feat: add specialist agents example and imageModel routing

[dorukardahan]

What

New examples/specialist-agents/ config with 9 agents: the 5 core agents from full-stack plus 4 domain-specific specialists. Also adds imageModel configuration and documentation for both features.

Specialist Agents

Agent	Primary Model	Role
devops	Codex	Infrastructure, deployment, shell scripts, monitoring
researcher	Gemini Pro	Deep research, fact-checking, long-context analysis
content-writer	Opus	Blog posts, documentation, copywriting
community	Flash	Community management, moderation, quick responses

Core agents vs specialists:

• Core agents are model-optimized — they pick the best model for a task type (code, research, fast)
• Specialists are domain-optimized — they have workspace isolation, custom AGENTS.md/MEMORY.md, and domain-relevant skills

Each specialist gets its own workspace directory so domain context stays separate.

Image Model Routing

Added imageModel to the defaults block:

"imageModel": {
  "primary": "google-gemini-cli/gemini-3-pro-preview",
  "fallbacks": [
    "google-gemini-cli/gemini-3-flash-preview",
    "anthropic/claude-opus-4-6"
  ]
}

Routes vision/image analysis to Gemini Pro (strong multimodal, 1M context), with Flash and Opus as fallbacks.

Files Changed

• New: examples/specialist-agents/openclaw.json — 9-agent config
• New: examples/specialist-agents/gemini-models.json — Gemini provider for specialists
• Updated: SKILL.md — specialist agents section + image model routing section
• Updated: examples/README.md — specialist setup guide with Gemini copy commands
• Updated: README.md — repo structure + subscription table

No private data

All workspace paths use generic ~/.openclaw/workspace-* convention. No credentials, keys, or personal info. Agent names are generic templates (devops, researcher, content-writer, community) that users can rename.

[dorukardahan]

Great expansion — the specialist-agents example is useful and mostly well-scoped. I reviewed this as stacked on top of PR #5 and checked README/SKILL/examples plus new files for consistency, factual clarity, and user confusion risk.

What’s strong:

• New specialist-agents example is generic and reusable (no user-specific identifiers; only ~/.openclaw paths).
• imageModel routing is documented in both SKILL.md and examples/README.md, with concrete defaults and fallbacks.
• README tree and pricing/config matrix now include specialist-agents, matching examples/README.md.
• Public-repo safety check: no keys/tokens/private hostnames/credentials exposed.

Cross-file consistency observations:

• ✅ examples/specialist-agents/openclaw.json includes 9 agents as documented.
• ✅ examples/README.md and README.md both describe specialist-agents and pricing range consistently.
• ✅ imageModel fallback stack in SKILL.md matches the specialist example (Gemini Pro -> Gemini Flash -> Opus).
• ✅ Gemini schema-workaround file (gemini-models.json) added for specialist-agents and referenced in docs.

Potential confusion / recommended follow-ups:

1. README tree indentation ambiguity (small doc bug):

   • In README.md, and appear at similar tree depth, and there is also a trailing line that can read as attached to specialist-agents or root depending on font.

   • Suggest making the tree explicit with a dedicated directory block:

2. Workspace naming inconsistency between docs and example:

   • SKILL.md example uses
   • examples/specialist-agents/openclaw.json uses (good)
   • But readers may expect (agent id) for symmetry.
   • Not wrong, but consider adding one line clarifying naming is arbitrary to avoid copy/paste confusion.

3. imageModel behavior statement may overgeneralize:

   • SKILL.md says agents without imageModel will use their primary text model for vision tasks.
   • If runtime behavior differs by provider/tooling version, this could become stale quickly.
   • Consider softening wording to typically or linking to canonical OpenClaw behavior docs.

4. Optional consistency tweak in specialist fallbacks:

   • Some specialist agents have only 2 fallbacks while core defaults have 3. This is valid, but docs emphasize robust multi-provider fallback strategy.
   • Might be worth adding a short note that specialist configs intentionally trade redundancy for simplicity.

Net: Strong PR, no blocking issues found; mostly minor documentation clarity improvements to reduce user misreads.

[dorukardahan]

Great expansion — the specialist-agents example is useful and mostly well-scoped. I reviewed this as stacked on top of PR #5 and checked README/SKILL/examples plus new files for consistency, factual clarity, and user confusion risk.

What is strong:

• New specialist-agents example is generic and reusable (no user-specific identifiers; only ~/.openclaw paths).
• imageModel routing is documented in both SKILL.md and examples/README.md, with concrete defaults and fallbacks.
• README tree and pricing/config matrix now include specialist-agents, matching examples/README.md.
• Public-repo safety check: no keys/tokens/private hostnames/credentials exposed.

Cross-file consistency observations:

• examples/specialist-agents/openclaw.json includes 9 agents as documented.
• examples/README.md and README.md both describe specialist-agents and pricing range consistently.
• imageModel fallback stack in SKILL.md matches the specialist example (Gemini Pro -> Gemini Flash -> Opus).
• Gemini schema-workaround file (gemini-models.json) added for specialist-agents and referenced in docs.

Potential confusion / recommended follow-ups:

1. README tree indentation ambiguity (small doc bug):
   In README.md, specialist-agents/openclaw.json and full-stack gemini-models.json appear at similar tree depth, and there is also a trailing gemini-models.json line that can read as attached to specialist-agents or root depending on font. Suggest making the tree explicit with a dedicated specialist-agents directory block listing both openclaw.json and gemini-models.json.

2. Workspace naming consistency note:
   SKILL.md examples and specialist config use workspace-research for the researcher agent. This is valid, but readers may expect workspace-researcher by agent-id symmetry. Consider adding one line that workspace naming is arbitrary.

3. imageModel behavior wording may age:
   SKILL.md states agents without imageModel will use their primary text model for vision tasks. If runtime behavior changes by provider/tooling version, this can become stale. Consider softening to typically or linking canonical OpenClaw behavior docs.

4. Optional fallback-policy note:
   Some specialist agents define fewer fallbacks than core defaults. Valid, but docs emphasize robust multi-provider fallback strategy. A short note that specialists trade redundancy for simplicity would reduce ambiguity.

Net: Strong PR, no blocking issues found; mostly minor documentation clarity improvements to reduce user misreads.