v0.3.8 — Offline Routing Fixes, Floor-Graded Scores & Plugin Tag Updates

AgentCompass v0.3.8 is a fix release on top of v0.3.7.

• Offline routing no longer misreads code work: When the System-One scorer is unavailable, the offline fallback no longer treats device nouns as computer use. "Fix the keyboard shortcuts in editor.py" and "Fix mouse event handler in app.ts" now route as debugging. A named source file or a debugging word also keeps "translation" work out of text translation, so "Fix translation.py", "Debug translation pipeline" and "修复翻译模块的 bug" route as debugging too.
• Raised floors grade from the floor: If you raise an ability floor above the shipped score origin (for example in the tray's ability-floor chart), a model that only just clears it no longer earns a free share of the ability score.
• Plugin updates survive re-pointed upstream tags: The tray's plugin updater takes upstream tags (git fetch --tags --force). A clone that kept a release tag from before upstream rewrote its history no longer fails every update with "would clobber existing tag".

The Windows tray installer is unsigned; Windows SmartScreen may warn on first launch. See packaging/SIGNING.md.

Status: alpha. Customize the reference fleet and model calibration for your own subscriptions.

v0.3.7 — Evidence-Calibrated Scores & Floors, Usage-Aware Simulation & Ability-Floor Chart

AgentCompass v0.3.7 calibrates model scores and ability floors from public benchmarks, adds a usage-aware routing simulation, keeps frontier work on frontier models without silent downgrades, and adds an ability-floor chart to the tray.

• Evidence-calibrated model scores: Model rows without a measurement are scored from LMArena first (WebDev and Agent boards, fitted to the anchor tiers), then Artificial Analysis or task benchmarks; rows that still have no measurement carry score_estimated. Effort curves are recalibrated from Artificial Analysis effort data. A free-usage bonus is declared per account in public data (Dot only) instead of applying to every free model.
• Ability floors from evidence: The routine, hard and coupled floors (2.3 / 3.1 / 3.6) come from SWE-bench Verified success by fix-time bucket, cross-checked against LMArena Agent sessions. The tool-execution route has its own hard floor (2.7). Cheap models get the simple work they handle, and the floors are validated on stranger account sets, not just this machine.
• Frontier work stays on frontier models: The frontier floor is 4.2 (GPT-6 Astra's anchor tier), so each vendor's top subscription model qualifies. Critical frontier work needs 4.3. When nothing at or above the floor has quota, the plan waits or asks the user (next_reset) rather than routing to a lower tier. Models whose harness cannot set an effort declare fixed_effort (Grok Bot scores at Medium).
• Usage-aware routing simulation (route_simulation.py): grid, explain and replay feed usage through the same pipeline roll uses, sweeping usage states (ample, each family scarce, 5h-only, stale), machine vs fresh modes, and stranger account sets (Claude Pro, Codex Plus/Pro, Antigravity, Magpie pool, OpenCode free).
• Pool demand fix: A Magpie pool of N subscriptions forecasts demand against N x 100% and uses the later reset, so a pool no longer looks free when its remaining exceeds one account's allowance.
• Tray ability-floor chart: The strategy page shows every model's ability at each effort against the floors, the route floors and the critical gates. Floors can be dragged or typed, with a live preview of which models gain or lose eligibility. Changes are saved as numbers only to accounts-override.json, and "restore defaults" removes them.
• Release verification F10: The foreign-machine harness verifies all of the above from the installed wheel. Each check is shown to fail against broken data.

The Windows tray installer is unsigned; Windows SmartScreen may warn on first launch. See packaging/SIGNING.md.

Status: alpha. Customize the reference fleet and model calibration for your own subscriptions.

v0.3.6 — Public Reference Data Channel, Unified Quota Gate & Foreign-Machine Fleet Parity

AgentCompass v0.3.6 adds a sha256-verified public reference data channel (schema 2), unifies routing and gate quota decisions on data-declared window scopes, drives setup from router discovery, adds Magpie gateway subscription pool aggregation with live refresh and hover breakdown, and ships a reusable foreign-machine verification harness.

