[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M76WWBY1Z2C8EM54EYCD1PH8":3,"skill-same:01M76WWBY1Z2C8EM54EYCD1PH8":51},{"skillNo":4,"name":5,"nameZh":6,"description":6,"descriptionZh":6,"repositoryFullName":7,"ownerLogin":8,"ownerAvatarUrl":9,"ownerAccountNo":10,"ownerDisplayName":11,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":15,"categories":16,"keywords":17,"publishedAt":18,"linkOnly":19,"markdownContent":20,"markdownBody":20,"frontMatter":21,"refName":22,"commitSha":23,"skillDirectory":24,"forksCount":25,"repositoryHtmlUrl":26,"repositoryHomepageUrl":27,"ownerHtmlUrl":28,"ownerBlogUrl":29,"ownerDescription":6,"updatedAt":30,"install":31,"resources":44},"01M76WWBY1Z2C8EM54EYCD1PH8","magus-multi-agent",null,"egladman\u002Fmagus","egladman","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F8713153?v=4","01M75H42QB5A4X82APZF4H44GM","Eli Gladman","01M75H42R1K59Z7JWAP621J849",3,"GPL-3.0","https:\u002F\u002Fgithub.com\u002Fegladman\u002Fmagus\u002Fblob\u002Fmain\u002Finternal\u002Fagent\u002Fskills\u002Fmagus-multi-agent\u002FSKILL.md",[],[],"2026-11-12 13:55:43",false,"# Splitting work across agents\n\nCount the WRITE SETS your change needs (the distinct groups of files to edit), not the\nprojects it invalidates. Getting this backwards is the standard fan-out failure. A\none-line edit in a central package invalidates half the workspace and is still one\nedit{{if .Full}}; fanned out, it puts several agents on one file{{end}}.\n\n`{{cmd \"affected\"}} \u003Ctarget> --plan` partitions VALIDATION: which targets to run, grouped\nfor runner balance. It is neither an edit assignment nor a proof of write isolation.\nSo it can veto a fan-out and never license one. One shard means keep the work local.{{if .Full}}\nSeveral mean the testing parallelizes, and the editing is still your question.{{end}}\n\nBefore any edits exist, the shard plan is empty and proves nothing. Finding the\ncandidate paths IS the partitioning work, done with the graph{{if .Full}}, not by intuition{{end}}:\n\n```sh\nmagus graph build                           # first in a fresh worktree, or refs answers \"unknown, not absent\"\nmagus refs \u003Csymbol> --occurrences -o json   # every edit site, column-precise\nmagus explain \u003Cnode>                        # one node's edges and blast radius\nmagus affected ci --plan --stdin            # plan PROPOSED paths, before editing\n```\n\nYou do not need to be asked to split work{{if .Full}}. \"Split this across agents\" is one way in;\nseveral disjoint write sets you can name is another{{else}}: several disjoint write sets\nyou can name are reason enough{{end}}.\n\n{{if .Full}}What costs is unbounded fan-out, not fan-out: a hand-on without a shrinking scope,\na job with no criteria, a principal model on mechanical edits. {{end}}Say what a round\ncosts when the user is deciding, and prefer the smallest fan-out that covers the work.\n\nFan out only after the collision check below REPORTS the jobs disjoint; looking\nseparate is not that check. With only one coherent write set, keep the work local.\n\nThe root agent owns the goal, the budget, the topology, integration and final\nverification, and never hands those out. Above the root is the human in the loop. What\nlands, what is published and what a round may spend are the person's calls: the root\nreports to them and does not decide them.\n\nWhile a row it handed out is live, the root does no groundwork: no reading, searching or editing\nof source, no runs of its own. It plans from the graph and moves jobs; a scout and integration are jobs.\nThe guard enforces this by role, and the role is the live lease: the root may take ONE row with\n`{{cmd \"job exec\"}} \u003Crow>` until `{{cmd \"job exit\"}} \u003Crow>` ([the root's role](reference.md#the-roots-role)).\n\n## Coalesce what the write sets allow\n\nDisjoint write sets LICENSE parallelism; they do not require it. Every worker carries\na fixed load before it reads a line of the diff{{if .Full}} (system prompt, instruction\nfile, routing index, skills, its own graph queries), so ask whether each unit is big\nenough to be worth one{{end}}.\n\n- A depends-on chain is ONE worker in sequence, not two workers in turn{{if .Full}}.\n  Two is two loads for one unit of work, and the second starts by rediscovering\n  what the first just learned{{end}}.\n- For the rest, read the grouping metric (`{{buzz \"workspace.grouping\"}}` through `{{tool \"client\"}}`,\n  or `grouping history` in `{{cmd \"describe job\"}} \u003Cjob>`) and weigh each measured shape\n  against the others{{if .Full}}: chained, split and solo each report the load before a\n  first write and the context, wall time and collisions per group{{end}}. It reports\n  and never decides{{if .Full}}; a person writes any standing grouping policy{{end}}.