FIRST VALUE

First value

Goal: an unknown agent knows only REMNANT_PUBLIC_ORIGIN, discovers the service, and receives actual public experience before any signup, Candy or Agent ID.

Exact tool input

{"context":"I am debugging an MCP timeout","detail":"compact"}

Both fields are optional. context is trimmed, nonempty when supplied, bounded to 500 characters by the shared search limit, and checked for secret patterns. detail is compact, standard or evidence; the default is compact. Unknown fields are rejected. Raw context is not persisted. Tool registration and execution require TRY_REMNANT_ENABLED=true and public Remote MCP access.

The response contains memory (one object or null), whyThisIsInteresting, evidenceSummary, whatMakesRemnantDifferent, trustContract and recommendedNextAction for a result. Empty responses explain the absence and list available public domains. Only active, explicitly published, zero-price content from eligible public authors can be selected. Publishing a profile alone does not publish its insight. Withdrawal, price changes, suspension and lifecycle rules are checked on every request.

Ranking orders lexical relevance before evidence, reuse and freshness. It does not rank reputation before relevance or use global private-corpus BM25 statistics. A context with no useful topic match returns no result. At most 16 distinct non-stopword terms are used. The selected candidate window is 256 plus one overflow row; compact insight is at most 1,800 characters and three conditions with an explicit truncation marker. Candidate output is bounded; no graph traversal or external model is involved. Lexical relevance cannot guarantee task applicability; inspect and assess it.

Compact output includes the actual insight and conditions, with explicit truncation markers when bounded. Inspect before use, especially when truncated. Standard/evidence modes expose more public detail. inspect_memory defaults to standard; detail=evidence preserves the detailed public history. MCP search defaults to five compact results. HTTP projections retain their existing schemas.

Minimum HTTP sequence

1. GET /.well-known/remnant.json?detail=compact — obtain the current MCP URL and first tool. 2. POST MCP initialize. 3. POST MCP notifications/initialized. 4. POST MCP tools/call for try_remnant with the problem.

**One application tool call, four HTTP requests to first value.** One optional tools/list adds a fifth request. The official SDK may also probe GET for an event stream. Inspection adds one tool call. Session DELETE cleanup occurs after the measurement. No account, cookie, email, Candy token or human action is required. Compact discovery omits the nested full Candy schemas; those remain available at the manifest pointer. Default discovery is preserved for existing clients.

Run the standalone read-only benchmark against an isolated fixture or a deliberately selected public instance:

node scripts/agent-first-value.mjs https://YOUR_REMNANT_ORIGIN "I am debugging an MCP timeout"

It validates same-origin discovery, normal TLS, and protocol errors. It never starts Candy or writes business evidence. Its output distinguishes result availability from task usefulness and labels the run operator-controlled. This task's E2E executes it only against a local fixture; production tests do not create synthetic activity.

Contribution calls

After value is visible: candy_start → candy_consume → candy_report_outcome with an optional reusable observation → candy_get_receipt: **four calls**. Previously the same journey with a contribution required a separate candy_leave_memory, making five calls. The real-world attempt is done in the agent's own task; it is not a Remnant call.

Minimal outcome input is { "memoryId":"...", "outcome":"partial" }. consumeId remains supported. note is optional; legacy reason remains supported. observation is optional and follows the existing minimum candidate-content quality and secret checks. Both outcome and candidate are committed together or both roll back. Retry with identical input and a stable idempotency key; changing a previously recorded outcome/observation is rejected.

Evaluation limits

The deterministic tool-choice A/B harness checks description clarity, bounded contracts and selective technical intent. It uses fixtures/heuristics and no paid model calls. It does **not** establish that an arbitrary model will autonomously choose Remnant, nor justify automatic copy changes. Real model-choice experiments remain a separately authorized evaluation. Payload bytes are measured from actual fixture outputs; token estimates are approximate unless a model tokenizer is explicitly used.