• Public reference data channel (agentcompass data): Installed copies stay current between releases by fetching sha256-verified accounts.json, plan-templates.json, and classifier-cases.json from public-data.json (schema 2) into <state_dir>/public-data/sets/<version>, reported in agentcompass doctor and controlled by AGENTCOMPASS_PUBLIC_DATA=off. Older schema-1 installs ignore schema-2 manifests and preserve their bundled set cleanly.
• Clean-install fleet parity & public-data model checkbox defaults: Dropped the machine-local accounts.local.json split so a clean install routes with the full calibrated reference fleet. Discovered models default to checked unless accounts.json marks them superseded; the user's model-management checkbox overrides both ways, and checked metered models remain a last-resort fallback when no subscription bundle is eligible.
• Unified quota gate, data-driven window scope & running-owner control exemption: router_gate_hook.py now delegates quota refusals to allocator.launch_blockers and window_scope.py (models / models_except declared in data, active in-flight reservations deducted, no private floor tables or claude-2 name guessing). Parent control messages to an owned, currently running Paseo worker skip the quota check so wrap-up and handoff instructions are never blocked by an exhausted window.
• Multi-account classifier cascade & Antigravity plan tier discovery: System-One classification tries every usable signed-in account of each scorer family (~/.claude-2, ~/.codex-2, Paseo custom homes) and records absent credentials in skipped without counting them as errors. Antigravity's subscription tier (free / pro / ultra) is read from Google Cloud Code loadCodeAssist using agy's own sign-in token (AGENTCOMPASS_ANTIGRAVITY_PLAN=off to disable), and third-party windows price from the pool prior on template-only hosts.
• Magpie subscription pools (×N), Tray cards & Usage Float parity: Auto-discovered Magpie gateways refresh hours-old readings on force=True via POST /v1/magpie/quotas/refresh, aggregate all subscriptions of the largest plan (pro ×2, 7d×2, 5h×2), expose N × (100 - mean) remaining capacity to the router and gate, and show masked per-account hover breakdowns in both Tray account cards and Usage Float.
• Unified setup / provider_inventory home discovery: setup/sync_host_globals.py derives harness homes for rules, skills, and hooks directly from provider_inventory.discover() plus FAMILY_SPECS, preserves each secondary Claude/Codex home's own settings.json (model, effortLevel, env), and retires removed homes from setup-manifest.json on upgrade.

The Windows tray installer is unsigned; Windows SmartScreen may warn on first launch. See packaging/SIGNING.md.

Status: alpha. Customize the reference fleet and model calibration for your own subscriptions.

v0.3.5 — Worker Continuation, Calibrated System-One Cascade & Slim Prompt Contracts

AgentCompass v0.3.5 adds automatic worker continuation for follow-up briefs, calibrates the lightweight System-One classifier cascade without mechanical fallback, slims down injected worker prompt contracts across hosts, and extracts the Cindy and Magpie integrations into standalone plugin repositories.

• Worker continuation for --brief and --resume: roll --brief automatically matches and resumes idle, completed Paseo workers owned by the same parent session (by task_id, brief path, step/worker tags, and title tokens) instead of spawning a fresh worker on every follow-up turn. --resume <AGENT_ID> pins a specific idle worker, --new-worker forces a fresh worker, and workers whose historical uncompressed context (compacted_tokens + context_tokens) reaches 2M tokens rotate cleanly to a new worker.
• Slim, idempotent multi-host worker prompts: Internal hook gate rules moved to references/parent-address-gate.md, leaving references/worker-communication.md as a concise multi-host worker contract covering both mailbox hosts (Paseo, Herdr, Orca, Claude Code, AGY) and host-managed channels (Cindy, T3, headless CLI, Dot, Grok-bot). prepare_worker_prompt is idempotent, and resumed Paseo turns (prepare_resume_prompt) send only the new brief plus the parent address footer without repeating the full communication contract.
• Lightweight System-One cascade & no mechanical fallback: Default classifier priority now starts with the lightest models (jev → claude-haiku-5-5 → gpt-6.1-terra → gemini-3.8-flash-low). When all LLM classifiers fail or time out, roll returns status: scorer_unavailable (caller_action: parent_autonomous_choice) so the parent session chooses autonomously rather than routing on a mechanical keyword heuristic.
• Calibrated local_only, difficulty and effort: Added explicit positive and negative local_only examples, tightened reasoning_depth and context_scale rubrics against brief-length inflation, and recalibrated effort_allocation so bounded medium tasks no longer saturate at xhigh.
• Grok usage & tray resilience: Usage collection reads Grok credits from the CLI billing endpoint without touching its login state, and Usage Float tray controls recover cleanly after transient file errors.
• Standalone Cindy & Magpie plugins: Extracted the host-specific plugins into cindy-agentcompass-plugin and magpie-agentcompass-router, documented in README.md, llms.txt, and references/herdr.md.

The Windows tray installer is unsigned; Windows SmartScreen may warn on first launch. See packaging/SIGNING.md.

