Merged commits since preview source 085d534:
- fix(review): a cache you exclude after the review stops blocking approval on a plan with no Units, and the refusal names what changed (#3032)
- fix(swarm): a plan approved again during a side-by-side build names the step that hands it to the build (#3023)
- fix(next): a word after a switch that takes no value is part of the person's request, so a folder or utility they name after it is not taken as AI-DLC's own (#3025)
- chore(lint): aidlc-lib.ts may grow to 2,000,000 bytes before Biome skips it (#3100)
- fix(plan-approval): an answer written in the plan's questions file is kept on the record, not lost when the engine marks the line (#3015)
- fix(locks): a config, update or plugin command killed while it took its lock no longer blocks every later one (#3030)
- fix(redo): a Unit's redo is recorded as a redo, not as the person's request when they never asked (#3028)
- fix(status): a status for a named piece of work says it waits on your code plan approval, instead of Running (#3027)
- test(kiro): on Windows Server the drive answers Kiro's V3 offer with Remind me later instead of waiting an hour (#3022)
- fix(copilot): while work runs, a question is asked in chat instead of Copilot's question box, so the answer counts (#3024)
- fix(jump): a scope typed with a stage jump on running work changes the scope first, then jumps, instead of being dropped (#3013)
- fix(questions): a choice that holds a comma comes back whole when a question is shown again in a new chat (#3014)
- fix(resume): a setting typed with /aidlc --resume when no work is selected is kept for the work you pick (#3021)
- fix(swarm): building the Units side by side from your own branch starts every Unit, instead of failing on AI-DLC's records (#3020)
- fix(recompose): undoing a stage you just skipped goes through, instead of a refusal when the plan leaves out what it reads (#3019)
- fix(config): a copied project's own config keeps its release and hooks when a native aidlc is on the machine (#3018)
- fix(record): an answer from a question box of several questions keeps the person's own picks, not the box's first pick (#3017)
- fix(approve): under strict, a stage's own question rounds are no longer read as changes at the gate, so the person's Approve is taken (#3016)
- docs(classic): the classic scope says where its approvals are, Unit by Unit in Construction (#3012)
- fix(kiro): a plugin's own agent on Kiro IDE and Kiro CLI v3 can run commands and use your MCP servers, like AI-DLC's own (#3011)
- fix(plugin): plugin select with --harness-dir changes the install you name, not the first one in the project (#3010)
- fix(single): running a stage on its own keeps a copy of your work's files for it, and says where (#3008)
- fix(questions): a question with more options than the question box is asked once, as a numbered list, never split in two (#3009)
- fix(plan): "Approve, and add User Stories next" adds the stage and runs it next, even when the approval is recorded first (#3007)
- fix(copilot): Copilot CLI starts without a settings warning in a repo with the Claude install, and runs AI-DLC's hooks once (#3006)
- fix(reverse-engineering): a workspace of hundreds of repositories no longer waits several seconds at the gate, a resume and each link (#3004)
- fix(questions): on Claude Code each question's options fit its question box, so none is refused or left out (#3003)
- test(scope-run): the strict classic walks on Kiro CLI, Codex, Cursor and opencode reach Construction, and three checks that hold on main run again (#3002)
- fix(jump): jumping ahead past a whole phase records that phase as skipped (#3001)
- fix(guard): on Windows, code waits for your plan approval when the project folder is spelled with a short name (#2996)
- fix(skills): the agent no longer tells you a step's internal name, and Kiro asks what you want instead without reading its rule aloud (#2997)
- fix(config): a check you turn off from the chat is said once, in your words, without the file, command and workflow id (#3000)
- fix(stages): Reverse Engineering and the other stages say the lead works alone, not "a lead-only run" with "the sole and final link" (#2999)
- fix(code-generation): every step of an isolated Code Generation run names that run's own next, not the workflow's (#2998)
- fix(jump): going back after a scope switch leaves the stages it dropped not started, so adding one back, returning to it or switching back runs it (#2995)
