Warning: Full Suite failed for this source. A Full Suite failure report ends these notes.
Merged commits since preview source cac57da:
- fix(engine): the agent reads "carry on": the carry-on word list leaves the engine, and words over work nobody selected are read before anything is offered (#2352)
- fix(opencode): AI-DLC ships no ask-everything rule; one ask before a recursive delete and a push (#2351)
- fix(stop-hook): a setting value next refuses still ends the turn when the agent sent it beside its own --resume (#2349)
- fix(settings): the review setting answers in plain words, with the way back (#2340)
- fix(guard): the phase-boundary verification report is written while the code plan is (#2341)
- fix(opencode): the /aidlc command carries the skill whole, so opencode's 50 KiB cut never reaches the agent (#2346)
- test(live): the routing picker is read by the option offered, and the widening question after a check switch is the person's to decline (#2344)
- fix(engine): a step opens with the words for the person and closes with the notes to the agent, so the Codex, Kiro CLI and opencode tool panels show the question first (#2345)
- fix(engine): a park keeps the person's words, and the step that resumes there carries them (#2336)
- fix(kiro-ide): on Windows, a request PowerShell or cmd.exe would change reaches next through the request file (#2333)
- fix(orchestrate): the code plan step is the conductor's own, and says so (#2339)
- fix(engine): the answer-mode line waits for the stage's first question, and a stage with none to ask gets none (#2343)
- fix(kiro): a step the hook ran for the person is handed over with their words first (#2342)
- fix(settings): a setting typed over the step in hand keeps that step current (#2338)
- fix(status): paused work reads paused, with the way back (#2337)
- fix(kiro): the conductor agent file sorts first, so a bare /aidlc + Enter stays /aidlc and aidlc leads the agent picker (#2335)
- fix(skills): the stage, rules and question rows say what the agent does, not which field to branch on (#2334)
- fix(engine): the option the person picks when the agent offers a change is their request for it (#2332)
- fix(kiro-ide): the prompt hook names the first engine step on an /aidlc line, so Kiro's Default agent starts AI-DLC (#2331)
- fix(dispatcher): a word of theirs that looks like a launcher flag stays their word (#2309)
- fix(kiro-ide): in Command Prompt every AI-DLC reply is kept in a file the result names, and split quoted values are rejoined (#2327)
- fix(status): finished work that ended on a skipped stage reads finished, and the reviewers setting is listed (#2328)
- fix(engine): every step carries what to say and what to do, so the agent reads an instruction instead of a field (#2323)
- fix(copilot): the words typed after /aidlc on Copilot CLI reach the record and the work (#2330)
- fix(codekb): the code scan writes nothing into the repository store, so it runs inside a harness sandbox (#2326)
- fix(codex): the hook replay cache lives in the project, bounded, instead of growing in the OS temp folder (#2329)
- fix: an approval given with "bring the specialists in" stands, and the switch applies from the next stage (#2322)
- fix(reviews): the reviewer is handed the time, so it runs no command for a date (#2321)
- fix(orchestrate): an approval and a skip typed in one message both happen once (#2317)
- fix: a second chat keeps what the person said and joins the work this machine chose, silently; a fresh clone stays as today (#2315)
- fix(records): a setting typed while a question waits is not recorded as its answer (#2319)
- fix(settings): a setting says what it changed, and the work carries on in the same turn (#2318)
- fix(workspace): a piece of work named alone after /aidlc is selected at once, words that may name one reach the agent with the names, and the switch says the way back (#2324)
- fix(memory): AI-DLC's shipped defaults no longer speak as the person's project (#2320)
- fix(guards): admit pre-plan summary confirmation receipts (#2274)
- fix(source): a path Git ignores that no Unit can claim is not an unclaimed change (#2244) (#2286)