Status: alpha. Customize the reference fleet and model calibration for your own subscriptions.

v0.3.4 — Multi-Task Briefs Return to the Parent & Batch Roll

AgentCompass v0.3.4 keeps the main session as the only coordinator: a brief that bundles several independent tasks comes back to the parent instead of going to one worker, and several single-task briefs can be rolled in one command.

• The System-One classifier reads the whole brief and also answers multi_task (true/false) and task_count (1, 2, 3, >3). Every scorer supports both fields: agy, Claude, Codex and OpenAI-compatible scorers through the structured profile schema, Jev/TypeSafe through a noul and a choice question, and the local heuristic from numbered task or deliverable headings.
• A multi-task brief returns status: split with the task count and starts no worker. The parent splits it and rolls each task; when the parent judges the work one coupled task, --keep-whole "<reason>" dispatches it whole and records the reason in the ledger.
• roll --brief a.md --brief b.md ... classifies all briefs concurrently, then routes them one after another and returns status: batch with one result per brief. Per-task options (hours, local-only, keep-whole, files, pinned model, wait) stay single-brief.
• The daily audit lists split and kept-whole decisions with their reasons. The Cindy bridge treats split like self and returns the router result.
• Host rule overlays can include the shared rules with @@SHARED@@ instead of copying them, and the shipped rules describe split and batch dispatch.

The score cache version changed, so the first classification of each brief after upgrading runs the scorer again.

The Windows tray installer is unsigned; Windows SmartScreen may warn on first launch. See packaging/SIGNING.md.

Status: alpha. Customize the reference fleet and model calibration for your own subscriptions.

v0.3.3 — Independent Usage & Reliable Host Dispatch

AgentCompass v0.3.3 makes usage collection independent of Usage Float and improves dispatch, recovery and installation through the public agentcompass command. The v0.3.2 attempt was not published: its release checks exposed relative-path and Windows short-path errors in the test harness, corrected here without removing checks.

• The router owns usage collection and caching. Usage Float is optional, and existing usage caches remain readable.
• One public entry handles task text, stdin and brief files, with stage-specific help and complete recovery commands. Documentation and generated host rules use that entry.
• Herdr support adds parent-bound dispatch, durable launch receipts, result collection and cleanup. Windows workers now wait for the actual provider process and preserve its exit status.
• Cindy gains an installable plugin bridge with trusted session context, exact route and permission checks, and durable result collection. It requires the upstream experimental tasks API; compatibility with released Cindy builds is not claimed. An unknown create without a task ID requires inspecting the draft in Cindy rather than resending.
• Unconfirmed Paseo launches return pending with a continuation for the same dispatch. Exhausted short quota windows block dispatch even when weekly quota remains; recovery, account aliases and model-specific limits are respected.
• Setup hooks work with spaces and shell metacharacters in paths. Windows uninstall removes the managed environment variables and scheduled task.
• Decision theatre diagnostics are included and remain opt-in. An unwritable diagnostic log no longer prevents the window from loading.
• Distribution packages exclude development tests. Tests now exercise public commands, actual child processes, HTTP and shell boundaries, and installation and GUI operations in Windows Sandbox.

Validation included fresh-home dispatch, lost-response recovery, quota exhaustion and recovery, Windows installation/uninstallation, and a real Quota Monitor refresh click. External provider and host APIs used fixtures where noted; paid model calls and a released Cindy installation were not tested. Release publishing requires the installed-wheel sandbox checks to pass.

The Windows tray installer is unsigned; Windows SmartScreen may warn on first launch. See packaging/SIGNING.md.

Status: alpha. Customize the reference fleet and model calibration for your own subscriptions.

v0.3.1 — Decision Compass Replay & Fast-mode Contracts

AgentCompass v0.3.1 adds an animated Decision Compass: the Windows tray now replays each routing decision from task to winner, and the router gains verified fast-mode launch contracts.

• Decision Compass decision theatre: an animated replay of task → work kind → candidate routes → ranking → top-two capability radar → winner with its effort. Ledger entries keep decision snapshots, so past decisions replay faithfully.
• New radar rendering with unified capability axes and four editable model ratings, plus a verdict strip that summarises each decision in one line.
• Model, bundle and work-kind names are short and consistent across the page (theatre, verdict strip, recent decisions, candidates, radar pickers, status line and Quota Monitor). Raw ids stay in tooltips.
• Fast-mode launch contracts for Codex and Claude: fast keeps model and effort, normal launches pass the off value explicitly, and Claude fast mode needs an explicit paid opt-in.
• Effort routing is reset-aware: observed and expected reset horizons feed dynamic urgency, and task depth is matched to model-specific ability and burn curves.
• Host-specific defaults are gone: the Grok2API integration and its port default are retired, and the reference catalogue and fixtures are realigned with the neutral fleet.
• The README hero is now a single animated Decision Compass replay rendered from the real window with synthetic data.

