todoerCapture is independent of AI. POST /api/v1/assist/capture saves one task and an exact original-text receipt before returning. It never calls an external provider. The original UTF-8 text is limited to 10 KiB and also becomes the task description. The first nonempty line becomes the display title; a line over 500 characters is shortened with an ellipsis. Text is never silently split into tasks.
Routing is deterministic: an explicit list ID wins, then accessible literal-prefix rules, then the named destination preference and accessible Inbox fallback. Conflicting matching rules produce a conflict requiring an explicit destination. Missing or ambiguous defaults also require a choice. Rules reference stable list IDs; revoked destinations are skipped. Rules can be managed only by the user's browser session or an unscoped admin credential. Their list IDs remain filtered by live write access.
The standalone Assist screen has been removed. /assist and /settings/assist redirect to /settings/ai, which holds capture preferences and AI instructions. Capture and reviewed AI drafts remain available through the regular task flows and the APIs below.
Provider configuration and privacy
Built-in inference requires an enabled account AI switch, a supported enabled model and a valid funding source. Managed credits additionally require administrator approval, sufficient available credit, an operator-configured instance key and a configured rate card. BYOK uses the user's encrypted OpenRouter key. TODOER_AI_ENABLED=0 stops new inference for the instance. Manual capture remains available.
The selected model and OpenRouter routing are visible before dispatch; the serving provider and token/cost metadata are shown afterward when available. Clients cannot supply arbitrary endpoints or unsupported model IDs. Settings and billing controls are browser-only. See AI, BYOK and credits for account controls, operator configuration and payment behavior.
Operators can supply private prompt configuration through TODOER_ASSIST_PROMPT_FILE. The runtime file is kept outside Git and image build contexts. It does not replace code-level authorization or schema validation.
Classification sends the saved capture text and at most 20 currently write-authorized destination IDs and short list titles (80 characters each). It does not send other tasks, comments, workspace descriptions, or account briefs. Drafting sends only the submitted brief. Avoid sending information you do not want processed by an external provider.
Requests use strict JSON schema output, require_parameters: true, data_collection: "deny", allow_fallbacks: false, and bounded provider routing prices derived from operator configuration where a conversion basis is available. Routing caps are independent of the retail quote shown to the user. Data-collection restrictions do not establish a universal zero-retention guarantee; review the selected provider's policies. See the official structured output documentation and provider routing documentation.
Classification has a two-second total HTTP timeout and 128 output-token cap; drafting has eight seconds and 4096 tokens. Responses are capped at 64 KiB. Redirects, application retries, provider fallback, tool execution, and model-selected destinations for drafts are disabled. Unknown destination IDs, extra output fields, malformed JSON, truncated completions, and invalid task trees are rejected. Model confidence is never used as authorization or as a reason to apply a suggestion.
Durable limits
Abuse/safety reservations are committed before dispatch using a conservative request-size and maximum-output bound. These safety reservations are separate from purchased credits and are not a wallet debit. Account, model and provider spend controls also apply; capture never needs a balance.
Managed inference uses an accepted retail quote, an atomic maximum-credit reservation and a final settlement/release. Failed or interrupted operations retain a recoverable accounting state. BYOK has no Todoer credit debit but can still incur OpenRouter charges. See usage and billing for the distinction.
A repeated request key returns its stored result or the current quote/recovery state. It never silently dispatches again after an uncertain provider outcome. A new explicit key requests another attempt. Reusing a key with changed input, model request, task revision or permission context conflicts.
REST contract
All routes use existing REST authentication and live authorization. Browser writes also require the existing CSRF token and same-origin checks. MCP credentials are not automatically REST credentials.
-
GET /api/v1/assist/status: account-aware enablement and behavior flags. -
POST /api/v1/assist/capture:{idempotency_key,text,list_id:null|ID,default_key:null|KEY}→{task,route_source,original_text_preserved:true}. Classification and drafting first return{status:"quote",quote}when inference is available. Review the quote's funding, model, maximum customer debit and expiry, then submit the identical request body and idempotency key withX-Todoer-AI-Quoteset to the returned quote ID. This explicit acceptance is required for dispatch. Quotes do not change task data. The native UI performs this confirmation before calling the provider. -
POST /api/v1/assist/classify:{idempotency_key,task_id,revision}→{status:"suggestion",task_id,revision,destination:null|{id,title},applied:false}, or disabled/unavailable status. Current task revision and every suggested destination are rechecked after the HTTP call. -
POST /api/v1/assist/drafts:{idempotency_key,brief,target}→{status:"draft",preview,suggested_list_title}, or disabled/unavailable.targetis{kind:"existing_list",list_id}or{kind:"new_list",workspace_id,title,description:"",restricted:true}. Restricted creation requires workspace-owner access; explicitly choosing inherited access requires editor access. -
POST /api/v1/assist/drafts/{id}/review:{revision,payload}→ reviewed preview.payloadcontainstargetandtasks:[{title,description,parent_index:null|earlierZeroBasedIndex}]. Review explicitly validates edited content, destination permissions, and the current audience snapshot. Maximum 50 tasks, depth five, 500-character titles, 10-KiB descriptions, and 64-KiB total payload. Dates, schedules, reminders, and priorities are not accepted fields. -
POST /api/v1/assist/apply:{preview_id,revision,expected_hash,idempotency_key}→{list_id,tasks}. A generated preview must first be explicitly reviewed. Review expires after ten minutes. Modified hashes, stale revisions, expired review, changed destination permissions, or an already-applied preview conflict. Only the server-stored reviewed payload executes. Common list/task mutation methods run inside one transaction, including history and idempotency records. A failure rolls back all created tasks and any new list. -
GET /api/v1/assist/rules,PUT /api/v1/assist/ruleswith{prefix,list_id},DELETE /api/v1/assist/rules/{id}: at most 20 literal case-sensitive prefix rules per user, prefixes 1–100 bytes.
Task and preview records preserve the original input for inspection; subsequent edits never rewrite the original capture receipt or draft brief. Apply retries recheck live access before returning their saved result. No conversational or hidden client state determines a destination.
Verification and limits of evidence
Tests use an isolated local HTTP fake provider and synthetic text. They verify real request transport, strict routing flags, output validation, visible-only classification, editable review, atomic rollback, stale/expired/hash conflicts, idempotent retries, and durable budget reservations. They do not evaluate a live model's quality, external retention, availability, or actual billing. No production API key or paid provider call is required for these checks.
Reproduce domain and route checks with cargo test --manifest-path app/Cargo.toml --locked assist --lib. With an isolated development server on port 8092 and no provider key, run npx playwright test app/browser/assist.spec.js --project=chromium --project=firefox --workers=1 to verify retired URLs redirect to AI settings without an Assist navigation entry.