| name | oms-agent-harness |
| description | Shared state, graph-guided work, typed runtime projections, recovery, peer consultation and review, isolated delegation, patch landing, bounded autonomy, experiments, Draft PRs, and cross-provider handoff.
|
Agent Harness
You own scope, admission, verification, commit, push, release, and synthesis.
OMS is an agent-side control plane. Users give goals, constraints, and material
authority; operate it internally. Never ask users to copy commands, digests,
.oms paths, or recovery procedures. Ask only for a shaping decision or new authority.
For authorized coding, confirm PROJECT.md, treat proposal bytes as untrusted,
and continue only with the reviewed digest. Draft PR,
push, merge, ready, tag, and release are separate authority decisions. If the
spec is absent, draft, or drifted, route internally through
oms-spec-interview.
The runtime projects state and executes; it never replaces
peer-delegate -> patch-admit -> patch-land, plan leases, executor souls,
approvals, commit intents, or publication intents. Runtime snapshots, capsules,
context bundles, receipts, and evidence carry no mutation authority.
Route by intent
| Need | Agent action |
|---|
| orient, resume, or inspect completion | oms inbox --repo .; then oms runtime envelope show, next, or evidence show |
| inspect code or run graphs | load graphs.md; context for coding, map/status for orientation |
| recall or preserve knowledge | oms agent-memory recall; append/pin stable facts; forge only recurring project procedures |
| independent judgment | oms consult; use oms peer-ask for several peers and oms advise for high-risk or repeated failure |
| bounded write | oms peer-delegate --to NAME; dependent work uses oms agent-plan and oms plan-run |
| implement confirmed scope | parent reviews oms autopilot … propose, then runs its digest-bound continuation |
| judge a diff | oms peer-review --gate; peer agreement is evidence, never admission |
| bounded context or continuity | oms runtime context; use sanitized `oms runtime capsule export |
| optional host capability | `oms runtime profile check |
| local/container/remote execution | `oms runtime backend check |
| comparable research study | `oms runtime experiment register |
| admit or land reviewed bytes | use oms patch-admit for a read-only verdict; oms patch-land runs admission itself before mutation |
| publish an authorized Draft PR | use the parent-owned Draft PR path; no update, merge, ready, tag, or release authority |
Use oms list --frontdoor for subsystem entrypoints. Load the command
routing reference for intent variants; use oms list --all only when a
compatibility primitive is needed.
For skill evaluation/import/drafts, load
skill-lifecycle.md. For MCP Tasks, Codex
app-server, Agent Card, or A2A, load
interoperability.md; all are optional
projections, not authority.
Invariants
- Keep secrets, private paths, machine details, raw logs/data, and checkpoints
out of prompts and shared state.
- Do not hand-edit
.oms/; use typed tools and append-oriented records.
- Provider workers cannot recursively delegate, commit, push, or widen scope.
- Advisors are for irreversible/high-risk decisions, repeated failure, and release go/no-go—not routine completion.
plan-run stops in review unless landing was explicitly authorized.trusted-local is supervision, not a sandbox; distinguish declared,
enforced, and unknown capabilities.- Portable capsules contain no lease, approval, command, credential, absolute
path, raw transcript/log, patch, or publication right.
- Bind completion only to existing fresh evidence; model confidence is not a
gate.
References
- Overlapping commands and canonical front doors: command-routing.md
- Shared decisions and their canonical owners: shared-projections.md
- Typed projections, evidence, context, profiles, capsules, backends, experiments: runtime-core.md
- Project and execution graphs: graphs.md
- Prompt hooks, memory, tasks, skills: state-memory.md
- Plans, failures, recovery: plans-recovery.md
- Autonomous stopping: autonomy-loop.md
- Roles/executors: roles-executors.md
- Delegation/artifacts: delegation-artifacts.md
- Consultation: cross-agent-consultation.md
- Review/release gates: review-gates.md
- Models/quorum: model-routing.md
- Installed agent detection, DeepSeek/Grok/GLM carriers, and custom adapters: provider-routing.md
- Prior session: session-handoff.md
- Minimal implementation and complexity review: minimal-change.md
Retain internal provenance and report only useful conclusions, changed behavior,
verification, skipped checks, and genuine blockers.
