Stijn

Replay recorded promotions through a tree's own rule without models or spend. Read which promotions the rule supports, what unsuccessful attempts missed by, their A/A controls, and hours and dollars per promotion. Compare a baseline and a candidate that changes one rule.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

Start, continue, keep, pause, copy, watch and stop authorised Anabasis v2 campaigns through `bun run runs`: launch a campaign under a name, go on with one from its log on another login or commit, keep one going across logins until it ends on its own terms, pause one at a clean point, copy a stopped one to another worktree, move one to another host, read each run's status and each login's room, reach the other hosts, and stop one and release what it held. Use when the operator asks to start, launch, resume, continue, move, keep, pause, copy or stop a campaign or a run, which logins have room, how the runs stand on any host, whether a campaign is ready for its next move, or why a move was refused. Owns the run's life on the host; run-improvement-campaign reads the table, and whole-run-investigation reads what a run did.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

The harnesses a campaign keeps, read outside a run: which there are and where each came from, what each is made of and what moved from its parent, what each was measured to be, trying one on a task, the audit before a promotion and the blind review at a handover, whether a part earns its place, and changing what makes a harness or what a seat is told about one. Use when the operator asks which harness is current, what a Builder changed, whether a harness is ready to promote or hand over, to run a kept harness on a task, or to change the harness contract. Reading one run is whole-run-investigation's, wording a claim attribution-and-proof's.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

Put a change on the one open stack of Anabasis v2 pull requests: find the real top, keep local work while moving to it, decide between appending and a new pull request, shape commits that each pass alone, run the checks CI runs (by hand where the runners do not), publish through the gating pre-push hook, write the body, repair your own children when a parent moves, and land a stack with `bun run land`. Use when asked to open, append to, stack, restack, rebase, hop to the latest head, sanitize or land a pull request, or when a push, a hook or a base refuses. v1's stack-hop and setup-pr in one.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

Measure whether a change made the tree simpler: line counts, branching, coupling, interface width, rule count, standing exemptions and model-visible prose across one diff or a stack of pull requests. Use before claiming a pass reduced complexity, when deciding which pull request added machinery, and to read a function's branch count, which no v2 gate reports.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

Run the Anabasis v2 improvement loop: score the five points, read each campaign at its milestone, find the link that holds its next promotion, fix that link's owner, prove the path, record a prediction, launch fresh, watch, read and decide. Use when the operator asks to keep campaigns going, run a pass or an hourly audit, read a promotion, investigate a campaign that is not promoting, or choose the next run. Owns the experiment and its evidence; AGENTS.md section 3 owns the operator's decisions.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

Review changed Anabasis v2 code for reuse, simplification and efficiency, then apply the fixes. Report correctness bugs separately. Starts from the lint plugins, the whole-tree census and the not-slop ledger, measures before and after with reduce-complexity's scripts, and holds a pass to its budgets and the repository's rules. Use as /simplify on your own diff before a commit, or when the operator asks whether code is simple enough, why something exists, to go one layer higher, or to remove something and find another way.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

Test an uncertain Anabasis v2 path at the smallest real layer that can decide it: one fact from a log or a dry run, a walk of one value across the layers, a replay through this tree's readers, a round or campaign rehearsed on a seed with scripted seats and no Referee, or one live seat from a seed. Use to test a join, revisit a recorded position, size a comparison, or check what a kill or stop leaves. Not for reading a finished run (whole-run-investigation) or starting a paid one (run-management).

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

Investigate one Anabasis v2 campaign from its log and turn what it shows into findings with one owner each. Answers whether a comparison's reading (step, no-step, not-scored) may be believed, whether a fail was the solver's or the exam's, whether the harness could know the exam, whether the tasks bracket the solver, whether the campaign is climbing, and where the hours and dollars went. Reads evidence; starts no solve, session or round, and changes no verdict.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

Launch and collect independent Codex sessions (gpt-6-luna at high, xhigh or max; gpt-6.1-sol at low to xhigh) for bounded parallel work, from Claude Code or Codex: one manifest, one detached launcher, one report a session, a drain while it runs, and a retry of incomplete sessions. Owns the Codex transport when another skill sets the questions, as whole-run-investigation does for its lanes. Use when the operator asks for Luna or Sol sessions, a swarm, many sessions, a reasoning effort, a Codex subagent from Claude Code, or a later collection of reports. Distinguish a request to launch from one to wait and collect.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

Zip recorded Anabasis v2 runs for handover at a chosen depth (--light, --medium, --verbose). Use when the operator asks to zip, bundle, archive or send a run, a round, its log, its harness or "everything the model saw", or names a size cap. One command, three cumulative levels, any number of runs, a README, a reading guide and a MANIFEST inside every zip.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

Hand a question about Anabasis v2 to an outside Oracle model as one task document and one zip. The Oracle reads GitHub; the zip carries what it cannot fetch: campaign logs and kept harnesses through `bun run runs zip`, a reader's output, predictions with how each resolved, patches for commits never pushed. Use when the operator asks to prepare an Oracle question, a report request, a problem statement or a second opinion on a run, a reading or a design, or to send the latest runs to an outside model. Produces local files; uploading is the operator's.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

Use after an Anabasis v2 comparison, checkpoint or system change, and before any sentence claims that a harness, a model or a patch improved something. States what one reading may be said to show: the identities it was made under, the counts kept apart, whether it was repeated, which part of a harness it belongs to, whether a mechanism was live in the run, and whether the measurement is valid as distinct from whether it was exploited. Ends in one fixed report.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

The commit-R protocol: a test-driven rewrite of one compartment of Anabasis v2. Commit what is in flight, remove the compartment's tests in an "R:" commit, write them again from a stated hypothesis so the production code fails on purpose, remove the production files in a second "R:" commit, then write the production code again, rewrite the neighbouring functions the old shape bent, and delete what served only it. Use when the operator says "commit R", "commit-R protocol", "remove and rewrite the tests" or "rewrite from scratch", or asks for a test-first pass that looks for a faster, simpler path to the same outcome.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.