Warning: Full Suite failed for this source. A Full Suite failure report ends these notes.
Merged commits since preview source fa8f913:
- fix(resume): the words you type with a resume reach the work, instead of the open stage running without them (#3139)
- fix(compose): a stage change you ask for at a tailored plan's gate reaches the plan, never dropped while the agent says it was done (#3093)
- fix(pick): the words you type in a clone reach the work picked for you, so a redo you ask for happens there (#3121)
- fix(units): on Windows the Units Generation gate says once when the Unit names make the folder too deep for their records (#3132)
- fix(compose): an approved tailored plan runs at the depth, settings and Guard Policy the plan showed (#3038)
- fix(answer): an answer you never gave is no longer recorded with your reply to another question (#3086)
- fix(sensors): a check that could not run is said once at the gate and recorded as could not run, never as passed (#3145)
- fix(review): a gate with open review findings says which ones Approve accepts as risk, and how to fix or reject one instead (#3142)
- fix(locks): a pinned hook that checks the install lock just as it is let go runs, instead of failing with ENOENT (#3164)
- fix(request): your request is recorded with the spacing you typed, so a quoted phrase with double spaces on Kiro CLI is no longer rewritten (#3152)
- fix(approval): an answer you gave before the stage's document was written does not approve a gate the agent never showed you (#3148)
- fix(record): a key or password you type is masked in the committed record, not written whole (#3117)
- fix(resume): on Copilot CLI a bare /aidlc at an open question is read as yours even when the chat's start is recorded after it (#3099)
- fix(resume): a new chat asks again the verification command or summary question you left open, and the recovery offer names no summary you never saw (#3083)
- fix(review): a cache you exclude in a recorded checkout's own .aidlc-source-paths.json stops blocking approval, and the refusal names that file (#3039)
- fix(plan): a skip on a Unit walk or of an owed Reverse Engineering says how long its undo holds, instead of any time (#3108)
- fix(config): on Windows, config and doctor say once when the project folder is too deep for AI-DLC's records (#3111)
- fix(compose): a plan approved for existing code keeps its Reverse Engineering when the folder has no code yet (#2985)
- fix(copilot): a redo you ask for at Code Generation runs on the first ask (#3059)
- fix(counts): after a plan switch, each stage's start, the progress line, status and the audit row say the count the switch line gave (#3162)
- fix(review): every helper sent is on record, so a review another agent wrote is said on every harness, not only Claude Code (#3118)
- fix(gate): after a setting you type at a gate, the agent is told it was a command, not your approval (#3089)
- fix(swarm): a build that holds unchecked work is no longer told to land it as it is after its plan is approved again (#3138)
- fix(reverse-engineering): a new chat at its last question records your answer and still brings you its approval (#3081)
- fix(reverse-engineering): the code scan runs in a project folder whose name has a space, and the folder keeps its name (#3147)
- fix(plan-approval): while the code plan question waits, plan approval goes off only on a message you send after it is shown (#3140)
- fix(routing): "separate, but use the classic plan" at the which-work question starts the new work on classic (#3134)
- fix(jump): going back to a stage asks you its summary once, instead of saying it can't go ahead (#3053)
- fix(code-generation): a revision picked up in a new chat keeps what you said should change (#3072)
- fix(swarm): a .env or log file Git ignores no longer keeps the side-by-side build from starting (#3095)
- fix(skip): a stage you ask to skip with a scope switch in the same command is skipped where it stands (#3054)
- test(kiro): the Kiro TUI driver answers a question labelled with who would answer it, so the live Kiro checks reach the stage again (#3146)
- fix(recompose): a stage you skipped, put back later in your own words or by a jump, goes through as the undo does, instead of a refusal (#3094)
- fix(code-generation): the approval of a build names the files you brought that it changed or deleted, and how to put each back (#3144)
- fix(plan): a stage you skipped where it stood and then ask back goes back on the plan and runs next, instead of "already on the plan" (#3154)
- fix(guards): a check you ask in chat to change lands on the first try, with no engine error first (#3143)
- fix(codex): a question box you submit with one question left blank keeps your other answers (#3124)
- fix(config): a check switched off in your name from the chat is said to you once, in the setter's own line (#3127)
- fix(code-generation): on Copilot CLI on Windows the code plan has its folder before it is written, so the plan question reaches you (#3141)
- fix(reverse-engineering): reusing the code knowledge base is recorded as your choice only once you have answered its question (#3110)
- fix(plan): approving and skipping a stage in one reply at Practices Discovery's gate no longer hands the agent a refused approval report (#3107)
