feat(models): support OMP as a model agent

[daoiqi]

What

Adds OMP (Oh My Pi) as a models target, so teamai models switch / restore / pull reach it like any other agent.

OMP is Pi's fork and keeps the same provider catalog, so the switch, restore and ownership paths are shared with Pi — piProvider, piSnapshot, piKeyOf and sameManaged are reused as-is. Only three things differ:

	Pi	OMP
File	~/.pi/agent/models.json	~/.omp/agent/models.yml (YAML)
Agent-dir override	honors PI_CODING_AGENT_DIR	not honored — default layout only
Env-backed apiKey	$VAR, expanded by Pi	!printenv VAR, run by OMP

Notes on the last two:

• PI_CODING_AGENT_DIR relocates Pi's agent dir, so reading it here would send OMP's catalog into Pi's directory (or the reverse) on a machine that sets it for either tool. This is the position omp-hooks.ts already takes for OMP, and it means OMP's path is constant — no sameAgentSettingsPath case is needed.
• OMP has no inline $VAR expansion, but it runs a !command apiKey through /bin/sh and uses its output. So an environment-backed key is written as !printenv VAR and stays out of the file, the same guarantee the other agents' forms give. A key stored locally with no environment reference (no variable to reference) is still written as-is. When the variable is unset, OMP fails at startup with a named error rather than sending an empty key.

OMP only reads models.yml; it never rewrites it, so TeamAI owns the bytes and the serialization is ours to choose.

Verification

npm run build, npx tsc --noEmit, npm run lint, npx vitest run (8087 passed), and the e2e suite are green. Real-CLI run against the built dist, in a scratch HOME, with the actual omp 18.6.1 binary consuming the file TeamAI wrote:

$ teamai models add tokenhub --name 'Tencent TokenHub' \
    --protocol anthropic,openai-chat-completions \
    --base-url https://api.model.haihub.cn --model deepseek-flash --from-env TEAMAI_E2E_KEY
✔ Added local model profile local:tokenhub.

$ teamai models switch local:tokenhub --agent omp
switched      omp switched to local:tokenhub

$ cat ~/.omp/agent/models.yml
providers:
  local:tokenhub:
    name: Tencent TokenHub
    baseUrl: https://api.model.haihub.cn/v1
    apiKey: "!printenv TEAMAI_E2E_KEY"
    api: openai-completions
    models:
      - id: deepseek-flash

$ omp -p "reply with the single word: ok" --model deepseek-flash
ok

$ omp models
local:tokenhub (1)
│ deepseek-flash │      1M │    384K │ low,high,max │ yes │

$ teamai models restore --agent omp
restored      omp model settings restored

restore returns the file to its prior bytes, including a member's own providers alongside TeamAI's. Also checked by hand: --dry-run writes nothing; an unparseable models.yml fails closed and is left untouched; a provider already owned by the member under the same key is refused rather than overwritten.

Docs

Bilingual, kept in sync: docs/designs/model-profiles.{md,zh-CN.md} (agent-writes table, key-reference sentence, path-pinning sentence), docs/product-overview.{md,zh-CN.md}, docs/usage-guide.{md,zh-CN.md}, and the Oh My Pi models column in all five READMEs.

Note for reviewers

writePi / piDoc gained an agent: 'pi' | 'omp' parameter, so Pi's call sites are touched even though their behavior is unchanged — Pi's four e2e tests and the full unit suite cover it. The rationale for keying on the agent rather than on a file-format capability object: the agent is one fact that determines both path and serialization, so the two cannot be mismatched.

[github-actions]

Findings

• [P1 blocking] src/models/switch.ts:1089 skips collision detection whenever OMP already has managed state. If a member owns local:b, then switches OMP to local:a and later local:b, writePi overwrites their provider; restore subsequently deletes it because only local:a’s original snapshot was captured. Check collisions for a newly targeted key even during re-switches.
• [P2 non-blocking] src/models/switch.ts:292 parses and re-stringifies the entire user-owned YAML document. The first switch permanently removes comments and original formatting, and restore cannot return the file to its prior bytes as claimed. Preserve the YAML document structure/comments or retain sufficient original content.
• [P2 non-blocking] docs/designs/model-profiles.md:82, its Chinese counterpart, both usage guides, and skill-data/core/SKILL.md:198 still say only Codex/OpenCode/CodeBuddy/WorkBuddy lose alias models when switched. OMP now follows that path too; update all affected bilingual docs and the deployed skill as required by AGENTS.md.

Testing

• The PR description documents sufficient testing, including a representative real-CLI run with OMP.