\n- Where the lens answers unknown, apply these defaults and say in the plan that you\n  did. Merge small disjoint units inside one project. Spawn only when more than one\n  independent unit survives the merge. Fill idle root time from that pool, never by\n  splitting finer.\n\n## Run the graph-engineering loop\n\nProve it before you plan it. A claim a plan or brief rests on cites the output ref of\na run that settled it. A guess nobody ran stays out of the plan.\n\n{{if .Full}}The loop: state the goal and acceptance criteria, partition, hand out, observe,\nevaluate against the evidence, revise ownership or ordering, then integrate. Graph engineering\nis a natural evolution of loop engineering: it improves the partition, never replaces the loop.\n\nAcceptance evidence is an output ref the root reopens (`{{cmd \"query output\"}} \u003Cref>`),\nnever a worker's prose: a worker that ran a filtered subset, or restated its criteria\ninto ones it passed, reports success either way. No job is complete until its criteria\nand assigned check pass, and the root, not a worker, decides whether the top-level goal\nis complete.{{else}}The loop: define, partition, hand out, observe, evaluate, course-correct, integrate.\n\n- Acceptance evidence is an output ref the root reopens (`{{cmd \"query output\"}} \u003Cref>`),\n  never a worker's prose.\n- A worker is not complete until its criteria and assigned check pass.\n- The root agent decides whether the top-level goal is complete.{{end}}\n\n## Keep worker records out of the user update\n\n{{if .Full}}Read a worker's structured result, not its whole event stream. Keep its job id,\noutcome, changed paths, validation refs and unresolved risks; reopen the relevant\nevidence when verification needs it. Persist the complete JSON or JSONL in the\nhost capture or result file. Do not echo it into another tool call or the reply.\n\nFor a job validation, project `problems`, `unchecked` and a count of `permitted`\nfrom `-o json` before it reaches your context. Keep the producer's failure status\nwhen piping. Inspect individual permitted deviations only when they affect a\ndecision; `--dry-run` text intentionally previews the full record. Prepare the\nspec once in a scratch file (`job_spec` below); do not retype its full exception\nlist on every retry:{{else}}Read worker's structured result, never whole event stream. Keep job id, outcome, changed paths, validation refs, unresolved risks.\nPersist full JSON\u002FJSONL in host capture or result file; never echo it into another call or reply.\nJob validation: project `problems`, `unchecked`, count of `permitted` from `-o json`; keep producer's failure status when piping.\nInspect a permitted deviation only when it affects a decision. Spec once in scratch file (`job_spec`); never retype exception list per retry:{{end}}\n\n```sh\nset -o pipefail\n{{cmd \"job fork\"}} --stdin --dry-run -o json \u003C \"$job_spec\" |\n  jq -e 'if .job.id then {job: .job.id, problems, unchecked, permitted_count: (.permitted \u002F\u002F [] | length)} else error(\"missing job validation\") end'\n```\n\n{{if .Full}}Tell the person what changed, what passed, what remains unverified, and any decision\nthey need to make. A routine claim, handled denial, or recorded deviation is not an\nupdate; collapse repeats into the current blocker and next action. A waiver that skips\nvalidation or accepts a material risk gets one plain sentence and its evidence link,\nnever pasted waiver objects, claim lists or guard transcripts.{{else}}Tell person: what changed, passed, remains unverified, decisions needed. Routine claim, handled denial, recorded deviation: not an update; collapse repeats into current blocker + next action.\nWaiver skipping validation or accepting material risk: one plain sentence + evidence link, never pasted waiver objects, claim lists, guard transcripts.{{end}}\n\n## Declare the interface before any job forks\n\nWhen a change adds a shared API, the ROOT names it before any edit. A shared\nAPI is a module several call sites use, a type, an event, or an exported\nfunction. Name its path, every exported name with its signature, and the EXISTING\nsymbols it must reuse, not restate.\n\nWorkers implement the names they were handed and never coin a public one. A name\ninvented at each call site is how one concept ends up with five spellings\n{{- if .Full}}, and how a plan that says \"collapse two state types into one\" ships\na third: each site made a locally sensible choice and nobody owned the set{{end}}.