hides_failure,success_ignores_outcome,
effect_never_runsandclaim_without_mechanismappear in their own
section, with how many functions were analyzed and how many could not be
decided. They never change the score or the status band. - Some placeholder findings move. An empty registered method is reported
unless it declares the no-op; a comment marking the body unfinished keeps
the finding; success text is never a declared answer; fixed text where the
docstring promises a state lookup is reported.ignored_parameteris
narrower for named bases and parameter families. Scores of the affected
files move by the existing penalty of each finding. - New outputs.
-o report.html,-o report.pdf(needs an installed
Chrome, Edge or Chromium),--format ai,--format ai-mdwith
--ai-mode,--ai-limit,--ai-token-budget, and--report-dirfor all of
them at once. - Analysis cache
analysis-cache-v53; entries from 3.13 are recomputed.
Known limitations
- The hollow channel follows calls inside one module. A call into another
module, a parameter's method or a value chosen at run time is open, and a
function that depends on one isEFFECT_UNKNOWN, not a candidate. On the
code measured during development that was more than half of the functions,
so many hollow functions are not found. - Docstring promises are checked for pagination, parallel work, aggregation,
rejection, validation, live data and computed results. Retry and cache
promises are extracted butNOT_EVALUATED. A parallel promise counts a
ThreadorProcessas its mechanism only when the function starts it
(.start()on the object, its name, or the list a loop starts); threads
appended to a list and started elsewhere leave the promise undecided. - JS/TS and Go item locators are lines: a line shift changes their ids.
- The token budget is an estimate (characters / 4). On a very large project
the index of every item alone can exceed a model's context; splitting a work
order is not available yet. - The agent work order states its preconditions; the detector does not check
them or the agent's response. - The limitations listed for v3.13.0 still apply.