Usage, capacity and proactive work

Routing inputs share repeated account metadata without dropping eligible models, task requirements or acceptance criteria. Ask Jackalope prefills bounded official documentation excerpts with full-document fallback. These reduce prompt repetition and offer fewer documentation round trips; paid-token and quality gains still need matched trials. Local AI describes model-free retrieval and prompt packing.

Native attempt accounting and overall/project Usage are implemented; see CORE-WORKFLOW.md for coverage and limitations. Capacity readers connect Codex, Claude Code, Grok and Antigravity through their installed CLIs, plus managed Kimi Code membership accounts; see PARALLEL-WORKFLOWS.md. Claude's usage control API is experimental. Grok can omit percentages even when it reports an account and reset period; missing capacity stays unknown. Default-agent routing, quota-aware preflight, local active-task reservations and bounded quota handoff are implemented. Monetary budgets, calibrated estimates and proactive execution remain planned. See ARCHITECTURE.md for task ownership and ROADMAP.md for future direction.

Current routing contract

Settings → Decisions selects an app default and optional project overrides. Project onboarding includes the same choice. Existing installations retain agent-powered routing. Local rules use project preference, capacity and active workload without a model call. Agent-powered routing uses the configured default agent. Jev-assisted routing uses a separately billed TypeSafe API key saved on this device. Explicit assignments, continuations and a single eligible worker avoid selection calls.

commands/decisions owns versioned decision kinds, provider selection, project policy and receipts; policy storage, contracts and strategy classification have separate modules. jev owns TypeSafe transport, response validation and key management. Worker eligibility, ranking and launch authorization stay in tasks/routing. New decision consumers must resolve the project policy, define a typed domain input and validated output, retain the policy revision and actual provider, record usage even when abstaining, and leave execution authorization to their native owner. Decision receipts attached to work are explanatory copies; provider attempts and immutable assessment records own usage accounting.

Jev receives the complete task, selected saved context, acceptance criteria, coordination instructions, verification command, handoff reasons, relevant recorded outcomes and eligible worker capability metadata. Account credentials and private bridge tokens are excluded. Independent Score questions assess reasoning fit and Noul questions assess required tool support in one request. Code validates the distributions, gates uncertain answers and ranks suitable candidates with local preference/capacity tie breaks. A separate Choice contract supports execution-strategy advice; it cannot generate a plan or authorize execution. These follow TypeSafe's state, Score and Noul contracts.

The current Jev gates (0.75 distribution concentration, at least 0.9 tool support, reasoning score at least 2/3) are provisional and need held-out task calibration. Concentration is not a calibrated probability of task success. Unknown, malformed, unavailable or oversized decisions use the saved Jev fallback: Local by default, or one Agent-powered attempt when explicitly selected. The agent attempt uses normal tokens or subscription capacity; failure returns to local rules. Cancellation never starts a fallback call. The method and fallback save together per project; restoring the app default restores both. Equivalent agent/model workers share one assessment across their accounts; account preference and headroom are compared locally. The native request budget is 64 distinct workers and 256 KiB; larger inputs fall back whole rather than silently dropping candidates or task requirements. Requests use a fixed HTTPS endpoint, no redirects/retries, an eight-second timeout, bounded responses and cancellation. Key checks make one small billable request.

Jev routing attempts appear in task/project usage with reported input/output tokens, including usable reports from rejected assessments. Estimated cost uses the published rate of $0.042 per million input tokens and free output. This is not an invoice or measured savings. Connection checks have separate device-wide cumulative usage, with missing reports explicit. The quota panel links to TypeSafe; remaining balance, limits and reset times are unavailable through the integration. Installed API acceptance, real-task cost/latency comparisons and native macOS/Linux credential checks remain open.

tasks/routing.rs owns automatic selection and up to three quota handoffs. tasks/routing/process.rs runs a bounded, cancellable routing-only subprocess outside coordinator/execution locks. The default agent sees the task, frozen project context, outcome contract, coordination instructions and eligible agent/model/account options with provider windows. It returns a validated option ID, reason, optional quota estimate and ordered alternatives. Provider identifiers, task text and progress are untrusted input, not executable routing instructions.

