Outerloop 0.2.0

What's Changed

• Lanes are the backend's word; dispatch always exists; sessions see no GPU by @renmengye in #366
• Brief: a launch snapshot is scope-checked like the final tree by @renmengye in #367
• Intake: every skip is logged, a task label vouches, new Apps read members by @renmengye in #369
• Dispatched wakes are on by default; the operator disarms by @renmengye in #370
• Line seal parents on the kernel's record of the line head, not the checked-out branch by @renmengye in #371
• Every climb ends through one terminal, fresh or resumed by @renmengye in #372
• Follow-up: post the measured number first; a seal with nothing to fold is the session's own commit by @renmengye in #374
• Design pass: the run lifecycle, its states, messages, and what the kernel decides by @renmengye in #375
• Stage 1 of the lifecycle: one inbox, one renderer by @renmengye in #376
• Stage 2 of the lifecycle: messages reach a parked author by @renmengye in #377
• Stage 3 of the lifecycle, first half: submit in review, one publish, a verdict is a message by @renmengye in #378
• Stage 3 of the lifecycle, second half: the end verb and the review top-up by @renmengye in #379
• Stage 4 of the lifecycle: three states, and the follow-up module deleted by @renmengye in #380
• Design docs: auto-merge is off by default and a per-repo opt-in, never the kernel's decision by @renmengye in #381
• Board: a parked run's pill says what it waits on by @renmengye in #382
• Lifecycle stage 5: CI results as messages, direct merge under merge: auto, memory guidance by @renmengye in #383
• outerloop permissions: an explicit way for adopters to update their App's access by @renmengye in #384
• Version 0.2.0rc1 by @renmengye in #386
• Board rows carry the whole hypothesis paragraph by @renmengye in #387
• Board: the squeue card scrolls inside its box on a narrow screen by @renmengye in #388
• Runs keep their hypothesis through review; the sibling view shows it (#360) by @renmengye in #389
• Local compute shares the machine's GPUs across jobs by @renmengye in #390
• The merge step logs why a blessed PR waits by @renmengye in #391
• Design note: what A2A and MCP would give Outerloop by @renmengye in #392
• Messages carry a global id, a context, a recipient and a reply reference by @renmengye in #393
• One message verb, sibling routing through the kernel, numbered headers and the chain view by @renmengye in #394
• Drop the pre-rename names for 0.2.0 by @renmengye in #395
• Version 0.2.0rc2 by @renmengye in #396
• A timed-out local job releases its GPUs once its process group is gone by @renmengye in #397
• A PR that will not self-merge says why on the PR by @renmengye in #399
• HOLD_LAUNCHES holds new kernel launches while everything else runs by @renmengye in #401
• Ending line snapshots push with the kernel's auth, once by @renmengye in #400
• The loop can run on a login node against Slurm, with a QOS and an apptainer path by @renmengye in #402
• start refuses to run when the author's CLI is not installed by @renmengye in #403
• Version 0.2.0rc3 by @renmengye in #404
• init installs the author's CLI, with a pinned installer for claude by @renmengye in #405
• The measurer waits for a finished job's files before calling it a job without a result by @renmengye in #406
• Version 0.2.0rc4 by @renmengye in #407
• Design note: Outerloop on several clusters by @renmengye in #398
• One configurable default Claude model for every role; Vertex pins the auxiliary model by @renmengye in #408
• Release 0.2.0 by @renmengye in #411

Full Changelog: v0.1.2...v0.2.0