\n\nThen make the declaration gradeable{{if .Full}}, so the names are a contract, not a suggestion{{end}}:\n\n- `symbol` + `present` for each new exported name.\n- `symbol` + `absent` for the name a second copy would predictably take beside a\n  symbol the work must reuse. No gate grades reuse itself{{if .Full}}: `present`\n  holds for a symbol that existed before the job began, and a reference count\n  cannot tell the defining file or an import from real use{{end}}.\n- `symbol` + `unreferenced` for each helper the shared API replaces.\n\nA job that adds or changes an exported API gets a `symbol` + `signature` goal by\ndefault. A stub job lands the stub, and the root runs `{{cmd \"graph build\"}}`. The root\nthen forks the `signature` goal empty. The fork records the indexed stub as the\ncontract, beside a `changed` goal on the same name.\n\nA job's `breaks` list names the flags, JSON keys, exported symbols or stored formats\nyou approved it breaking. A worker that finds a break outside it stops and reports.\nEach unreferenced symbol `{{cmd \"job fork\"}}` lists in the write paths\n(`deletion_candidates`) or `{{cmd \"job wait\"}}` reports is finished or deleted. Confirm\nfirst{{if .Full}}: the graph cannot see reflection, interface dispatch or outside consumers{{end}}.\n\nCheck each declared name with `{{cmd \"refs\"}}` before forking.{{if .Full}} A bare name that resolves\nto more than one definition is graded against one of them, silently.{{end}}\n\nThis applies with zero workers. A solo change across several call sites is a one-job\ntree. Skipping the declaration because nothing is forked is the same failure{{if .Full}}: write\nthe table, check it with `{{cmd \"refs\"}}` when the edits land, then report the change done{{end}}.\n\n## Set one topology boundary\n\nBefore spawning, state the topology: the model per job, and whether isolated worktrees are available.\n\n{{if .Full}}- Fan-out and depth are not capped unless the workspace sets a cap. Spawn as many\n  jobs, nested as deep, as the partition supports.\n- A limit exists only when magus.yaml's `jobs` section sets one. Read it with\n  `{{cmd \"config view\"}}` or `{{tool \"config\"}}`.\n  - `max_depth` and `max_live` make `{{cmd \"job fork\"}}` refuse past them, naming the key.\n  - `default_timeout` bounds a fork that names no `--timeout`.\n  - Honor a cap the user states the same way.\n- Editing costs the workspace nothing; VALIDATION contends: the magus runs a job\n  triggers, which share the machine's build budget. A run kept out of that budget is\n  waits in its own call up to `budget_wait` (default 30m; `--budget-wait 0`\n  refuses at once, MGS3009) and names who holds it; a spawn or `{{cmd \"job fork\"}}`\n  into a full one says how full it is.\n  Serialize validations that share a worktree, even when their write sets are disjoint.\n\nAssign the check from the pipeline the workspace composed, not from convention.\n`{{cmd \"describe target\"}} ci \u003Cproject>` names what `ci` chains and in what order.\n\n- A job gets the narrowest check that covers its edits: the target from that\n  decomposition. Scope it to the job's own tests with the runner's filter after `--`\n  where one exists. A go test check naming exactly the packages of the write paths\n  (`go::go-test . -- .\u002Fpkg\u002F... -run 'Store'`) may filter their tests too. Never the full suite.\n- The integration job owns the full `go-test` and `lint`, and runs every check a merged\n  worker filed as owed. The integrator re-runs the described order, and `{{cmd \"affected\"}} ci` re-proves the whole composition.{{else}}- No cap on fan-out or depth unless magus.yaml `jobs` sets one (`{{cmd \"config view\"}}`, `{{tool \"config\"}}`): `max_depth`, `max_live` refuse past them; `default_timeout` bounds a fork with no `--timeout`. User-stated cap: same.\n- Editing costs nothing; VALIDATION contends for machine build budget. A run kept out waits in its own call up to `budget_wait` (default 30m; a worker's check passes `--budget-wait 0`, so it is refused at once); spawn or `{{cmd \"job fork\"}}` into full budget says how full.\n- Serialize validations sharing a worktree, even with disjoint write sets.\n\nAssign check from composed pipeline, never convention: `{{cmd \"describe target\"}} ci \u003Cproject>` names what `ci` chains, in order.\n- Job: narrowest target covering its edits, scoped with runner filter after `--` (`go::go-test . -- .\u002Fpkg\u002F... -run 'Store'`). Never full suite.\n- Integration job: full `go-test`, `lint`, every check a merged worker filed owed; re-runs described order; `{{cmd \"affected\"}} ci` re-proves composition.{{end}}\n\nA job's check is that narrow target, never the gate. The root gates ONCE, in its own\ntree, after every unit lands{{if .Full}}, which is what stops seven fanned-out workers from\neach running the whole pipeline concurrently on one machine{{end}}.\n\nDecide each job's validation PLANE with its target. Some worker environments cannot\nexecute magus at all{{if .Full}}. Examples are an isolated tree with no usable binary, or a guard\nrouting raw tools to targets it cannot run{{end}}. Such a worker cannot validate what it\nwrites. Mark that job's check ROOT-DEFERRED when you fork it:\n\n- the worker writes the tests, stops at the static checks its environment does run,\n  and says so;\n- the root executes the job's target centrally before verifying.{{if .Full}} Otherwise each\n  worker reports \"done\" with nothing executed, which `{{cmd \"job wait\"}}` refuses to pass.{{end}}\n\nA worker may fork part of its own job. It may not hand it out without shrinking the\nproblem{{if .Full}}; that is the shape that does not terminate, and\nthe cost people attribute to \"multi-agent\" is almost always this{{end}}. These instructions give it a\ndefinitive end without capping depth:\n\n- **Every level narrows.** A child's scope is a strict subset of its parent's. A\n  worker that would hand on its whole job does the work instead.\n- **Every job carries acceptance criteria down.** A child inherits its parent's\n  criteria plus its own.{{if .Full}} A job nobody can evaluate cannot end; that, not the nesting,\n  makes depth dangerous.{{end}}\n- **A retry is `{{cmd \"job fork\"}} \u003Csame id>`, never a new id.** The next holder reads the previous\n  attempt in `{{cmd \"describe job\"}}`, never a transcript. One marked not retryable names\n  what changes first (`{{cmd \"job apply\"}}` to widen, a config edit, or you). Make it, or\n  do the job locally; never re-fork it unchanged.\n\n  A goal unmet twice marks it so: the root re-scopes the job, or serializes it behind\n  whatever keeps breaking it{{if .Full}}. Two jobs with an undeclared dependency each break\n  the other's criteria, and re-forking the failing one satisfies every instruction above\n  while alternating forever. The budget is what ends it{{end}}.\n- **Whatever the parent does not hand out, the parent still owns.** A strict subset\n  leaks by construction.{{if .Full}} Split \"no caller of X remains\" into per-project jobs, and\n  the callers in no project belong to nobody. Every job passes and the goal is\n  unmet.{{end}} Carry a remainder row at each level and close it explicitly.\n\nSTART EVERY SPAWN AT THE ECONOMY TIER, and SAY which model you picked. Move a job up\nonly for work the table below assigns higher. The fit still runs both ways: an\nambiguous API boundary does not get the cheapest model because it looked like less\nwork{{if .Full}}. Judge cost per COMPLETED job: a cheap worker that needs a rerun is not\ncheaper{{end}}.\n\nNaming it is the checkable half: every spawn names a model, or an agent definition\nthat names one.\n\n- Inheriting the parent's model ON PURPOSE is fine and often right{{if .Full}}, such as a hard review under a cheaper coordinator{{end}}.\n- Inheriting it BY OMISSION is the failure. A host defaulting to \"same as the parent\"\n  makes every unnamed spawn the most expensive one, and records no choice.\n\nWhen the tier is unclear, start at economy. A job that fails its criteria there is\nre-forked one tier up, not retried at the same one.{{if .Full}} Economy is a starting\npoint, not a ceiling: same-strength offload is legitimate.{{end}}\n\nTiers are the workspace's{{if .Full}}: a magusfile enum (`magus\\modelTiers`) maps each to every\nharness's model alias, and the guard compares a spawn's tier by rank{{end}}. `--model` takes a tier\nor a model. magus never picks a model live: re-fork to change a tier.\n\n| Shipped tier | Assign |\n|---|---|\n| principal | architecture, ambiguous ownership, public APIs, migrations, security, integration |\n| standard | isolated implementation with a clear contract and a bounded set of projects, and tracing behavior through code |\n| economy | mechanical edits, fixtures, docs, inventory, and lookups the graph answers (exists, depends-on, is-generated) |\n\nIf the host cannot select models or reasoning effort, keep its default. Tool access\nis a separate axis: evidence gathering, scouting, and review get read-only tools\nwhere the host offers them. Never downgrade the root integration pass or final\nrelease gate.\n\nTo prove something, brute-force it at economy{{if .Full}}. Many economy workers finish about as\nfast as one principal worker, for far fewer tokens{{end}}:\n\n- Fork one read-only job per claim, input, or variant, and run them all at once.\n  Give each a script check, so `{{cmd \"job wait\"}}` can pass it.\n- Spawn the `magus-scout` agent for each where your harness installed it.\n- Each reports its command and output ref{{if .Full}}, so evidence settles the claim and not a worker's reading{{end}}.\n- The principal tier only reconciles the claims where workers disagree.\n\nKeep economy jobs short-lived. The brief points at evidence (paths, output refs, the\nrow) and carries no transcript.{{if .Full}} Some providers raise a tier's price past a prompt size.{{end}}\nWhere the brief names that size, the worker reports and stops before passing it.\n\nNesting is allowed when the host supports it, but it creates no new budget and no\nprivate ownership map. A child reports its proposed descendants to its parent before it\nspawns them.{{if .Full}} The store carries each with its parent, model, criteria and write paths.{{end}}\n\nDescendants inherit the ancestor's deny paths and may subdivide only its write paths. A\ncap the user sets applies to the whole tree, not once per parent.\n\nKeep one integration owner per landing checkout, the integrator, never the root. A\nchild coordinates its descendants but may not accept changes outside its own job or\nrelax the root's criteria.\n\n## Seed the partition with magus\n\nChoose the target that will validate the work. `ci` is the release gate, but any\ntarget `{{cmd \"affected\"}} \u003Ctarget>` accepts can be planned:\n\n```sh\nmagus affected \u003Ctarget> --plan\n```\n\nFor a proposed change whose paths are known but not yet edited, plan those paths\ninstead of the current diff (`--stdin --plan`, one repo-relative path per line).\nRead the JSON fields `count`, `max_parallel`, `source`, and `matrix`.{{if .Full}} A shard is a\nhistory-balanced execution group, not an edit assignment or proof of write isolation.{{end}}\nUse a shard only as the first partition. If paths are unknown, query the task's symbols,\nfiles, and projects before producing a stdin plan.\n\n## Prove that jobs do not collide\n\nClassify the union of every job's proposed paths in one call:\n\n```sh\nmagus describe file \u003Cboth jobs' paths>... -o json\n```\n\nRead the facts:\n\n- `overlaps` lists each declaration covering more than one proposed path: a shared\n  write set by construction.\n- `claims[].target` names the target that regenerates a path. Generated outputs have\n  one integration owner and are never hand-edited by workers.\n- `depends_on` carries the owner's direct edges.\n- Affinity stays with `{{tool \"client\"}}` (`{{buzz \"history.affinity\"}}`). Without MCP,\n  run `{{cmd \"buzz\"}} -e` printing `magus\\history().affinity()`.\n- Use `{{cmd \"refs\"}} \u003Csymbol>` when two jobs may touch the same API.\n\nA read-only job has no write set, so it is outside this analysis.\n\nTwo jobs may run together only when:\n\n- The combined classification reports no overlaps. Source write sets and declared\n  outputs are disjoint, or share a file by claiming DIFFERENT declarations of it\n  (`\u003Cfile>#\u003Cdeclaration>`, the default). `--depends-on` is for the SAME one.\n- Neither consumes an API or generated artifact the other changes.\n- Shared manifests, lockfiles, schemas, and agent instructions have one owner. A\n  magusfile or magus.yaml is shared by key claims (`magusfile.buzz#project.targets.test`).\n- Dependency and temporal-affinity evidence does not say they should move together.\n\nONE WORKTREE PER WORKER wherever a write path touches workspace configuration: any\nproject's magusfile, its `magus.yaml`, or a spell source a magusfile imports.\n\nmagus READS those files to load the workspace. A half-saved one stops the workspace\nloading for everybody in that checkout{{if .Full}}, who lose `{{cmd \"run\"}}`, `{{cmd \"ls\"}}`\nand their own tests over an edit they cannot see{{end}}.\nWhen `{{cmd \"job fork\"}}` refuses one, the answer is a separate worktree, not a narrower\nboundary: fork with `--checkout \u003Cdir>` outside the workspace, never by hand.\n\nFork records what it could prove in `write_proof` (`alone`, `disjoint` or\n`overlapping`), and `{{cmd \"ls jobs\"}}` prints it. An overlap is recorded, not refused:\nsequencing two jobs onto one path is your call{{if .Full}}, and the row records\nwhether anybody checked{{end}}. Answer the guard's shared-checkout reminder with a proof or a\nworktree, not a judgment that it looks fine.