[akiiwa-games]

Moves packages/agentcompass from 0.1.4 to 0.2.0: the scoring engine is now AgentCompass v0.6.0's, ported module by module in the source repository (akiiwa-games/magpie-agentcompass-router; src/ here is byte-identical to its commit ac494f3).

Contains the continuity change of #791. The package also holds the 0.1.5/0.1.6 work (session hold on any pooled runtime, a replayed state has no age, classifier-only profiles, account states, routing contract), because the source already includes it. This PR and #791 overlap in continuity.js and middleware.js; please decide which to review. Merge either one, or this one and close #791.

What changes for users

• Routing picks <model>:<effort> by a task profile's four capability categories (divergence, coding, aesthetics, tool_operation, demand 0..3), each effort's time and price, and session continuity. The quota economy gates on 5h windows and prices the longer windows.
• Ten new modules under src/: axis-scores, candidate-eval, candidate-ranking, fleet, plans, quota-economy, quota-pools, reference-data, task-intake, user-scores.
• New option userScores: your own category scores for a model.

Behaviour changes you will notice

• Profile shape. axis_demand, access, wants_fast and read_only_review are required. context_scale, hallucination_risk, error_cost, divergence and realtime_loop are gone. A profile in the old shape is refused (invalid_caller_profile); there is no compatibility read.
• Removed options and exports. realtime_loop (option, request field, session hold). From catalog.js: DEFAULT_MODELS, DEFAULT_SCORING_POLICY, KIND_TRAITS, CANONICAL_ROUTES, EFFORT_CURVES, resolveModelSpec. From index.js: computeWindowCost, demandForecast, evaluateModelQuota, CANONICAL_ROUTES, healthOnlyQuotas. From continuity.js: effectiveRealtimeLoop. New: ROUTES. The capability floor and the tier-based raise in the session hold are removed.
• Data gates. A row with a demanded category without data, a billed account without a price, unknown image input, no Time/Task at that effort, or a window of unknown capacity is out of automatic routing. Naming the model still dispatches it, with the data error in warnings. Family and unknown names (claude-opus-9) have no score of their own, so only naming them dispatches them. An account whose plan is not reported has unknown capacity: put plan in quotas.
• Bare names in a pool. If the provider has a dispatch-effort table and the model is not in it, the row has no efforts. antigravity and gemini no longer borrow another provider's table.
• A quotas file broken for over 6 hours excludes only pool/quotas, with a warning.
• Not ported (Magpie only routes): launching workers, harness order, fast variants.

How it was tested

• HOME=<empty dir> bun test packages/agentcompass: 140 pass, 0 fail (Windows, bun 1.4.4).
• bun scripts/check.mjs agentcompass: registry.json passes.
• The package tests run on their own, with no AgentCompass checkout and no network. They assert rules (data gates, exclusions, quota monotonicity, order invariance, entry points routing a profile alike, the no-text-classification scan, the documented contract), not a chosen winner.
• In the source repository, on the same src/:
  • the comparison with AgentCompass's Python allocator.plan at v0.6.0 (a 58,016-row grid of profiles and pools; the 19,339 sampled rows match on model, effort and score to 1e-9);
  • the Linux sandbox run in a real gateway, Magpie 0.1.4022 (magpie-cli-linux-amd64, sha256 46e96145…128dd0, matching the release's SHA256SUMS): CLI install and options, market listing and GUI API, Mode A, Mode B through /v1/systemone, classifier unavailable, session continuity and the turn lock, 7 pool shapes × 4 profiles × 4 protocols (112 requests, every route equal to the engine's), the four categories, pins, caller profiles, and no-profile requests. All passed; the record is docs/linux-sandbox-e2e-report-0.1.4022.json there.
• Not run: the scenario against real upstream accounts (no sign-in was used in the sandbox).
• Files left out of the package: the parity test and its fixtures, the generator script, the pin file, the sandbox scripts, docs.

registry.json: only the agentcompass summary changed (same text as the source's registry entry).

中文摘要

agentcompass 从 0.1.4 升到 0.2.0：评分引擎换成 AgentCompass v0.6.0（四类能力 divergence / coding / aesthetics / tool_operation，各档位耗时与价格，配额经济，数据门禁）。本 PR 包含 #791 的会话连续性改动（0.1.5/0.1.6），两者在 continuity.js 与 middleware.js 上重叠，请维护者决定审哪一个。升级影响：画像字段变了（旧形状会被拒绝）、删除 realtime_loop 及若干导出、缺数据的行不进自动路由（点名仍可派发）、未上报 plan 的账号容量未知。包内测试独立运行（140 通过）；同一份 src/ 在源仓库里与 Python 引擎对拍一致，并在真实 Magpie 0.1.4022 的 Linux 沙箱里跑通了全部场景，只有真实上游账号那一个场景没有跑。

🤖 Generated with Claude Code