gemini-smart-search（Gemini smart 搜索）

Gemini Smart Search

Use this skill when Gemini should be the search backend, but gateway-level web_search config is too static or too disruptive to change.

Purpose

This skill is a script-backed search workflow, not a gateway tool override.

It exists to provide:

• dynamic Gemini model selection
• quota-aware fallback
• a separate Gemini API key path if desired
• structured JSON output
• no gateway restart requirement for model changes

Modes

Model routing is split into two layers:

• display chain: human-facing preferred model family labels
• candidate API ids: the actual model ids to probe, especially for 3.x preview-era models

Current display chains:

• cheap

- Prefer gemini-2.5-flash-lite - Then gemini-3.1-flash-lite - Then gemini-2.5-flash

• balanced

- Prefer gemini-2.5-flash - Then gemini-3-flash - Then gemini-2.5-flash-lite

• deep

- Prefer gemini-3-flash - Then gemini-2.5-flash - Then gemini-3.1-flash-lite

For 3.x models, do not assume the UI label is the raw API id. Probe candidate ids such as preview-suffixed names when needed.

Invocation

Run the Python script or the shell wrapper via exec and request JSON output.

Python is now the canonical entrypoint because it also loads repo-local .env.local when present. The shell wrapper remains a convenience layer.

Primary example (preferred):

python3 skills/gemini-smart-search/scripts/gemini_smart_search.py \
  --query "BoundaryML context engineering" \
  --mode cheap \
  --json

Wrapper example (convenience only):

bash skills/gemini-smart-search/scripts/gemini_smart_search.sh \
  --query "BoundaryML context engineering" \
  --mode cheap \
  --json

python -m gemini_smart_search may work when run from the scripts/ directory, but it is not a supported interface for agents right now. Do not depend on it.

Output contract

Expect JSON with at least:

• ok
• query
• mode
• model_used
• fallback_chain
• display_chain
• answer
• citations
• error
• escalation

Notes:

• model_used is the actual probed API model id (for example gemini-3-flash-preview), not the human-facing display label.
• Citation URLs may initially be Google/Vertex grounding redirect URLs instead of canonical source URLs; treat that as a known current limitation.
• With --json, supported runtime paths should return structured JSON on both success and error. Invalid CLI arguments now also return JSON when --json is present.

API key policy

The script should prefer a dedicated key path for this skill, then fall back to the standard Gemini key.

Required key resolution order:

1. SMART_SEARCH_GEMINI_API_KEY (primary declared env)
2. GEMINI_API_KEY (compatibility fallback)

If neither key is present, the agent must explicitly ask the human for a Gemini API key before claiming setup is complete.

Do not store the key in tracked repository files. Prefer a repo-local, gitignored file such as .env.local.

See references/config.md.

When to use this skill instead of built-in web_search

Use this skill when:

• you want Gemini-only search
• you want to test or isolate quota pools
• you want model routing without touching gateway config
• you want predictable JSON output for downstream orchestration

Do not use this skill when:

• a normal built-in web_search is sufficient
• you need non-Gemini providers
• you only need to fetch and read a known URL (web_fetch)
• you need logged-in or JS-heavy page interaction (browser)

Fallback policy

Fallback only for errors like:

• quota exceeded / 429
• model unavailable
• transient upstream failure

Do not silently fallback on obvious local/script bugs or invalid arguments.

References

• references/config.md — environment variables and design notes
• references/qa-test-plan.md — focused QA scope for v1 behavior and release gates
• references/qa-results-2026-03-12.md — CLI-oriented QA outcomes from the current release cycle
• references/agent-qa-cases.md — adversarial agent-style misuse review
• references/model-id-recon.md — verified callable Gemini model IDs and mapping notes
• references/escalation-design.md — when to return a GitHub issue URL for human escalation
• references/release-checklist.md — artifact release checklist with current completion status
• references/development-goals-v0.1.1.md — next small version scope and artifact policy
• references/release-notes-v0.1.1.md — release notes for the current artifact
• assets/example-output.json — expected response shape
• scripts/smoke_test.sh — non-destructive local smoke checks for the scaffold
• scripts/prepare_artifact.sh — deterministic clean artifact export helper

Status

Python implementation is now wired for a first real version:

• direct Gemini API call path
• Google Search grounding enabled
• mode-based model routing
• Python-side repo-local .env.local loading
• fallback across Gemini Flash-Lite / Flash variants for retryable upstream errors
• structured JSON output for orchestration

This is still intentionally minimal: it does not yet expose advanced tuning flags, caching, or richer citation post-processing.

Before publishing an artifact, consult references/release-checklist.md, review references/development-goals-v0.1.1.md, run scripts/prepare_artifact.sh, and publish the release note in references/release-notes-v0.1.1.md alongside the artifact.