akiiwa-games/AgentCompass / skills/agent-compass
agent-compass
UnverifiedMulti-agent task router and subscription quota scheduler. Choose 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/effort-matching.md
View on GitHub# Effort and the decision ledger
The router picks an effort by scoring each effort bundle like any other candidate: the fit of the model's four category scores
(`creativity`, `coding`, `aesthetics`, `tool_operation`) at that effort to the task's demand, minus the quota economy's price for
the spend at that effort. Effort is not a function of task level, and no gate or floor decides it.
## Model side: per-effort scores and Time/Task
Per-effort capability and duration live in `references/five-axis-scores.json` (`calibration/five_axis.py`):
- **Per-effort board readings.** Artificial Analysis boards (`aa.terminalbench_4`, `aa.frontiercode_2`, `aa.mmmu_pro`, and `aa.ttfa` read as `-ln(seconds)`) publish separate rows per effort (`low`, `medium`, `high`, `xhigh`, `max`). Missing efforts on an AA board are filled from the nearest same-series sibling's step on that board, else the fleet median step (`n >= 3`), else held at the nearest listed effort.
- **Single-variant boards.** Boards that list one variant per model (`arena.agent_code`, `arena.agent_alignment`, `lmarena.*`, `designarena.*`) shift across efforts using the model's (or same-series sibling's) AA curve on the same facet when an AA board measures that facet; when none does, the board stays flat across efforts and its status is marked `measured_partial`.
- **High from one-sided effort data.** A measured facet whose own curve lacks High gets it before the bound along effort, never over a measured value: between its nearest measured efforts below and above High by linear interpolation in effort rank (status `interpolated`), or, measured on one side only, the nearest measured effort plus the panel's median own step to High on that facet (`n >= 2`; status `estimated`, method `panel_effort_step`). `measured_efforts` lists the efforts the boards measured.
- **Every effort, and no inversion.** Every facet with a value gets one at every effort of the scale, `none` (reasoning off) to `ultra` (own-curve interpolation, the panel step, the fleet's edge step beyond Low and Max), and every curve is bounded by one rule, `user_scores.bound_along_effort`: High the anchor, a lower effort capped by the next higher one, a higher effort floored by the next lower one, in the facet's direction (rules in `calibration/README.md`).
- **Manual scores.** A manual score (`user-scores.json`) sets High; the facet's measured efforts keep their values, its other efforts follow its own curve moved to the manual High, an effort its curve lacks takes the published panel or edge step, and the curve is then bounded by the same rule with the set efforts (High, and a declared fixed effort) held at the manual value; an effort that bound moves is no longer listed as measured.
- **`default`.** A launch at `default` (and a bundle that names no effort) reads the effort the data declares for it (`five-axis-scores.json` `effort_readings`: High, where the calibration stores a board's effort-less listing).
- **Time and spend.** The hours a task takes at an effort are the Artificial Analysis Time/Task at that effort (`duration_multiplier.by_effort`, as a multiple of `claude-opus-5-5` at High), and the spend is those hours at the model's `usd_per_hour`. A model AA lacks takes its `arena.ai` Agent Time/Task moved onto that scale by a cited fit (`require_skill=True`); a model with no time of its own takes its nearest same-series measured sibling's per-effort time (`series`); missing efforts interpolate or step via the fleet's cited ratios (`basis_by_effort`). A model or effort with no time is out of automatic routing (a named dispatch still runs).
## The match
- Targets come from the panel (the fleet models of `anchor.panel`, counted by `anchor.panel_n`): level 1 (`t1`) is the panel's weakest (and the zero point of `five_axis_fit`), level 2 (`t2`) is the panel median, and level 3 (`t3`) is the panel's best (see `docs/technical-reference.md`).
- A routine task (`demand = 1`) is met by lower efforts (`responsiveness` is also higher at lower efforts when `tool_operation >= 2`) and pays less quota; a demanding task (`demand = 2..3`) climbs to higher efforts when the extra fit (`min(z, t) - t1`) outweighs the quota cost.
- Expiring quota makes the extra spend cheap (the economy prices it near zero under the window's recovery line), so effort rises.
- `decision.task.max_recommended_effort` is the effort of the chosen bundle.
## Ledger snapshot
Every non-dry-run `plan` line in `router-ledger.jsonl` carries `decision_snapshot` (`ledger_snapshot.decision_snapshot`): the task
demand, the scoring weights and the ranking. The best 6 eligible rows keep every component, `expected_pct`, `expected_usd`,
capacity summary and notes; the next 8 eligible and 16 excluded rows are compact. A plan line stays under about 24 KB so the
gate's 1 MiB tail read (and `read_ledger`'s 8 MB tail) still reaches recent routes.
The tray replays that snapshot (`router_data.replay`, `theatre_source == "snapshot"`) and never calls `allocator.plan()` for it.
A ledger row without a snapshot is recomputed against current quota (`theatre_source == "recomputed"`) when its task is still
valid, otherwise shown as unreplayable; the UI says which.
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 |