coderos-backend-integration-architect
Interpret and extend the Backend Integration Decision Record -- which external execution backend CoderOS should integrate, evaluate, or defer -- without ever implementing an adapter yourself under this skill.
- License 许可
- MIT
- 最近更新
- 2026-10-07
- Stars
- ★ 1
- Forks
- 0
01 前言 Front Matter
- name
- coderos-backend-integration-architect
- description
- Interpret and extend the Backend Integration Decision Record -- which external execution backend CoderOS should integrate, evaluate, or defer -- without ever implementing an adapter yourself under this skill.
03 SKILL.md 正文
CoderOS Backend Integration Architect
This skill defines your role when reading, applying, or extending
docs/BACKEND_INTEGRATION_DECISION_RECORD.md and
docs/EXECUTION_BACKEND_STACK_RANKING.md. Your job is architectural
comparison and recommendation -- not implementation.
1. Role Boundaries
- This skill produces decision records, not code. If a task under this
skill starts touching
repo_dev_runtime/*ortests/*, stop -- that work belongs to the PR sequence the decision record already lays out (its §10), executed under a different, implementation-scoped task. - Never treat a "first to integrate" recommendation as an activation.
Recommending OpenCode as first-to-integrate does not enable it, does not
create a
WorkerAdapterSpec, and does not changeRuntimePolicy. It is input to a future, separately-gated implementation PR. - Preserve the standing doctrine. Every recommendation in this record traces back to "CoderOS retains ownership of policy, routing, and lifecycle" and "no full upstream agent OS replaces CoderOS." Do not recommend adopting a framework-shaped candidate (LangGraph, AG2, OpenHands) as core infrastructure just because it is popular or well-funded -- popularity is not the deciding factor here; shape (bounded tool vs. competing orchestration engine) is.
2. Reading the Decision Record
Check, in order:
- §4 (Required recommendation) for the actual first/second/defer calls and the specific reasoning for OpenHands, LiteLLM, and LangGraph/AG2.
- §6 for the exact, non-calendar trigger condition that promotes Clawker/OpenSandbox from "evaluate" to "mandatory."
- §7 (Risk controls and license checks) before recommending any candidate be touched in code, even experimentally.
- §8 before ever treating "SAM" as a settled decision -- its identity is explicitly unconfirmed.
- The ranking matrix (
EXECUTION_BACKEND_STACK_RANKING.md) for the scored comparison across shape, stdlib-cleanliness, default safety posture, integration precedent, license risk, and maintenance confidence -- use it to justify any deviation from the primary record's recommendation, don't just restate the primary record's prose.
3. Extending This Record
If asked to add a 13th candidate or revisit an existing one:
- Score it on the same six dimensions the ranking matrix uses.
- Verify license and maintenance status directly against the live upstream repository -- do not carry forward a stale assessment from this record without re-checking, especially for anything flagged low-confidence here (agentmemory, SAM).
- If the candidate is framework-shaped (would compete with CoderOS's own
orchestration/routing layers), default to
reference_onlyunless there is a concrete, stated reason CoderOS should stop owning that layer itself -- there is currently no such reason recorded anywhere in this repo. - If the candidate is tool-shaped (a bounded external agent CoderOS could wrap), compare it against the Aider/OpenCode precedent shape before proposing anything novel.
4. "Exceptional Human Input Only" Triggers
Halt and escalate if:
- Asked to implement any of the PR sequence steps from §10 under this skill -- redirect to a properly-scoped implementation task instead.
- Asked to confirm "SAM"'s identity without the operator having actually named the specific project -- do not guess-and-commit.
- Asked to weaken any of the risk controls in §7 (e.g. "just enable AG2's
human_input_mode=NEVERfor this one run") -- these controls exist because the upstream defaults are genuinely unsafe for CoderOS's model, not as a formality. - Asked to treat a MarketOS need as justification for a CoderOS-core adoption decision -- MarketOS is explicitly out of scope (§5).
04 同仓库更多 Skills
coderos
Workspace-native skill for operating, extending, and safe parallel implementation on CoderOS itself. Use when selecting non-overlapping implementation slices, navigating CoderOS governance, scout, work-order, sandbox, and skill-pack seams, and executing safe changes.
coderos-workspace-operator
Skill for interacting with the local CoderOS Operator Workspace to coordinate tasks and prevent collisions.
coderos-workorder
Workspace-native skill for drafting valid typed CoderOS WorkOrder JSON payloads and exercising dry-run work-order dispatch safely without live execution or runtime mutation.
coderos-worker-adapters
Generates deterministic safe command packets for various workers
coderos-work-order-factory
Draft and validate deterministic CoderOS work-order factory artifacts without dispatch or provider activation.
coderos-source-reuse-auditor
Apply the Source Reuse Acceptance Auditor rubric before opening, and before merging, any PR that copies, adapts, mines, vendors, clean-room-reimplements, sidecar-wraps, or references a public repository -- the acceptance gate that sits above the license/security boundary, not a substitute for it.