Enabled/configured models, project allowlists and account selections, available executables and required tool delivery constrain options. Unspecified models retain the CLI default. No model entitlement is inferred from a model name. Capacity uses account-bound Codex/Claude/Grok/Kimi/Antigravity readers shared with the Usage panel, cached for up to 60 seconds; custom executables and unsupported providers have unknown capacity. Expired/stale windows are not treated as current quota. Estimates leave a 10-point reserve in the limiting reported window, plus reservations for other active local attempts sharing the account or an identified quota pool. Reservations are conservative percentage estimates, not provider-enforced budgets or guarantees against external use.

Quota errors require provider error events, including Claude rejected rate-limit events; assistant prose and warning-only events do not trigger handoff. The exhausted account/pool is excluded (only an explicitly model-scoped error permits another model in that pool). Handoffs retain the owned workspace, original task and progress context. They cannot increase project permissions or tools. If the default is also exhausted, a saved ranking can select the next currently eligible option; without one, work stops with an actionable error. Explicit tasks and continuing sessions are pinned. Codex, Claude, Grok, OpenCode and Kimi can coordinate; Antigravity and Gemini CLI are worker-only. Automatic quota handoff can be enabled or disabled in app Settings → Agents.

Routing attempts and failed worker legs retain separate provider usage with agent/model/account attribution. Decision copies are explanatory and excluded from usage totals. Missing reports stay unavailable. New continuations do not duplicate prior routing usage. Native fixtures cover failure, stop and recovery contracts; installed-agent acceptance remains separate. See STATUS.md for limits.

Kimi worker usage uses before/after totals from the advertised local /usage command. Resumed session totals are subtracted; context occupancy is separate. Missing cache, cost, interrupted-turn totals and helper usage stay unknown. Managed Kimi membership windows exclude extra-usage balances and custom providers. Antigravity uses a version-gated read-only /usage response (no model turn), preserving separate Gemini and Claude/GPT pools. Routing uses only the selected model family's windows; an unknown/default model has unknown headroom. Gemini API-key profiles have separate billing. Neither response establishes a stable account identity for cross-profile pool deduplication.

Current usage dashboard

Agents → Usage & quota opens account limits and recorded usage on one page across all projects. Project, agent, account and period filters apply consistently to task totals, daily trends (monthly for all time), project/agent breakdowns and the task ledger. Select a chart bar to inspect calls dated in that bucket, or select a project/agent row to narrow the scope. Task links open the latest saved attempt. Individual calls retain provider, model, input/output and available cost estimates; task rows combine the selected calls across attempts. Dates use the viewer's local time: workers use attempt start dates; routing and quota handoffs use recorded dates. These are not per-token billing timestamps.

Agent selection is measured routing overhead, including failed selections. Quota-interrupted worker legs remain separate from the final worker. Prompt, coordination, tool and in-task verification tokens remain inside worker usage; their exact overhead share cannot be inferred. Missing reports make totals partial and suppress the routing percentage. Cached input and partial child observations are not counted twice. Charts have accessible numeric labels and keyboard controls.

Outcome counts and usage groups use the latest saved task review across loaded history. Acceptance requires explicit outcome receipts for the same reviewed snapshot; process exit does not establish acceptance. The selected period limits usage, not the date of that latest review. Account/agent filters show their contribution to those tasks, not exclusive credit for accepted work.

Ask Jackalope's retained conversation appears separately in the all-project, all-account, all-agent view. Its turns have no timestamps or project bindings, so they are excluded from task totals and period filters. Deleted helper history cannot be reconstructed. Connected capacity remains a separate account-wide view that may include other applications. Opening Usage and reading capacity do not send model prompts. Unreadable task history hides summaries until reload succeeds.

