awslabs/aidlc-workflows v2.11.1-preview.20261008.2

AI-DLC Workflow 2.11.1-preview.20261008.2

on GitHub

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

[2.11.0] - 2026-10-08

AI-DLC 2.11.0 rolls up the user-visible changes merged since 2.10.0, including replies read and kept in the person's own words from any chat, requests carried out at once with the way back, Plan Approval asked by the engine, Construction that keeps finished Unit work when the plan changes, leaner scope defaults, configuration that runs beside open work and keeps the team's own files, far fewer permission prompts, a named step when a tool's hooks never run, and corrections on every harness. Upgrade: run aidlc update, then run aidlc config --yes in each project to refresh its harness runtime (aidlc config --harness <name> --yes for each harness when a project has more than one); open work carries on through the refresh. Manual-copy users must replace the complete runtime/<harness>/ tree from aidlc-copy-runtime-2.11.0.tar.gz, which now needs Bun 1.3.8 or newer. Existing intents keep their recorded Guard Policy, sensors, learnings, and summary confirmation; Plan Approval and Collaborators, which they never recorded, follow their scope's new defaults, so outside enterprise an in-flight intent runs its remaining stages with the lead agent only unless /aidlc --collaborators on is typed.

• aidlc version reports 2.11.0 on Claude Code, Codex CLI, GitHub Copilot, Cursor, Kiro CLI, Kiro IDE, and opencode.
• New work starts leaner on every shipped scope except enterprise: a change after an approval is said once and the run carries on, with no stood-aside line on every write, and each stage runs with its lead agent only; /aidlc --collaborators on brings the support agents back for one piece of work. /aidlc --plan-approval on|off sets whether each code plan is shown for approval (only the person turns it off), express and poc ship with it off, and bugfix no longer asks the learnings question or for a summary confirmation. Approval gates and human-turn checks are unchanged.
• Plan Approval is now asked by the engine, which keeps the plan's fingerprint itself. An approved plan is asked about once, even after the chat compacts, the work parks, or the stage rules arrive in parts. While the plan waits only the build waits: a review, a setting, a commit, or another small request still runs and the question stays open. If the approved plan or its test instructions change before the build, the person hears what changed and can say "go back to the approved plan".
• Replies are read in the person's own words at every question (stage gates, summary confirmations, Unit checkpoints, guard questions): "approved", a number, or "looks fine but rename the handler" each do what they say, and every recorded answer keeps what the person wrote. A question can be answered from any chat, picks from a Claude Code or Codex question box are kept as given (one box can answer several questions), a reply already kept is never asked for again, and "How would you like to answer them?" is asked once per piece of work.
• A bare /aidlc in a new session carries on with the work in progress, with no Resume / Redo / Jump / Start Fresh menu, and a fresh clone names the committed work and picks it up in one answer. What the person asks for is done at once, with one line naming the way back: /aidlc --skip <stage> and --add <stage>, a scope change, a Construction setting said in chat ("turn checkpoints off", "build one unit at a time"), and settings typed over parked work. Settings typed with new work reach the work the person picks, a reply answers only the question on screen, and a flag-shaped word inside a request ("add a --version flag") stays part of the request.
• Construction keeps finished work: Units already done stay done across a scope change, a switch between one Unit at a time and stage by stage, checkpoints turned on or off, a jump, or a redo of one Unit's step, and a jump back redoes only what the person named. An approved Unit stays approved on a fresh clone, an interrupted build picks up where it stopped, and the person can approve over a review that did not finish (Approved. The <Stage> review did not finish.) unless team memory locks Guard Policy strict. When Units are built one at a time with Unit checkpoints off, one question covers the per-Unit stage approvals due after the last Unit.
• A composed plan now belongs to its piece of work. To keep it as a reusable scope, pick Approve and save as scope at the gate or say "save this plan as quick-fix" later. The plan offer is short, with the stage table and scores on request, each listed plan shows the engine's own stage count, a tailored plan keeps the name the gate showed, and a custom plan starts from classic's settings with Guard Policy off. A code-findings report is planned on bugfix or security-patch unless the person names another scope.
• /aidlc --project-type <greenfield|brownfield> says whether the work is a new project or existing code; typed mid-workflow, or said in plain words, it scans again and runs Reverse Engineering for existing code, and a new-project folder that gains code before Construction is asked about once. A PDF or Word file the person names is onboarded without asking them to move it, /aidlc --status names the project type, depth, other open work, and only the checks someone switched off, Requirements Analysis sweeps each component against a corner-case checklist, and Reverse Engineering leaves .NET bin/ and obj/ output unread.
• When a tool runs no AI-DLC hooks, the first message now stops with that tool's one step in its own words (in Codex type /hooks and trust all, in VS Code turn on Chat: Use Hooks for the folder, in Kiro CLI type /agent and pick aidlc, in Kiro IDE trust the folder and reload the window) instead of approvals going unrecorded; setup and doctor name the same step, and a Copilot chat says when the hooks have never run. AI-DLC's own workflow commands run without a permission prompt or approval card on Claude Code, Cursor, GitHub Copilot in VS Code, Kiro CLI, and Kiro IDE, while commands that change the machine's AI-DLC install ask first. Agents do file work with their file tools, and the new read-only aidlc engine now, aidlc engine log answers --stage <slug>, and aidlc engine audit history replace date -u and hand-read audit files.