- fix(questions): after "do not ask me anything", the work no longer asks whether to turn a check off for the whole project or what to keep for next time (#2994)
- fix(answer): an answer in the person's own words at a question with choices records the choice they made, with their words kept beside it (#2972)
- fix(recompose): undoing a stage you just added goes through, instead of a refusal when a later stage needs what it makes (#2993)
- fix(start): a bare /aidlc in a new project asks what to build, instead of showing an error (#2992)
- fix(code-generation): the README and other files the person brought stay as they wrote them, instead of being replaced by the app's (#2991)
- fix(stages): the first summary says the depth and test strategy the work runs at, not a level the agent picked (#2990)
- test(journeys): the release journey fails when the person's files change before their pick of a plan starts the work (#2987)
- fix(switches): asking a new chat to turn a check back on does it in one step, instead of a refused first try (#2989)
- fix(compose): the plan offer says once when the plan skips a stage that a stage it keeps uses (#2988)
- fix(routing): "Separate new piece of work" says the size of the plan it starts, and the agent picks that plan from the real list (#2986)
- fix(jump): going back to an earlier phase no longer records the phase you left as completed (#2984)
- fix(swarm): building the Units in parallel on Copilot hands them out instead of stopping on a size line (#2971)
- fix(record): a question box submitted with nothing picked no longer records its question as the person's words (#2983)
- fix(guard): AI-DLC's own steps run when Windows spells the project folder with a short name (#2981)
- fix(copilot): on Windows a project under a short folder name such as ADMINI~1 keeps its plan approval instead of stopping at Code Generation (#2980)
- fix(record): an answer's and a summary confirmation's record keeps the reply given for it, not the replies to the stage's earlier questions (#2979)
- test(t139): the native evidence reader reads a Windows path quoted the cmd way, ending in a backslash (#2978)
- fix(resume): asking for a redo at a gate redoes the stage, instead of the turn ending with nothing said (#2974)
- test(runner): the merge queue's macOS unit shards run up to three files at once (#2969)
- test(journeys): a golden journey fails when the person's files change before their pick of a plan starts the work (#2976)
- fix(reclassify): "existing code, but skip the reverse engineering" no longer hears that the code gets documented next (#2975)
- fix(copilot): AI-DLC's own Stop note that Copilot sends back to the agent is no longer recorded as your message (#2973)
- fix(record): an answer the agent chose on the person's message says the agent chose it, reading their message, not that they asked (#2970)
- test(answers): the live answers check opens the routing question and the plan offer once, from the person's first line (#2957)
- test(live-fuzz): the live run opens the routing question and the plan offer once, from the person's first line (#2954)
- fix(code-generation): an isolated run's early report names the isolated run's own next, not the workflow's (#2967)
- fix(questions): "do not ask me anything" beside the recommended answers turns the summary confirmation off for the work (#2965)
- test(kiro): the live Kiro CLI files answer Kiro's V3 offer with Remind me later and run on (#2966)
- test(weights): the four unit test files the reading wave added get their Linux and Windows weights (#2968)
- test(fixtures): the mid-inception state marks a skipped stage as a real run does, so adding it is no longer refused in the live fuzz (#2964)
- fix(summary): a message sent before the agent wrote the summary does not confirm it (#2962)
- test(kiro): the Other-then-alternative routing case checks the person's words, not how many next calls the agent ran (#2955)
- fix(review-freeze): picking Fix findings at a blocking check lets the agent fix the reviewed document, then has it reviewed again (#2961)
- fix(recompose): skipping Reverse Engineering after saying it is existing code takes it off the plan, instead of a refusal (#2956)
- fix(reading): the person's replies count from any chat on the work and are kept by their message, every setting they ask for names their message, and engine code cannot read their words as text (#2960)
Source commit: fa8f913