awslabs/aidlc-workflows v2.11.4-preview.20261105.1

AI-DLC Workflow 2.11.4-preview.20261105.1

on GitHub

Warning: Full Suite failed for this source. A Full Suite failure report ends these notes.

Merged commits since preview source 17c123c:

• feat(knowledge): a PDF is read where pdftotext is missing, with nothing to install (#3548)
• fix(reading): a question about where the work stands is answered while a gate or question waits, and the decision stays open (#3551)
• fix(record): a Claude Code wake-up (a /loop tick or a scheduled task) is no longer recorded as your reply at a gate (#3552)
• feat(knowledge): one line names the documents the knowledge base cannot read yet, and doctor lists them (#3523)
• fix(kiro-ide): a PowerShell terminal whose shell integration did not load keeps AI-DLC's replies in a file and says what to check (#3446)
• fix(kiro): switching a Kiro CLI (V2) project to Kiro removes the agent files Kiro upgraded from AI-DLC's (#3344)
• feat(units): in team mode each hand-over between people is offered as one question, not a command to type (#3558)
• fix(doctor): the diagnostic export no longer carries a paused Unit's reason or next-action text (#3534)
• feat(plugins): a plugin's house rules land in a section of their own, at the level it names or you choose (#3545)
• test(user-stories): t238 checks the step hands the lead the team's knowledge, not whether the agent opened it (Fixes #3556) (#3557)
• fix(state): work an older AI-DLC recorded resumes after an update, brought up in place with nothing asked (#3500)
• fix(rules): the rules each stage hands the agent leave out the rule files' template comments (#3550)
• fix(reading): a question about where the work stands is answered from the status, not asked about as new work (#3543)
• fix(build-and-test): Build and Test says which Units have no code, at its gate and when you approve it (#3547)
• fix(jump): a jump ahead in Construction names the Units whose steps it skips, in its own line (#3544)
• fix(stop-hook): under Continue automatically, the turn ends while the agent's background helper works (#3538)
• feat(knowledge): the knowledge base reads Word and Excel files itself, with nothing to install (#3537)
• test(units): the units-block scope run asks the Units Generation approval once, at the gate the engine opened (#3540)
• fix(compose): Approve creates the plan you saw, whatever base the agent types for it (#3533)
• fix(doctor): the diagnostic export shows Construction autonomy, walk order, jumps and Unit progress, with Unit names hashed (#3532)
• test(team-unit): on Windows the merge harness traces git on a file's retry, not its first attempt (#3528)
• test(weights): the CI test weights catch up with two weeks of new test files (#3522)
• test(team-unit): own-writes-not-dirty copies its built seeds instead of rebuilding them (#3525)
• test(team-unit): build each replan file's starting project once and copy it per case (#3524)
• feat(inception): Inception reads the team's documents and asks only what they leave open (#3513)
• feat(plugins): a plugin's commands run as aidlc in any shell and /aidlc in the chat (#3515)
• test(e2e): a passing Windows Codex file keeps a copy of fixtures still held, instead of being failed (#3520)
• feat(audit): each decision of yours records your git name, and status says who approved (#3514)
• feat(audit): a question's record keeps the choice the agent recommended, beside its choices and your answer (#3508)
• fix(config): a project an older AI-DLC copied in refreshes, its own unchanged files told from the person's (#3503)
• fix(doctor): Claude Code's trust row passes on AI-DLC's own entry and says once when every Bash command is allowed (#3504)
• fix(rules): on Claude Code auto mode lets AI-DLC's agent dispatches run, and the agent gets the stage's rules as it starts (#3505)
• fix(config): config's lines that name the command again never print the release mirror (#3502)
• fix(tests): a question box's answers back a choice the agent left to itself naming their message (#3424) (#3425)
• fix(units): each Unit's stage notes are kept in the Unit's own folder, not a stage folder beside it (#3494)
• fix(validity): under strict, an edited finished document is asked about once in a chat, not on every step (#3499)
• fix(config): config names .claude/settings.json when it is not valid JSON, and leaves it as it is (#3496)
• fix(check): the typecheck runs tsc with an 8 GB heap, so it finishes instead of crashing at Node's default ceiling (#3517)
• fix(install): the Windows install works when the profile folder has a non-ASCII letter (#3495)
• fix(codex): diagnose root instruction overrides in doctor (#3338)
• fix(units): Request Changes at a team Unit's end-of-Unit approval keeps the Code Generation you skipped (#3497)
• fix(code-generation): a second Request Changes in a row at a side-by-side build's checkpoint hands the agent your second words (#3498)
• fix(claude): a settings.json without AI-DLC's hooks gets config in the hooks-off line, and doctor gives each file its repair (#3490)
• fix(validity): under strict, lining up a finished stage's Markdown table is no change (#3378)
• fix(doctor): doctor no longer says the runtime graph is out of date after aidlc config (#3489)
• fix(test-pro): sensor tools export main(argv) so native fires do not crash (#3340)
• fix(dispatcher): a folder with no AI-DLC setup yet names aidlc config for that folder (#3492)
• fix(usage): price Sonnet 5.5 cache reads at $0.10 and add a Haiku 5.5 row (#3458) (#3459)
• fix(plan-approval): approve the plan and skip a stage in one reply: the skip is made and the approval is kept (#3488)
• fix(units): a Unit added, split or dropped mid-Construction is offered the go-back to Units Generation first (#3481)

Source commit: bcbf4ca

Nightly test report

Full Suite failed for bcbf4ca3bccf in run 42119561914. The preview build does not wait for these tests.

Failed legs: live_linux (failure), live_macos (failure), live_windows (failure).

Failing tests

• t56-workflow-forward-jump (1 failed assertions) in live-claude-sdk-2-macOS
  • t56 /aidlc --stage requirements-analysis forward jump (sdk) > forward jump marks reverse-engineering [S], pivots to requirements-analysis, logs STAGE_JUMPED/FORWARD, phase stays INCEPTION

Failed jobs

• Linux / claude-sdk (1): Linux / claude-sdk 3/5
• Linux / result (1): Linux / result
• macOS / claude-sdk (1): macOS / claude-sdk 2/2
• Windows / claude-sdk (1): Windows / claude-sdk 3/3

Full logs are in the run 42119561914 full-suite-* artifacts.