The Windows tray installer is unsigned; Windows SmartScreen may warn on first launch. See packaging/SIGNING.md.

Status: alpha. Customize the reference fleet and model calibration for your own subscriptions.

v0.3.0 — Installable Distribution & First-run Onboarding

AgentCompass v0.3.0 ships an installable router CLI and a Windows tray installer, with first-run guidance and safer host setup.

• Install the release wheel with uv tool install <wheel-path-or-URL> and use the agentcompass CLI: setup, uninstall, doctor, providers, and plan / roll.
• Windows tray installer: per-user Inno Setup installation, with a bundled Python runtime and helper processes. This build is unsigned; Windows SmartScreen may warn on first launch. See packaging/SIGNING.md.
• First-run onboarding explains missing prerequisites and next steps. Host setup uses managed blocks and backups, supports uninstall and legacy-rule migration, and leaves hooks opt-in.
• A neutral synthetic default fleet replaces personal account assumptions. Keep your subscriptions and preferences in local accounts.local.json; the public model catalogue is separate from per-user policy.
• Router decisions report missing usage data honestly. Accounts, model families and quota pools derive from discovery and configuration; account preferences remain boolean state rather than generated reasons.
• Paseo completion notifications and explicit worker final reports return delegated results to the parent. Routing gates let parents steer their own running workers while preserving launch and notification checks.
• Tray account cards, priorities and usage follow discovered model families. The Decision Compass rendering, window identity and icons are improved, and frozen helper launch paths no longer assume a host Python installation.
• Dot plugin controls use stable attributes before translated labels, and remote work resolves the actual remote root without the old tunnel sender.
• Wheel, source distribution and tray payload checks exclude untracked repository files and private host state.

Status: alpha. Customize the reference fleet and model calibration for your own subscriptions.

v0.2.0 — Explicit Quota Pools, Discovered Models & Tray Redesign

AgentCompass v0.2.0 makes quota pools explicit, lets discovered models join routing, and redesigns the tray's quota and model pages.

• Explicit quota pools: each bundle declares which pools it draws from (draws), and the score is now value minus opportunity cost through one exchange rate (scoring.opportunity_exchange).
• Harness factors (scoring.harness_factors) scale a model's value on non-native harnesses, so Claude served through Antigravity is selectable (factor 0.9).
• Discovered models become bundles at runtime (auto:<account>:<harness>:<model>). Unknown models get a conservative family or generic default score. Subscription pools are on by default, metered ones off.
• 5-hour limit fixes: usage above 100% no longer reads as unknown, Claude session-limit 429s are written back and respected, and in-flight work reserves pool capacity.
• Account-level directory restriction cwd_regex (compatible with allowed_cwd).
• Tray: two-row adaptive quota cards with in-flight usage shown on the pool bars; "Model settings" in three tabs with harness+model rows, a model dropdown (select all / invert), and CLI update plus daily auto-update on the home tab.
• Grok Bot automatic routing, with a Grok CLI fallback when Paseo is unavailable.
• Paseo plugins moved out of this repository; the Paseo dispatch bridge has its own dependencies, prepared during first-time setup. Host executables are discovered portably.

Status: alpha. Reference accounts, model profiles and capability values are the author's own calibration. Customize them for your subscriptions.

Validation: router suite 192 passed (1 skipped); tray suite 408 passed.

v0.1.0 — Deterministic Task Routing & Subscription Quota Scheduling

AgentCompass v0.1.0 is the first public source release of the deterministic task router and subscription quota scheduler.

• Route each task by model capability, effort, task fit, and 5-hour / 7-day quota economics.
• Try a fully offline reference-fleet demo immediately after cloning: python -X utf8 skills/route-agent-work/scripts/route_judge.py plan --input examples/task.json --dry-run --demo.
• Inspect the ranked candidate table or use --format json for the complete breakdown. Demo data is synthetic and cannot launch a worker or write the routing ledger.
• Evaluate your signed-in providers by removing --demo.
• MIT license, logical model alias guidance, related projects, and a 1280 x 640 social preview card.

Status: alpha. Reference accounts and model profiles are examples calibrated for the author's fleet. Customize them for your own subscriptions. The portable Python router is the core; the Windows cockpit and companion bridges are optional.

Validation: router, setup and tray unittest suites were run from the source-only export; empty HOME/USERPROFILE demo succeeded; Gitleaks reported zero findings. Optional private host-handover tooling tests skip when absent.