awslabs/aidlc-workflows v2.10.1-preview.20261007.1

AI-DLC Workflow 2.10.1-preview.20261007.1

on GitHub

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

Warning: Updating from the last release to this preview failed in its checks. Stay on your release or wait for the next preview.

Merged commits since preview source f228a77:

• fix(state): a Construction setting asked for while the plan waits keeps the plan question open (#2077)
• fix(scope): a scope change in Construction keeps the Unit work already done, also on "Continue automatically" (#2061)
• fix(lines): offers are plain questions, and a way back after your own action is said, never typed (#2069)
• fix(units): a review request over an unreadable units block hands the agent the repair, not the person (#2073)
• fix(kiro-ide): a helper's report_progress passes while Code Generation waits (#2071)
• fix(config): a clone with Windows line endings is not a wall of config conflicts (#2057) (#2070)
• fix(switches): the engine carries out a command you typed, never loses a switch in silence, and offers instead of instructing (#2074)
• fix(rules): Kiro IDE steps name the rules the chat already holds (#2023) (#2068)
• fix(routing): words sent to new work are no reply to the question the work in progress has open (#2062)
• fix(summary): a summary answered before its question was recorded counts (#2067)
• fix(plan-approval): while the plan waits, only the build waits; the person's other asks run (#2050)
• fix(switches): a check you type off reaches the work you meant, and "set by you" says where you set it (#2056)
• fix(construction): a change for one Unit at the one late question redoes only that Unit (#2064)
• fix(checkpoints): a Unit review never asked for names its first request, and approve as it is goes over a recovery review (#2063)
• fix(validity): an edit to a finished stage's own document names it and says it stands (#2060)
• fix(units): a units block the engine cannot read is the agent's to fix before Units Generation's gate, never the person's (#2055)
• fix(checkpoints): a Unit the person approves as it is goes on over a review that did not finish (#2048)
• fix(resume): a stage whose questions are answered keeps them on resume (#2054)
• fix(jumps): a jump back redoes exactly what the person asked for, one Unit included (#2053)
• fix(guards): "Turn this check off" says the agent turns it off, with no command to type (#2051)
• fix(park): a setting typed over parked work is applied, not dropped (#2046)
• fix(construction): with reviews off, a Unit done but never recorded is told its receipt step, not the stage again (#2045)
• fix(scope): preserve active Unit work across plan changes (#1926)
• fix(routing): a reply answers only the question on screen, and "go back to the approved plan" works from any chat (#2047)
• fix(rules): a step names the rules the chat already holds instead of sending them again (#2023) (#2043)
• fix(kiro): a helper's refusal in the person's own Kiro CLI chat names /agent and aidlc (#2044)
• fix(kiro): a Kiro CLI answer that did not reach AI-DLC names /agent and aidlc (#2042)
• fix(agents): AI-DLC commands keep their short path, so no approval card asks for them (#2041)
• fix(review): the person's approval goes over a recovery review that never finished (#2040)
• fix(guards): the check setter names the check in words, never as a fence (#2039)
• fix(kiro-ide): one hook card before a tool call and one after a shell command, none for a read (#2038)
• fix(review-freeze): the way out names the Unit whose document was refused (#2036)
• docs(construction): receipts are required from the first Unit when checkpoints apply; name the printed receipt step (#2032)
• fix(requests): a scope, skip or add request on parked work is applied, and a terminal guard switch names the step that works (#2037)
• fix(construction): "approve the plan, and run on its own from here" does both, in either order (#2035)
• fix(checkpoints): a Unit checkpoint answered in a new chat approves in one reply (#2034)
• fix(guard): the agent does its own ways on; the person is asked only to decide (#2024)
• fix(sensor): stop false traceability failures in plans with no Units (#2019)
• test(windows provisioning): a handoff file held by an on-access scan is read once it is free (#2033)
• chore(lint): raise the aidlc-lib.ts size cap to 1.8 MB (#2029)
• fix(questions): the way to answer the person chose is reused, whatever the question's wording or their words (#2028)
• fix(config): six fixes to what config, uninstall, a download and the skill tell the person (#2014)
• fix(next): a plan the person names is used, not asked about again (#2025)
• fix(answers): one question box reply backs an answer for each pick it carried (#2026)
• fix(answers): a picker answer carries a note of what was picked, and the protected event list matches the CLI (#2020)
• fix(skill): the composer block loads only when a plan is being composed (#1985)
• fix(hooks): tools work in a folder holding the AI-DLC install, and AI-DLC names the step when asked (#2018)
• test(guard matrix): the document changes (an approved design or code plan edited or deleted while Unit 2 builds, also stage by stage) (#2016)
• docs: make the Layer 1 run command actually LLM-free (#1265)
• test: a nightly golden journey for the fix and prototype scopes, and t276 and the TUI driver hold from a tab (#2015)
• fix(routing): a report stays on a fix scope, a typed park needs no prompt, re-use choices recommend first, picked work says its name (#2013)
• fix(approvals): an approved plan is built as it is, an approval is never a rejection, and Unit checkpoints keep their answers and notes (#2012)
• feat(inception): sweep each component against a corner checklist when writing requirements (#1872)
• fix(plain-words): six fixes to what the person reads, one commit each (#2010)
• feat(questions): ask the interaction-mode question once per piece of work (#1584)
• test(live guards): a typed Guard Policy switch ends at the engine's already-applied step (#2009)
• fix(config): a Codex refresh changes only AI-DLC keys in config.toml and keeps the rest (#1999)
• fix(config): set up and diagnose a pinned project against its pinned release (#1470)
• fix(install): verify release provenance on github.com when gh defaults to an Enterprise host (#1491)
• fix(review brief): a first review needs no empty Prior table, Request changes says what happens, and the person's findings table stays narrow (#2002)
• fix(kiro): read only the person's own turn when Kiro Crew wraps the prompt (#1870)
• fix(guards): a check turned off with a request is for that request, never the open work (#2006)
• fix(config): Codex setup names the hook trust step in Codex, not config trust (#2005)
• ci(merge-queue): macOS runs its selected unit files in two jobs instead of twelve (#2011)
• fix(guard-policy): a check the person asked to turn off reads "set by you", and an unsourced switch makes no claim (#2003)
• fix(counts): every stage count the person reads during a run counts the stages after Initialization (#1971)
• test(person turns): a menu answer backs the question the agent logs after it (#2007)
• fix(answers): every answer keeps the person's own words, and a question box's reply is on record as given (#2004)
• fix(skills): a change line is said once, and no summary or recap tells it again (#2001)

Source commit: 994eb96

Update from the last release

aidlc update from the last release to this preview failed in the preview run.

Nightly test report

Full Suite failed for 994eb96a8984 in run 37620866838. The preview build does not wait for these tests.

Failed legs: scope_runs (failure), live_windows (failure).

Failing tests

• t-guard-matrix-classic (1 failed assertions) in guard-matrix-macOS
  • classic as it ships, two Units > Unit 2 edits Unit 1's approved file without claiming it, Guard Policy off, review advisory, plan approval on, classic as it ships: nothing refuses or asks the person again, and the run reaches done
• t-tui-t101-memory-lifecycle.serial (1 failed assertions) in live-claude-tui-3-Windows
  • t-tui-t101 (memory.md start→approval lifecycle through a driven gate) > approval-handoff stage creates a faithful memory.md and renders its gate

Failed jobs

• Linux / result (1): Linux / result
• macOS / guard-matrix (1): macOS / guard-matrix / test
• Windows / claude-tui (1): Windows / claude-tui 3/3

Full logs are in the run 37620866838 full-suite-* artifacts.