- fix(plan): after a skip, the line says until when you can undo it, instead of any time (#3056)
- fix(routing): a request in a teammate's clone gets one question, which work it belongs to, naming the plan, on every tool (#3112)
- fix(guard): Request Changes about one Unit's design no longer throws away every Unit's finished design (#3102)
- fix(sensors): Fix findings in work built Unit by Unit has only the Units the check named reviewed again, and a pick after another question counts (#3071)
- test(plan-approval): the redo and change-reopen cases put the reopen in a later second than the person's message (#3163)
- fix(checkpoint): your Request Changes at a Unit checkpoint stands through the agent's own question, instead of being asked for again (#3133)
- fix(resume): a jump or redo naming a Unit the work does not have asks which one, instead of "accepted" and then a stop (#3079)
- fix(questions): a choice that names a price is recorded as written, with no permission prompt for AI-DLC's own command (#3135)
- fix(design): designing the Units side by side asks each Unit's questions in your chat first, so their designs are built from your answers (#3130)
- fix(opencode): a hook that ends before it reads its input no longer throws a broken-pipe error into opencode, so opencode run does not exit 1 for it (#3131)
- fix(park): with the hooks off, a stop you ask for under Continue automatically pauses the work instead of being refused (#3122)
- test(cursor): the stamped-chat turn checks count turns without acting as the chat, so they pass on a slow box (#3128)
- fix(plan-approval): the change you ask for with a Unit's redo answers its code plan question, instead of being asked for again (#3125)
- fix(park): stopping for now takes back only the message that asked to stop, so your earlier reply to the code plan question stays (#3120)
- fix(gate): an answer you gave before the stage's document was written no longer approves its gate (#3114)
- fix(swarm): a plan you approve again while a sent-back batch is rebuilt reaches its build, instead of a dead end (#3119)
- fix(summary): a document the agent saved before your confirmation is saved again by the agent, not asked of you (#3050)
- fix(guards): a revision reported before its review gets the review, instead of asking you to start the stage over (#3123)
- fix(code-generation): a build, its review and its completion check leave another open piece of work's files in the same folder as that work's (#3105)
- fix(kiro): on Kiro CLI on Windows a guard's refusal stops the tool call, as on Linux (#3075)
- fix(routing): "2, you pick its plan next" goes on to the plan question, instead of the agent picking a plan for the new work (#3113)
- fix(validity): under strict, a change you ask for in the reply that approves a stage is not asked about again (#3116)
- fix(review): an approval that asks a finding fixed records that a fix was asked, and the next review checks it again (#3115)
- fix(code-generation): under strict, an edit no Unit claims after the last review is named first, so its one recovery review claims it (#3067)
- fix(scope): confirming a summary and switching to a plan without that stage writes the documents you confirmed first (#3106)
- fix(jump): asking for a stage or phase makes the jump and starts it in one step (#3005)
- fix(state): a pull that leaves git's conflict in AI-DLC's record of the work carries on from your own side, and says where the other side stands (#3103)
- fix(swarm): Units built side by side keep their edits to the design records when they land, instead of losing them with their folders (#3104)
- fix(hooks): a chat started in a folder inside the project is told to start Claude Code or Codex in the project folder (#3101)
- fix(plan-approval): with plan approval off, the line naming the code plan says you can ask to go over it, instead of asking while the build already runs (#3098)
- fix(review): when the wrong reviewer writes a review, the agent is sent back for the right one instead of rewriting its name (#3096)
- fix(questions): a lone letter at a question shown with numbers or in a box is asked about once, not recorded as a pick (#3092)
- fix(swarm): building side by side from your own branch lands each Unit back on it, instead of a refused landing on main (#3091)
- fix(compose): Approve and save as scope always saves the plan, instead of sometimes starting the work with no scope saved (#3052)
- fix(plan-approval): after a resume, your code plan approval is recorded on the message you gave it in (#3090)
- fix(plan-approval): while the code plan question waits, a refused switch no longer offers your earlier "Approve" as the message that asked for it (#3084)
- fix(guard): Reverse Engineering you put back before Code Generation runs when the step names it, instead of a loop with the plan-approval guard (#3088)
- fix(swarm): a committed link to a file outside the project no longer stops the parallel build from starting (#3063)
- fix(plan): "switch this to the express plan" mid-run switches the plan in place, as --scope does, and never sends it to the composer (#3087)
