akiiwa-games/AgentCompass / skills/agent-compass
agent-compass
FlaggedChoose owner, model and effort, outbound authorization, advisor use, and continuation for delegated work. Reads local usage and effective account policy, profiles the task with the configured System-One scorer, then allocates a compatible worker. Use for independent worker or advisor work; ordinary local execution needs no routing.
references/gemini-decisions-api.md
View on GitHub# Gemini Decision 1 (Decisions API)
Google's Gemini Decisions API answers closed questions (yes/no, one option, a rating) about a shared input with
calibrated probabilities and no generated text. Reference: <https://aistudio.google.com/docs/decisions-api>
(Gemini Developer API, `v1beta`, model `gemini-decision-1`).
Like [TypeSafe Jev](jev-api.md), it may be selected as the
**System-One task scorer** (`scorer_type: gemini_decision`): the router then sends
`POST {base_url}/v1beta/decisions` (default `https://generativelanguage.googleapis.com`, model `gemini-decision-1`)
with the key in the `x-goog-api-key` header. The request carries:
- `input`: the whole task text and the dimension rubric, one JSON object serialized as text (the API takes a JSON
object only as text). Nothing is cut or summarized.
- `questions`: the same 10-question set Jev gets. `choice` questions for `work_kind`, `secondary_kind` (`none` or a
second distinct atomic kind), the effort band, `access` and `task_count` (`1`, `2`, `3`, `>3`); `score` questions for
`primary_level` (`1..3`) and `secondary_level` (`0..3`); `noul` questions for `local_only`, `needs_visual` and
`wants_fast`. `task_intake.validate_profile` projects `axis_demand` (`creativity`, `coding`, `aesthetics`,
`tool_operation`) and derives `needs_gui`, `read_only_review` and `multi_task` deterministically. The model never sees
a question key, so every instruction is a whole question.
The answers become the task profile the same way as Jev's: a `choice` gives the option, a `score` the
probability-weighted level rounded to a level, a `noul` is true from 0.5. The answer names no model, so the scorer
metadata reports the configured one. The per-answer `confidence` is not used: no threshold of our own turns it into a
profile field.
Key, cost and when it runs:
- The key comes from `~/.agent-compass/scorer-config.json` (`gemini_api_key`), else `GEMINI_API_KEY`; it never lives
in the repository, and it is never sent to another scorer's endpoint.
- The API bills input tokens (no output tokens). It runs only when it is the chosen scorer type: a key alone,
in the file or the environment, never adds it to the cascade.
- Without a key it is skipped. A failed call (HTTP 400 covers a limit, a blocked input and a rejected key) moves on to
the next scorer in the cascade `claude_haiku_low`, `codex_terra_low`, `agy_flash_low` (only the configured scorer
is tried when `fallback_to_agy` is false). When none answers, `score_task` returns no profile and
`source: scorer_unavailable`; no keyword-based or default profile is made up.
- It answers task profiles only. Other classifier questions (free-form schemas) go to the next scorer in the cascade.
Configure it with `agentcompass scorer --action set --type gemini_decision --api-key -` (the key is read from stdin)
or in AgentCompass under Settings > System options, and check it with `agentcompass scorer --action test`.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis. | Stars: 472k | Updated | Claude CodeCodex | Verified | Install |
| diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness. | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependencies | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is available | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |