Merged commits since preview source 1ee8439:
- fix(ask): a question with one choice is said as its numbered question, never as a question box Claude Code refuses (#2883)
- fix(review): under Guard Policy strict, a redo of a Unit built one at a time, or a team Unit sent back, gets its changed code reviewed with checkpoints off too (#2879)
- fix(scope): a plan switch while a stage runs carries on with the work, instead of ending the turn for the person to say go on (#2876)
- fix(workspace-sync): a token in a repos.json clone address no longer shows on screen (#2877)
- fix(cursor): an edit that arrives as StrReplace is recorded and checked like any other edit (#2873)
- fix(review): a Unit's findings are recorded on the Unit's own approval, and an approval nobody gave records them as Approved autonomously, not as a risk the person accepted (#2863)
- fix(sensors): a team's test-results.md template is checked at Build and Test and Performance Validation (#2875)
- fix(pipeline): Reverse Engineering's step names each link's exact command, and Request Changes no longer reruns a scan the Guard Policy kept (#2865)
- fix(kiro): offer once to pause and carry on in a new chat when Kiro's own record shows the chat nearly full (#2574) (#2864)
- fix(knowledge): the examples the requirements writer and reviewer read show IDs as FR1 and US1.4, the form the traceability sensor reads (#2872)
- fix(review-freeze): a write the agent made on its own after a final review is left and said at the gate, not put to the person as a question (#2871)
- fix(config): a refresh in a folder that is not a Git repository goes ahead without asking to set AI-DLC up there, and doctor drops a false default-scope line (#2866)
- fix(summary): an Approve at the gate of a per-Unit stage run at stage level, open when summary confirmation turned on, goes through (#2870)
- fix(learnings): the learnings questions are asked as AI-DLC prints them, so the record names the question the person saw and the option they picked (#2867)
- fix(plugin): removing a plugin takes its scope back out of the core stage, so the next config goes through (#2868)
- fix(ci): the credential broker ends a response whose upstream sends no data for two minutes, so a stalled model service no longer holds a CI test for an hour (#2859)
- fix(plugins): project {{INVOKE}} and {{TOOL_PREFIX}} in composed plugin prose (#2619)
- fix(engine): publish authority for Code Generation record recovery (#2558)
- fix(counts): the creation line, /aidlc --help and a saved scope say the stage count the plan offer gives on existing code, without Initialization's three setup stages (#2861)
- fix(review): after a Unit's review is recorded the agent goes on to the Unit's next step instead of ending its turn (#2851)
- fix(stages): Build and Test declares its strategy-dependent outputs as optional (#2811)
- fix(hooks): the runtime-integrity refusal names the path that matched (#2806)
- fix(doctor): check the hook file the native engine runs, not the project copy (#2175)
- feat(scripts): a trace audit reads the live legs of a Full Suite run, the passing ones too, and a judgement pass names the friction a person met (#2858)
- fix(orchestrator): emit summary ceremony only when the stage owes it (#2483)
- fix(practices): deduplicate project rules across affirmation dates (#2534)
- fix(review): a route folder claim like src/app/items/[itemId]/ is not refused for an ignored sibling (#2059)
- fix(copilot): the person's redo of a built Unit's Code Generation runs the reopen next --stage prints, instead of being refused every round (#2860)
- fix(checkpoint): retire owned checks on timeout and interruption (#2494)
- fix(sensor): required-sections matches emoji-decorated template headings (#2417) (#2609)
- fix(review): only the reviewer finishing clears the no-reviewer mark (#943) (#2732)
- fix(rules): the preserved-token example shows a requirement ID as FR1, the form the stage files and sensors use (#2852)
- fix(compaction): AI-DLC no longer keeps a compaction note that only repeated the audit row, and the guide no longer promises a warning that never came (#2855)
- fix(summary): an Approve at a gate that was open when summary confirmation turned on goes through, under every Guard Policy; work that runs after the change still asks the summary (#2857)
- fix(ci): the credential broker drops reasoning.summary from Mantle requests, so CI's Copilot journeys get an answer, and logs one line per request (#2856)
- fix(source): resolve embedded Git HEAD from jj-named packs (#2641)
- fix(audit): a history or answers read no longer binds the calling chat (#2805)
- fix(engine): a stage jump's source baseline reads back under every workspace stage (#2543)
- fix(review): under Guard Policy strict, a Unit sent back at its checkpoint or started again gets its changed code reviewed, instead of a refusal whose way on comes back to it (#2854)
- fix: refuse setting flags with next --single (#2654) (#2659)
- test(live drives): the golden journey's person answers a question the agent asked from the stage's questions file without logging it (#2853)
- fix(review): the step after a stage's last review pass names the review brief's command, and the brief tool answers --help (#2850)
- fix(copilot): a subagent's calls are its own and the conductor's are the conductor's, whichever shape Copilot sends (#2582) (#2738)
- fix(build-and-test): distinguish real-browser acceptance evidence (#2489) (#2496)
- docs(install): document removing AI-DLC from a project (#984) (#2473)
- fix(windows): a file the agent writes with a byte order mark is read as written, not refused (#2452) (#2727)
- fix(config): Codex hooks trusted on Windows count as trusted, so setup stops asking for a Trust all already done (#2849)
- fix(review): a finding the work was changed to fix after its review is recorded as Changed after review at Approve, not as a risk the person accepted (#2848)
- fix(guard): on Copilot, a jump held by a step gone out of date is told the next --stage that lets it through (#2847)
- fix(copilot): a copied attempt id is dropped, not refused (#2663)
- fix(kiro-ide): Kiro CLI's own Workflows switch is asked about, checked and turned off like Kiro IDE's, so AI-DLC's reviews and helpers run in Kiro CLI chats (#2840)
- docs(state-machine): a start record missing on this machine carries on with one line under every Guard Policy, as #2843 made it (#2845)
- fix(composer): a tailored plan on existing code keeps Reverse Engineering on every run, and its scores are kept on its first audit entry (#2846)
- fix(doctor): an idle intent's hooks no longer read frozen just because doctor ran (#2824)
- fix(sensors): a file written the way its stage says no longer reports heading findings at the gate; a failed check says what is missing (#2721) (#2844)
- fix(doctor): an update number from an old check says when it was checked, and --export refreshes it (#2830)
- fix(doctor): the diagnostic export timeline shows a skipped stage as skipped, with its skip time, instead of incomplete (#2829)
- fix(doctor): a workflow pulled from another clone is a warning that names that machine, not a hooks error aimed at this one (#2828)
- fix(doctor): a clean project is no longer told to commit the audit entries doctor wrote about itself (#2825)
- fix(state): under Guard Policy strict, a fresh clone no longer stops Code Generation's completion for AI-DLC's own missing start record; it says one line and carries on, and a start record that is here but changed still stops (#2843)
- test(live): the golden journeys run on Windows too, while each takes under the 25 minutes a retry allows there (#2842)
Source commit: f7af394