@quikit/quikpilot-runner 9.0.0 → 9.2.0-rc.2

package/README.md

@@ -26,6 +26,33 @@ Releases follow [SemVer](https://semver.org/): a major version bump means the wo
 shape changed (inputs, permissions) and requires a new pull request regardless of this package's
 version; minor/patch releases are runner-logic fixes safe to pick up on the next re-bootstrap.
 
+**9.2.0 — the AI can look at the code before it chooses which files to change (off by default).** Set
+`QUIKPILOT_TARGETING_MODE: loop` in the runner step's `env:` and the targeting step stops guessing from the file listing
+alone: over a few short turns it may use three read-only tools, `list_files`, `read_file` and `grep` (a plain-text
+search, not a pattern), and then gives the same answer as before, which goes through the same checks. A denied path is
+never read, listed, searched or counted; nothing outside the checkout and no symbolic link is followed; every limit
+(turns, calls, characters read) is told to the AI as an ordinary answer, never an error. Each turn's
+request is also kept under a ceiling of 90,000 tokens (counted conservatively: a path listing and non-ASCII text cost
+far more than characters divided by four): a result that would push the next request over it is not sent, the AI is told
+to answer with what it has, and if even that would not fit the run falls back to the single call. Characters read per run
+are capped at 60,000. If the loop cannot finish
+for an ordinary reason (an older runtime, a timeout, a malformed reply) the run falls back to the single call and says
+so in its warnings; a protected path in the answer, or the runtime refusing the run itself, is never retried. The log
+gains `targeting.loop_start`, `targeting.loop_turn`, `targeting.loop_refused`, `targeting.loop_done` and
+`targeting.loop_fallback` lines, and the targeting record gains an optional `loop` field listing what was read (paths
+and counts only, never content). Without the variable nothing changes. The workflow file is unchanged; the `env:` line
+is the only edit, and it is optional.
+
+**9.1.0 — the code-writing step sees the files the changed file imports.** For every existing file a run changes (code or
+test), the runner now reads that file's own imports (JavaScript and TypeScript, one hop, never what those files import)
+and sends the repository files they point to as extra context. They come after the files targeting named and never in
+place of them. Packages, path aliases (`@/…`), stylesheets and images are not followed; a denied path is never read; the
+extra files share the existing size limits, and if the runtime still finds the context too large the call is retried
+without them. The `codegen.context` log line now also carries `importedPaths`, what was skipped and why (`import: too
+large`, `import: over budget`, `import: unreadable`, `import: denied`, `import: not in workspace`) and how many path
+aliases were not followed. The workflow file is unchanged, so a repository on 9.0.0 picks this up with its next
+"runner version" pull request.
+
 **9.0.0 — a failed run keeps its files.** When a run fails after code generation, the runner writes the files it
 generated, a copy of the last test report and a short `manifest.json` (run, targets, files, the reason it stopped;
 no credentials) to `quikpilot-failure/`, and the workflow's new `Keep the files of a failed run` step uploads that folder