• Any aidlc config run while work is open is now done instead of refused: a refresh, an added harness, a pin, or a settings change prints what changed and its undo, and a refresh ends with Updated. Your open work (<name>) carries on. --download lets a config command fetch exactly the release a project needs (natively, or aidlc-copy-runtime-<version>.tar.gz on a copied project), verifying its checksum and, when gh is installed, its attestation. aidlc config --show shows every section without changing anything, a refresh lists the files it removed, and config and doctor (Harness trees on different releases) name each harness tree on another release with the command that brings it level.
• Refreshes keep the team's own files: .claude/settings.json keeps the project's hooks, permissions, and statusLine; a Codex refresh changes only AI-DLC's keys in .codex/config.toml; opencode.json gets only AI-DLC's entries; AI-DLC's part of .gitignore is one # AI-DLC: local working files block after the team's rules; a clone with Windows line endings no longer reports config conflicts; and a settings file saved with a byte order mark no longer stops aidlc config. The copy archive leaves out .gitignore, AGENTS.md, opencode.json, Claude Code's .mcp.json (a copy starts with no MCP servers), Copilot's .vscode/settings.json, the team's memory files, and aidlc/active-space, so copying a newer release over a project keeps them.
• A plain aidlc update never installs a release older than the one running: on a stable machine running a newer preview it changes nothing and names both ways on (aidlc update --channel stable or aidlc config --channel preview). Update checks stay fast with gh installed, and release verification uses github.com even when gh defaults to an Enterprise host. At a terminal, aidlc uninstall and aidlc system versions prune say what they remove and proceed with no [y/N] question; without a terminal they still need --yes. On Windows, an extensionless bin\aidlc launcher lets Git Bash find aidlc, and a machine switched back to 2.10.0 can switch forward again. A lock left by a stopped command clears even when its PID was reused, transaction locking works on mounts without hard links, and Bolt worktrees work in repositories with submodules and past the Windows MAX_PATH limit.
• On Kiro, the kiro-ide install now serves Kiro IDE 1.x and Kiro CLI's v3 engine from one .kiro/ tree and needs Kiro IDE 1.1.70 or newer, or Kiro CLI 2.24.1 or newer (the kiro install still needs Kiro CLI 2.6 or newer); aidlc config --harness kiro-ide (or --harness kiro) switches a Kiro project between the two installs in place. On Kiro CLI, setup asks for the session model and saves it with the preset's effort in your personal Kiro settings (aidlc config models --session-model <id> sets it without prompts), /aidlc --doctor, --version, and --help work on a native install, and a reply typed to another agent names /agent and aidlc as the way back. On Kiro IDE, the first run walks through trusting the folder, reloading the window, and choosing the aidlc agent, a bare /aidlc that Kiro's menu turns into a specialist still starts AI-DLC, AI-DLC's own commands run with no approval card, and hook cards drop to one before a tool call and one after a shell command, none for a read.
• Codex CLI no longer warns "Under-development features enabled" at every start; setup, doctor, and the hooks-off stop name Codex's own trust step (type /hooks, press t to trust all, then Esc), and the hook trust AI-DLC seeds now matches what Codex checks. Replies and refusals name $aidlc instead of /aidlc, a question box that runs out never answers for the person, and in a project shared with another tool Codex uses its own install. On opencode, setup and refreshes add only AI-DLC's entries to the team's opencode.json, the end-of-turn nudge and the /aidlc command text no longer show in the chat as the person's words, a rejected command ends the turn, and a pasted request with quotes or line breaks starts the work on every shell.
• On GitHub Copilot, AI-DLC's own commands run without an Allow click in VS Code, every step fits VS Code's 20,000-character terminal result, a Windows project carries on across VS Code's two drive-letter spellings, the chat moves on after an approval, and stage runners start only when the person types them. Cursor never throws away the person's prompt, says once which work a chat is on, says when a project is outside a git repository, and keeps background agents out of the foreground workflow. On Claude Code, the generated stage and scope runners start only when typed, and a new chat now gets the state of the work and the offer to rejoin it at session start.
• Long runs carry less: on Claude Code, Codex CLI, Kiro CLI, Kiro IDE, and opencode a step names the rules the chat already holds instead of sending them again, each harness ships only its own tool instructions, and Code Generation no longer runs the advisory linter and type-check sensors after every generated file (type-check still runs at Build and Test). Reviews are refused less often: the engine keeps each stage's findings list, so a decided finding is never reworded or asked about again, an empty placeholder row or a stray ## heading no longer forces a second review, and the traceability sensor stops failing per-Unit stages in plans with no Units.
• Changed default: Guard Policy is off on every shipped scope except enterprise (security-patch and infra were strict, the others relaxed); a team memory holding Mode: strict still wins.
• Changed default: the new collaborators setting is off on every shipped scope except enterprise, so each stage runs with its lead agent only unless /aidlc --collaborators on is typed.
• Breaking for scripts: aidlc engine testing-posture fingerprint --reapprove is retired and fails with a line that says to run next; the engine asks for Plan Approval again when the plan changed.
• Breaking for scripts: a composed plan writes no scope file; aidlc engine scope save --name <name> writes one when the person asks to keep the plan.
• Breaking for scripts: on Windows, install.ps1 in a "Run as administrator" window under UAC warns and asks instead of refusing, a non-interactive run there stops unless it passes -Yes, a full-token session (the built-in Administrator, or UAC off) installs with no warning, and AIDLC_ALLOW_ADMIN_INSTALL is no longer read.

Source commit: 71e7aa5

Nightly test report

Full Suite failed for 71e7aa5bb565 in run 37848615187. The preview build does not wait for these tests.

Failed legs: live_linux (failure).

Failing tests

• t-journey-question-once.sdk (1 failed assertions) in live-claude-sdk-1-Linux
  • t-journey-question-once (sdk): a question still open at the end of a turn is asked once > the routing question left open by Chat about this: nothing is recorded or asked again before the person answers, and their answer is taken

Failed jobs

• Linux / claude-sdk (1): Linux / claude-sdk 1/4
• Linux / result (1): Linux / result

Full logs are in the run 37848615187 full-suite-* artifacts.