- test(source): align cached admission fixture with stage-level plan (#2304)
- fix(engine): next reads the person's own /aidlc line when the agent's command carries none of it (#2316)
- docs: align rule delivery, memory layers, the presence gate, and review caps with the engine (#2291)
- test(answers): the answers corpus through real agents, judged on what was recorded and what the person was told (live check + baseline) (#2314)
- fix: a setting the person asks for in plain words is recorded as set by them, with their words (#2313)
- fix(stop-hook): a chat turn that reads the record file is not an engine run (#2311)
- fix(stop-hook): on Claude Code the carry-on line is Stop hook feedback, not a Stop hook error (#2312)
- fix(kiro): the aidlc agent carries the skill in its context from the start of every chat (#2310)
- fix(kiro-ide): every Command Prompt chat is told what the terminal does, the setup offer says what breaks, and a kept Command Prompt is explained once (#2305)
- fix(review): the reviewer says whether it was told its verdict, and the record keeps its answer (#2308)
- fix(kiro-ide): a message Kiro hands over late is recorded whole, and a turn nothing arrived for is never recorded as empty (#2306)
- fix(kiro): a skill already in the chat is not read again, and the new-work question has no heading (#2307)
- feat(sessions): the gate record points at your message when the engine cannot name the chat (#2302)
- fix(flags): a reply reaches the question it answers, and no word of theirs is spent before anyone reads the line (#2293)
Source commit: d484df9
Nightly test report
Full Suite failed for d484df9bd701 in run 38175815181. The preview build does not wait for these tests.
Failed legs: scope_runs (failure), live_linux (failure), live_macos (failure), live_windows (failure).
Failing tests
t-live-aidlc-answers-corpus.serial(Explicitly selected file executed no test cases (all skipped or empty). For production guard journeys, rerun with --production-guards; for live gates, enable the documented live variable and install/authenticate its CLI.) inlive-claude-sdk-2-Linux,live-claude-sdk-2-Windows,live-claude-sdk-2-macOSt-live-aidlc-input-corpus.serial(Explicitly selected file executed no test cases (all skipped or empty). For production guard journeys, rerun with --production-guards; for live gates, enable the documented live variable and install/authenticate its CLI.) inlive-claude-sdk-1-Windows,live-claude-sdk-2-Linux,live-claude-sdk-2-macOSt-scope-run-moves-late-question-change(1 failed assertions) inscope-runs-macOSa change for one Unit's stage at the one late question > Guard Policy strict: the change reopens beta's NFR Design only; alpha is asked nothing; the question comes back once
t-tui-t29-env-scope.serial(1 failed assertions) inlive-claude-tui-1-macOS,live-claude-tui-2-Linux,live-claude-tui-3-Windowst-tui-t29 env-scope (AWS_AIDLC_DEFAULT_SCOPE seeds new-workflow scope on disk) > invalid env value (AWS_AIDLC_DEFAULT_SCOPE=bogus) errors and writes no state
t176-new-work-offer-second-intent(1 failed assertions) inlive-claude-sdk-1-Linuxt176 P4 new-work offer (orchestrator offers a 2nd intent, sdk live) > describing unrelated new-work while an intent is active offers a 2nd intent; confirm → intent-create → 2 registry rows
Runner errors
live-claude-sdk-1-Windows:INCOMPLETE: required selected coverage was not fully exercisedlive-claude-sdk-2-Linux:INCOMPLETE: required selected coverage was not fully exercisedlive-claude-sdk-2-Windows:INCOMPLETE: required selected coverage was not fully exercisedlive-claude-sdk-2-macOS:INCOMPLETE: required selected coverage was not fully exercised
Failed jobs
- Linux / claude-sdk (2): Linux / claude-sdk 1/4, Linux / claude-sdk 2/4
- Linux / claude-tui (1): Linux / claude-tui 2/3
- Linux / result (1): Linux / result
- macOS / claude-sdk (1): macOS / claude-sdk 2/2
- macOS / claude-tui (1): macOS / claude-tui 1/3
- macOS / scope-runs (1): macOS / scope-runs / test
- Windows / claude-sdk (2): Windows / claude-sdk 1/2, Windows / claude-sdk 2/2
- Windows / claude-tui (1): Windows / claude-tui 3/3
Full logs are in the run 38175815181 full-suite-* artifacts.