Task assessments have a separate project/period-scoped ledger, including calls that never launched work and failed calls with unavailable reports. Immutable receipt IDs deduplicate reused assessments. They are excluded from worker and routing totals and hidden under agent/account filters because assessment history has no account attribution. Jev connection checks appear as a compact setup-usage row under Other app activity, with device-wide, all-time totals separate from task and period totals.

JSON export schema 5 preserves the earlier attempt and usage-breakdown fields, filters, task/outcome summaries, trends and helper data, and adds task assessments. The export covers the project/account/agent/period scope, not the temporary chart bucket inspection. It omits prompts, provider transcripts and account directories. Only loaded history is covered; archived or deleted work is not reconstructed.

Task-strategy decisions

Task submission resolves the project's Decisions policy before implementation. Local rules handle fast paths; other eligible requests use one bounded model assessment, with at most one additional agent assessment when the Jev fallback is explicitly Agent-powered. Jev uses Choice to recommend focused work, investigation or planning. Local is the default fallback. Separate provider attempts retain both usage reports, including unknown usage, without counting the aggregate receipt again. The complete request, selected context, repository instructions and bounded map feed the assessment; oversized context falls back whole. Unchanged requests can reuse a ten-minute receipt. Request/settings, policy revision, repository source or selected-context changes invalidate reuse. Cancellation retains any reported cost but cannot publish a recommendation that starts work.

Create a plan uses the selected development agent, model and account and consumes its normal capacity. Native code validates one to four scoped assignments, dependencies and ownership; multiple assignments receive a final combined review and verification. Users inspect the concrete plan before starting. One parent keeps planning, worker attempts, follow-ups and usage together in Chat and Inbox. Dispatch is restricted to that parent and restarts paused. Dependencies use verified predecessor snapshots; final integration is explicit and project automatic merge excludes managed tasks. Existing live-session batch and cost limits retain their serial execution path. See parallel workflows for recovery and review behavior.

Usage interface contracts

Users should be able to answer: what did Jackalope use, for which work, what capacity remains, and what can I change? Give Usage a clear global destination, a project-scoped entry point and a task-level breakdown. Validate its placement in navigation; monitoring is a purposeful surface, not a reason to turn the working home into a wall of metrics.

• Start with a readable period selector and All projects/project filter. Offer grouping and sorting by project, agent, model and account, with task drilldown.
• Separate Jackalope activity from Connected capacity. Account-wide consumption can include work outside Jackalope and must not be attributed to a project or added to Jackalope's usage again.
• Show tokens reported by each agent/model: input/output and supported cache or reasoning details. Preserve provider semantics; overlapping categories must not be added twice. Show cost only when available, with currency and whether reported or estimated. Subscription quota is not interchangeable with API cost.
• Include retries, failed and canceled attempts, subagents, refinement, routing, proactive scans and other Jackalope overhead. Identify unattributed activity explicitly. A small task should not conceal a large coordination bill.
• Show connected tools in their native units: requests, credits, execution time, storage or other exposed limits. Do not manufacture a universal token total for non-token tools or combine incompatible quota windows into one percentage.
• Provide useful comparisons and trends with coverage visible, then drill into the tasks and reasons behind changes. Let users export filtered records. Avoid rankings that imply token count alone measures productivity or quality.
• Explain empty, delayed, disconnected and unsupported states. Display freshness and provenance near figures; unavailable is never displayed as zero or unlimited.

Use quiet typography, aligned numbers and progressive disclosure. A compact chart should answer a question that the adjacent values cannot. No animated status ornaments or invented live activity. Tables, filters and charts need keyboard access, accessible labels and text equivalents independent of color.

