v2.108.2: fix(shared): avoid flailing false positives during browser checks

aidd v2.108.2

[2.108.2] - 2026-07-11

Fixed

• Browser verification no longer trips the flailing guard when snapshots are separated by clicks or other tool actions. Repeated-action detection now requires consecutive identical calls, while mixed server-diagnostic commands still use the existing sliding window and file edits reset both detectors.

v2.108.1

aidd v2.108.1

[2.108.1] - 2026-07-11

Fixed

• Interview question generation now receives its actual instructions. The prompt compiler routed every interview iteration to the Q&A answer template, so the generate-questions step sent the agent an answer-format prompt with an empty question instead of the "write the questions file" directive; agents dutifully wrote response files over and over (observed: 30 identical iterations and ~6.5M input tokens before the iteration cap ended the run as an apparent success). Generation now compiles its real directive, counts clean attempts that produce no parseable questions file, escalates the instructions on retry, and ends the run as flailing (exit 75) after three misses instead of looping to max iterations and exiting 0. Custom --interview FILE targets are named correctly in both the first and retry directives, and a file with no parseable questions counts as a miss instead of passing now and breaking the next run.
• A CLI process that dies without unwinding (uncaught exception or unhandled rejection) finalizes its run records before exiting: the terminal failed heartbeat and a fallback ledger entry are written with the scrubbed crash reason, so the web panel reports the real failure instead of reaping the run minutes later as heartbeat-stale with no ledger entry. In the observed incident the backend had already finished and committed its feature while the records froze at "started".
• Wall-clock budgets are enforced before spend, not only during a backend stream: a rate-limit backoff that would sleep across the run's deadline ends the run as a wall-clock timeout immediately (observed: a 31-minute sleep past a 3h budget followed by a doomed 39-second iteration), a loop-top guard stops any re-entry path from claiming work or spawning a backend past the deadline, and triumvirate runs get the same treatment.
• Iteration summaries and the run ledger agree about exit codes. A backend that exits 0 but is reclassified (for example, no AIDD_RESULT emitted records as 73) now gets a summary that leads with the classification and brackets the raw exit ("no AIDD_RESULT emitted [backend exit 0]") instead of a success-looking "finished with exit code 0", and the ledger's backendExitCode field carries the raw backend exit it always claimed to hold. Modes can also declare a run unrecoverable (ModeResult.fatal) to end it with an honest non-zero classification instead of exhausting max iterations.

v2.108.0: fix(shared): use node:path/win32 for win32-platform shell paths

aidd v2.108.0

[2.108.0] - 2026-07-11

Added

• Generated prose now has a single style contract: a new humanize-docs ingredient defines how documentation, changelogs, release notes, and diary entries should read (plain natural language, no AI filler or em-dashes, and structural rules against the rhetorical patterns that make text read machine-written). Eleven prose-writing ingredients (changelog-rewrite, codebase-analysis, dance, deployment-readiness, devdiary-update, diary-entry, document-changes, feature-coverage-audit, grill-with-docs, spernakit-bump, update-screen-map) reference it as their contract, so their output follows one voice instead of each drifting on its own. The ingredient is also invocable directly to de-AI existing text. Non-technical dev-diary entries switch from narrative storytelling to plain-language fact sheets.

Changed

• Eleven single-step wrapper recipes (coderabbit, coderabbit-pr, dance, deepreview, doc2feature, grill-with-docs-init-existing, hygiene, knip, test-application, validate-build, validate-tests) are gone; each merely wrapped one ingredient call, and the one-shot ingredient path already covers that. Launch those ingredients directly from the Ingredients page or a recipe step instead. The recipe index and CodeRabbit guide now point at the ingredients, new-app-from-idea calls validate-build as an ingredient step, and a cross-validation test keeps the catalog's Recipe badges in sync with the recipes that actually remain.

v2.102.0

aidd v2.102.0

[2.102.0] - 2026-07-07

Added

• The project Code tab now renders tracked image files inline. Selecting a PNG, JPEG, GIF, WebP, AVIF, BMP, or ICO file displays the image in the viewer (centered on a dark backdrop, with the usual path/size header and copy-path action) instead of the previous "Binary file." dead end. Images up to 12 MiB are served as data URLs; larger images are declined with a message explaining they are too large to render.

Changed

• Code tab file-tree rows are more compact: each file now renders on a single line with its size beside the name, instead of a two-line entry.
• The Code tab file viewer now shows the backend's specific explanation for files it cannot render (for example, an image over the size cap) instead of always falling back to the generic state message.