\n\nProject boundaries alone are insufficient. A declared dependency means group the work\nor serialize producer before consumer. Treat strong hidden affinity as a warning.\nWhen evidence is incomplete, reduce parallelism.\n\n## Fork the job, then spawn the worker\n\nThe order is fixed: fork the row, spawn the worker with a description naming it, and\nonly then does the worker take the lease. A spawn whose description names no live row\nbinds its child to nothing{{if .Full}}, so every write it makes is graded as\nan editor magus cannot attribute{{end}}.\n\n1. Record the checkpoint you are handing out: `{{cmd \"vcs checkpoint\"}} -o name` prints\n   the revision, plus a dirty-patch digest when the tree is not clean.\n2. Fork one job per unit, from flags:\n\n   ```sh\n   magus job fork api-store --model \u003Ctier or model> \\\n     --criteria \"accounts move to the new store; nothing names the old one\" \\\n     --write-paths 'api\u002Fstore.go,api\u002Fstore_test.go,db\u002Fmigrations\u002F*.sql' \\\n     --check \"go-test api -- -run TestStore\" \\\n     --checkpoint \u003Ccheckpoint>\n   ```\n\n   Or fork from a record with `{{cmd \"job fork\"}} --stdin`, the form that carries goals;\n   [reference.md](reference.md) shows one.\n\n   Before every fork, run `{{cmd \"job fork\"}} --stdin --dry-run`: it lists every problem with its code and fix, and writes nothing.\n   `{{cmd \"job fork\"}} --schema` prints every field and the newest `api_version`.\n3. Spawn the worker with the description `\u003Cparent>\u002F\u003Crole> \u003Cjob>`: two words, a parent id and a role joined by `\u002F`, then the job id.\n   - A root job has no parent row, so its first word takes any label\n     (`orchestrator\u002Ffeat api-store`).\n   - A child reads `api-store\u002Ffix migrate`; the guard resolves the job as `migrate`,\n     else `api-store\u002Fmigrate`.\n   - The role is a word for the work (`feat`, `fix`, `review`). A workspace spawn\n     rule may restrict it, and its refusal names the list.\n   - A title naming a row that no fork or apply validated is refused (`spawn-unvalidated`).\n   - Resume first: follow-up on code a worker just wrote goes back to that worker as a message while its prompt cache is warm, at any context size; a fresh spawn is refused (`spawn-warm-worker`), and [reference.md](reference.md#resume-the-warm-worker) says when fresh is right.\n\nWrite paths name exact FILES; a glob too broad is refused as MGS3043 unless permitted. The\ncheck is `\u003Ctarget> \u003Cproject> [-- args]`, never the gate. [reference.md](reference.md)\ncovers claiming one declaration in a shared file and forking from a leased worker.\n\nA fork refuses a job whose sources regenerate an output outside its write paths (MGS3044).\nAdd the output to the write paths, or permit it when the root regenerates it after integration.\n\nRender the prompt FROM the row; never type it. `{{cmd \"describe job\"}} \u003Cjob>` prints the\njob's own criteria, boundary and check, plus what the workspace knows and nobody wrote\ndown{{if .Full}}: the projects the write paths reach, the declared\noutput globs that land inside them, the paths a sibling job is holding, the build\ninputs and workspace configuration that have one owner, and the projects that\nchange alongside the leased ones without declaring a dependency{{end}}.{{if .Full}} Two renders of one\nrow are byte-identical; hand-typed prompts are not.{{end}}\n\nThe guard grades every command a spawn prompt presents in a shell fence or a code span. One\nit would deny refuses the spawn{{if .Full}}, since the worker runs its prompt's\ncommands as written{{end}}. A line that names a command to forbid it (\"never ...\") is not graded.\n\n### What a spawn prompt leaves out\n\n{{if .Full}}A spawn prompt is the parent's own words: the JOB ID, the goal and why, the decisions made, and evidence refs.\nIt leaves out skill loads, command style, the `BAGGAGE` export, write paths, the check,\nhow to commit and push, and attribution: the guard, the harness and `{{cmd \"describe job\"}}`\nsupply or enforce each. A restated guard rule costs every worker tokens and goes stale;\nwhen a worker keeps missing something, fix the guard rule or its message, not the prompt.\nA follow-up message to a bound worker carries only what changed.{{else}}Spawn prompt carries only: JOB ID, goal and why, decisions made, evidence refs.\nLeave out: skill loads, command style, `BAGGAGE` export, write paths, check, how to commit and push, attribution; guard, harness, `{{cmd \"describe job\"}}` supply or enforce each.