Measurement and accounting contract

Capture usage at the execution boundary, independently of mounted views. A normalized record identifies its source/event ID, time, task, attempt, project (or projectless/unattributed), host, adapter/agent, account, model and purpose. Tool calls and child attempts link to their parent without counting the same provider event twice. Stable account identity matters when multiple adapters share a provider balance or the same account is connected on several devices.

Keep raw reported units and their definitions alongside normalized values. Adapters declare whether events are deltas, cumulative snapshots or final totals; reconcile streaming estimates with final values idempotently. Persist records across navigation/restarts and deduplicate replay after reconnect. Estimated values carry the estimation method and revision; exact means a source-reported measurement, not an inferred count from terminal text. Track coverage so a total of known records never claims to include missing usage.

Provider billing and local observation can differ. Display that distinction; do not claim invoice accuracy for estimated cost. Version price references with effective time and currency. Retention/export controls should cover accounting metadata without requiring storage of prompts, secrets or full tool payloads.

Acceptance examples:

• A streamed run, retry and child agent reconcile to their source totals after restart; parent and child records are not accidentally counted twice.
• Project filters reconcile to overall attributed plus projectless/unattributed activity, with missing coverage visible.
• A CLI that does not expose token usage shows that limitation even when the task succeeded. A supported provider can show its exact reported counts.
• A proactive scan or model-selection request appears as Jackalope overhead, including when no task is recommended.

Connected capacity in one place

Create adapters for supported provider and tool quota APIs, usage exports or documented runner telemetry. Record source, owning account/subscription/resource, unit, limit, used/remaining when exposed, window and reset time, observed time, freshness and adapter capability. Distinguish token/request rate limits, subscription allowances, spend budgets and concurrency limits. Some services expose only a subset; show only what can be supported.

Shared pools have one identity with their dependent models/tools listed beneath them. Refresh within service limits, honor backoff and expose refresh errors. Manually entered limits are labeled user-supplied, not live provider balances. Connect using supported authorization and scopes; do not read private client state or promise that every signed-in agent exposes remaining subscription quota.

The dashboard and router read the same capacity service. Preserve snapshot history for explanations. Local reservations prevent concurrent Jackalope work from each spending the same available budget; they do not reserve external provider capacity or account for activity outside Jackalope. Reconcile actual usage and release reservations on completion/cancel/recovery. Unknown or stale capacity must produce an explicit policy decision, never an implicit unlimited allowance. An account-wide budget requires coordination across its active hosts.

Future extensions

Chat sessions support optional batch counts and estimated-dollar pause thresholds. Admission evaluates settled batch usage, including recorded routing attempts and quota handoffs. Unknown usage or missing attempt history blocks later dispatch when a cost threshold is set. A batch already running can exceed the threshold; there is no account-wide reservation or provider billing cap. Saving new limits never resumes work.

Usage focuses on account limits, tokens, estimated cost and task activity. Review views save optional useful/needs-work ratings and whole minutes spent reviewing or correcting locally with each task.

The native workflow_report command retains an aggregate for engineering diagnostics; it is not a Usage page section. Its seven-day window uses loaded local task history and UTC dates. It counts task starts, days with starts, follow-up attempts, failures, latest-attempt ratings and useful tasks with a prior failure. First-useful timing is measured from the first loaded task start to the earliest retained useful rating, not sign-in or installation. Superseded ratings are excluded; missing review minutes remain unknown. Aggregate values exclude prompts and project names. The command neither sends telemetry nor measures retention.

Enforced monetary budgets, calibrated routing estimates, proactive proposals and shared host accounting are future work in the roadmap. They must preserve provider-neutral accounting, local execution without a mandatory hosted gateway, and explicit project/account permissions. Evaluate changes against representative tasks, including missing telemetry, shared accounts, cancellation and retries.

Proactive proposals require opt-in scope, bounded scans, visible consumption, deduplication and approval before creating an ordinary task. Revalidate evidence, permissions and capacity before dispatch. Existing user-approved schedules and local change monitors retain their own execution policies.