- fix(swarm): the file your git hook refused can be fixed while the parallel build waits for it, instead of a refusal loop (#3082)
- fix(resume): coming back between a stage's finished review and its approval brings the approval, not the stage again (#3074)
- fix(review): under relaxed or off, a review the agent starts after your change no longer tells you the Unit can't go ahead or asks you to redo it (#3080)
- fix(request): the work you start is recorded in your own words, even when the agent retypes them (#3078)
- fix(guards): under relaxed, the line beside each build step says plan re-approval is off, not plan approval, which is still on (#3077)
- fix(guard): on Windows, Code Generation makes its own record folder from PowerShell on Copilot CLI and Kiro CLI, without the wait (#3076)
- fix(swarm): a write held while the Units are built in parallel says why it waits and when it can be made (#3073)
- fix(guard): reading your work's files through a shell variable is no longer refused while the code plan waits (#3070)
- fix(learnings): a note you chose not to keep is never written to project memory or recorded as learned (#3069)
- fix(code-generation): approving the code plan with a change makes the change first, so the approval covers the plan they approved (#3068)
- fix(reverse-engineering): under strict, a clone or copy of the project no longer sends Reverse Engineering back to a full rescan (#3064)
- fix(swarm): a parallel worker writes its Unit's records where AI-DLC reads them, with no stray construction folder in your project (#3062)
- fix(single): a stage you run on its own after the work finished records your answers (#3066)
- fix(construction): a yes to building the Units in parallel applies it, instead of leaving Construction one Unit at a time (#2982)
- fix(construction): after you redo a built Unit's Code Generation, its plan comes back to you once instead of looping on a revise (#3065)
- fix(scope): approving and switching to another plan in one reply records your approval right after the switch, in the same turn (#3055)
- fix(intents): a list of work edited by hand is no longer written over by the next new work (#3061)
- fix(construction): asking for the Units in parallel reads as one change, not stage by stage and then in parallel (#3060)
- fix(review): after Request Changes, a finished review is recorded the first time, whatever round number the reviewer wrote (#3058)
- fix(units): Units planned side by side get the dependency their tests need, so each Unit's tests can pass in its own copy (#3057)
- fix(routing): another plan's name typed alone beside work in progress is read, not dropped for the stage in hand (#3051)
- fix(copilot): on the Copilot CLI, a message you send right as a helper starts counts as yours (#3045)
- fix(record): a subagent's hand-back on Claude Code is no longer recorded as words the person typed (#3049)
- fix(sensors): after you say to approve despite a blocking check, the engine opens your gate instead of sending the agent back into the stage (#3048)
- fix(routing): on Kiro, new work in a clone that needs a tailored plan goes on to the plan, not "Unknown scope compose" (#3047)
- fix(resume): a jump back to an initialization stage no longer says "Jump accepted" before the refusal (#3044)
- fix(opencode): with your editor's file open beside opencode, AI-DLC keeps the words you typed, not opencode's note about the file (#3043)
- fix(opencode): with two chats in one window, a file write and a to-do update are recorded on their own chat's work (#3041)
- fix(routing): a bare yes at the where-does-this-belong question is no longer taken as starting separate new work (#3040)
- fix(approval): an approval reported for a gate you never saw, in a chat where you said nothing, is not recorded (#3033)
- fix(guard): tests and a dependency audit run in a project whose path names audit, instead of being refused as an audit-trail write (#3031)
- fix(single): an isolated Reverse Engineering run in a project with no work records its links and finishes (#3037)
- fix(swarm): continuing a parallel build from a Unit's folder goes on with the build, instead of saying the hooks are off and to restart Claude Code (#3036)
- fix(review): after the recovery review passes, the approval you already gave is recorded instead of asked for again (#3029)
- fix(routing): an answer in the person's own words at the question about where new words belong reaches the plan reshape it leads to (#3026)
- fix(kiro): a message typed while Kiro CLI works is recorded as yours, so a stop parks the work on your words (#3035)
- fix(design): a parallel Functional Design whose builders skipped the questions asks you them, instead of a question none of whose answers moves on (#3034)
Source commit: da59031
Nightly test report
Full Suite failed for da5903196e0a in run 40255612064. The preview build does not wait for these tests.
Failed legs: live_windows (failure).
Failing tests
No failing test files were recorded in the downloaded evidence. The failed jobs below may have stopped before or outside their tests.
Failed jobs
- Linux / result (1): Linux / result
- Windows / codex (2): Windows / codex 1/2, Windows / codex 2/2
Full logs are in the run 40255612064 full-suite-* artifacts.