\nRestated guard rule costs every worker tokens, goes stale. Worker keeps missing something -> fix guard rule or its message, never prompt.\nFollow-up message to bound worker: only what changed.{{end}}\n\nEvery worker prompt includes its row, its JOB ID, relevant graph evidence, and\nany cap the user set for the tree. Require the worker to:\n\n- export `BAGGAGE=magus.lease=\u003Cits id>` before it works: the guard grades its writes only\n  when that is set{{if .Full}}. Add `magus.spawner=\u003Cyour label>` to the baggage, and\n  `TRACEPARENT` where the host has one: magus records them as claims for\n  `{{cmd \"session\"}} ls`, and no verdict is keyed on them{{end}};\n- preserve unrelated changes, stay inside write paths, and avoid generated outputs;\n- run only its assigned magus target, in the foreground with a timeout that covers it, and never end its turn while a run is pending;\n- never wait for the build budget. Its check runs with `--budget-wait 0`, so a refused check (MGS3009) returns at once; it commits and files the check as owed, `\"owed\": \"no seat\"` in `validation` or that goal's `verification_evidence`, with no output ref. `{{cmd \"job wait\"}}` passes that result and the integration runs the check;\n- never exit its row while it waits on a reply from you: ending its turn with the question as its final text leaves the row open and the worker resumable.\n\nEnd every spawn prompt with the step that files the result, as the worker's LAST act\nbefore it reports back. `stop-unfiled` refuses a leased child's first stop until it files\none, and a prompt that leaves it out gets a prose report and a row that never\nmoves{{if .Full}}: the store keeps the write paths held, `{{cmd \"job wait\"}}`\nhas nothing to verify, and nobody notices until a sibling is refused over a path\nthe finished worker no longer needs{{end}}. The step: `{{cmd \"job exit\"}} \u003Cjob> --stdin`, filing the shape `{{cmd \"job exec\"}} \u003Crow>` printed.\nNever inline the result fields in a prompt; the row prints them filled in{{if .Full}}, and\n[reference.md](reference.md) shows a record{{end}}.\n{{if .Full}}A typed result makes verification mechanical. {{end}}Keep unresolved risks mandatory{{if .Full}}, so a\nmis-scoped worker can say so instead of widening silently{{end}}.\n\nA worker writes only its own row and the children it forks. Every other store write\nis the orchestrator's; a refused worker reports it as an unresolved risk and stops.\n\nAfter the spawn, the row carries the job:\n\n- A worker records the base it LANDED on, which can differ from the checkpoint you\n  handed it.\n- A denied worker coordinates and never works around.\n- Every row ends in pass, fail, or NO-RETURN, and the root writes which: silence is\n  not a pass.\n- Only a row with a check or a goal can pass. Fork a scout `--read-only`; it passes only\n  through a script check (reference.md) or by filing a result: a read-only row whose holder files a result is\n  recorded pass, and only a row with nothing filed ends no_return.\n\n[reference.md](reference.md) covers bases, read paths, releasing a path, moving a live\njob's boundary, how magus ends abandoned jobs, and what `{{cmd \"job wait\"}}` checks.\n\n## Declare the criteria magus can check for you\n\nA GOAL is the part of a job's criteria magus grades itself, from evidence the worker\ncannot author. It is a check that passed, or paths or symbols changed, present, absent\nor unreferenced. Write goals in the record's `goals`. `{{cmd \"job wait\"}}` refuses pass until\neach verifies. [reference.md](reference.md) has the table and the rename pattern.\n\n## Observe through the correct control plane\n\nRun workers non-blocking; block on one only when your next action needs its result.\n`{{cmd \"server status\"}} --watch=15s` is for a person, never an agent's background task: an agent waits\nfor its own run's completion notice.\n\nThe root reads `{{cmd \"describe job\"}} \u003Cjob>` or `{{cmd \"job watch\"}} \u003Cjob>` only when it needs to know.\nNever message a worker to ask how it is going, only to CHANGE what it was handed.\n[reference.md](reference.md) covers blocked workers and auditing what a job ran.\n\n## Integrate and verify\n\nThe root never builds. The integrator is a role, never a check: spawned `\u003Cparent>\u002Fintegrate\n\u003Cjob>` unisolated in the landing checkout, it cherry-picks, builds, regenerates and commits.\nIts merge (`hack\u002Fdev\u002Fmerge-job-branches.buzz --apply`) then recomposes the landing\nbranch's unpushed commits into one per merged job, each output beside its source, with no\nmerge commit left and the merged tree byte-identical. A commit a remote ref already holds stays.\n\nAs jobs finish: compare the store to the ACTUAL diff since each checkpoint; run\n`{{cmd \"job wait\"}} \u003Cjob>` before reading each result; reopen evidence with\n`{{cmd \"query output\"}} \u003Cref>`; resolve shared seams once; regenerate declared outputs\nonce; re-plan with `{{cmd \"affected\"}} \u003Ctarget> --plan` if the partition broke; read\n`{{cmd \"diff\"}} --impact` for context only; validate once per wave; finish with\n`{{cmd \"affected\"}} ci`. [reference.md](reference.md#integrate-and-verify) has the\nmechanical wait checks and the wave loop.\n",{},"main","5f0df014d9eaf4aea1a930b0bd13c3338e2642de","internal\u002Fagent\u002Fskills\u002Fmagus-multi-agent",0,"https:\u002F\u002Fgithub.com\u002Fegladman\u002Fmagus","http:\u002F\u002Feli.gladman.cc\u002Fmagus\u002F","https:\u002F\u002Fgithub.com\u002Fegladman","eli.gladman.cc","2026-11-14 12:23:44",{"source":7,"skillName":5,"commands":32,"prompt":42,"zipAvailable":43},[33,36,39],{"tool":34,"command":35},"npx","npx skills add egladman\u002Fmagus --skill magus-multi-agent",{"tool":37,"command":38},"bunx","bunx skills add egladman\u002Fmagus --skill magus-multi-agent",{"tool":40,"command":41},"pnpm","pnpm dlx skills add egladman\u002Fmagus --skill magus-multi-agent","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fegladman\u002Fmagus\u002Fblob\u002Fmain\u002Finternal\u002Fagent\u002Fskills\u002Fmagus-multi-agent\u002FSKILL.md",true,[45],{"path":46,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":49,"sourceHtmlUrl":50},"reference.md","REFERENCE","参考文档",29637,"https:\u002F\u002Fgithub.com\u002Fegladman\u002Fmagus\u002Fblob\u002Fmain\u002Finternal\u002Fagent\u002Fskills\u002Fmagus-multi-agent\u002Freference.md",[52,58,64,70,76,82],{"skillNo":53,"name":54,"nameZh":6,"description":6,"descriptionZh":6,"repositoryFullName":7,"ownerLogin":8,"ownerAvatarUrl":9,"ownerAccountNo":10,"ownerDisplayName":11,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":55,"categories":56,"keywords":57,"publishedAt":18},"01M76WWZ75J70NCG0TR23XDDE6","magus-workspace-rules","https:\u002F\u002Fgithub.com\u002Fegladman\u002Fmagus\u002Fblob\u002Fmain\u002Finternal\u002Fagent\u002Fskills\u002Fmagus-workspace-rules\u002FSKILL.md",[],[],{"skillNo":59,"name":60,"nameZh":6,"description":6,"descriptionZh":6,"repositoryFullName":7,"ownerLogin":8,"ownerAvatarUrl":9,"ownerAccountNo":10,"ownerDisplayName":11,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":61,"categories":62,"keywords":63,"publishedAt":18},"01M76WWX72SY78KG03AVYFE768","magus-vcs-hygiene","https:\u002F\u002Fgithub.com\u002Fegladman\u002Fmagus\u002Fblob\u002Fmain\u002Finternal\u002Fagent\u002Fskills\u002Fmagus-vcs-hygiene\u002FSKILL.md",[],[],{"skillNo":65,"name":66,"nameZh":6,"description":6,"descriptionZh":6,"repositoryFullName":7,"ownerLogin":8,"ownerAvatarUrl":9,"ownerAccountNo":10,"ownerDisplayName":11,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":67,"categories":68,"keywords":69,"publishedAt":18},"01M76WWTJYGRCKKXH948KN6RR3","magus-upstream-source","https:\u002F\u002Fgithub.com\u002Fegladman\u002Fmagus\u002Fblob\u002Fmain\u002Finternal\u002Fagent\u002Fskills\u002Fmagus-upstream-source\u002FSKILL.md",[],[],{"skillNo":71,"name":72,"nameZh":6,"description":6,"descriptionZh":6,"repositoryFullName":7,"ownerLogin":8,"ownerAvatarUrl":9,"ownerAccountNo":10,"ownerDisplayName":11,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":73,"categories":74,"keywords":75,"publishedAt":18},"01M76WWRTBBAPNKM0PA2NG5Q3N","magus-upstream-docs","https:\u002F\u002Fgithub.com\u002Fegladman\u002Fmagus\u002Fblob\u002Fmain\u002Finternal\u002Fagent\u002Fskills\u002Fmagus-upstream-docs\u002FSKILL.md",[],[],{"skillNo":77,"name":78,"nameZh":6,"description":6,"descriptionZh":6,"repositoryFullName":7,"ownerLogin":8,"ownerAvatarUrl":9,"ownerAccountNo":10,"ownerDisplayName":11,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":79,"categories":80,"keywords":81,"publishedAt":18},"01M76WWNTF03H9S5CN22QBX5ZK","magus-test-design","https:\u002F\u002Fgithub.com\u002Fegladman\u002Fmagus\u002Fblob\u002Fmain\u002Finternal\u002Fagent\u002Fskills\u002Fmagus-test-design\u002FSKILL.md",[],[],{"skillNo":83,"name":84,"nameZh":6,"description":6,"descriptionZh":6,"repositoryFullName":7,"ownerLogin":8,"ownerAvatarUrl":9,"ownerAccountNo":10,"ownerDisplayName":11,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":85,"categories":86,"keywords":87,"publishedAt":18},"01M76WWJX1FPSWYD1RB8JXNF3C","magus-sdk","https:\u002F\u002Fgithub.com\u002Fegladman\u002Fmagus\u002Fblob\u002Fmain\u002Finternal\u002Fagent\u002Fskills\u002Fmagus-sdk\u002FSKILL.md",[],[]]