@coinrithm/mcp-trading 0.7.16 → 0.7.17

package/CHANGELOG.md

@@ -1,811 +1,885 @@
-# Changelog
-
-All notable changes to `@coinrithm/mcp-trading` are documented here. The package
-ships two binaries — `coinrithm-mcp` (the MCP server) and `coinrithm-agent` (the
-self-host agent runner) — versioned together. The CoinRithm **API contract** is
-versioned separately (see `openapi.yaml` `info.version`, currently `1.7.0`).
-
-## 0.7.16 (unpublished)
-
-- Preserve independently timed funding-by-venue context in watchlist and
-  discovered entries, including regular and compact prompts. Keep raw rates,
-  interval normalization and unknown source/next-funding times distinct from
-  settlement funding.
-- Bench reports retain up to two safe route attempts with response source,
-  failure class, actions-string diagnostic and admission categories. Legacy
-  missing evidence remains unknown; error text, latency, prompts and raw model
-  output are excluded. Record explicit truncation for nonstandard providers.
-- Carry open interest, positioning, sampled liquidations, macro proxies,
-  chain TVL, stablecoin supply and sampled order-book depth through the runner
-  with their separate freshness, coverage and provenance limits.
-- Keep news source URLs/categories and every related PM coin in context; add
-  the opt-in per-coin-horizon PM cap. Objective guidance stays within the
-  authored strategy and hard caps. Use the supported MCP SDK 1.31 dependency.
-
-- Bench: the PM baselines now preserve the recorded agent's PM board policy
-  (for house agents: depth 30, churn removed), the original opportunity set.
-  Before, they asked for the uncurated 12-row page a live mechanical agent
-  reads, and a cassette recorded without it (house Mia) reported the market
-  as missing. Originally mechanical recordings keep their uncurated board;
-  live mechanical agents are unchanged. New recordings need only the original
-  pass and mark PM baseline readiness after successful PM reads, rather than
-  assuming that an enabled PM venue guarantees recorded evidence.
-- Bench: two more PM baselines, `baseline:base-rate` (uninformative 50%) and
-  `baseline:random` (seeded 20-80%), next to `baseline:skip` and
-  `baseline:market`. They replay the same recorded market pass, so existing
-  corpora need no re-recording; futures/spot-only corpora still get skip only.
-- Add `coinrithm-agent label --corpus <dir> [--horizon-hours 24] [--overwrite]`:
-  writes `labels/<cassetteId>.json` price bars from candles published after
-  each cassette's asOf (reads only). It waits for the horizon to elapse, uses
-  the finest bar size the cassette's age still allows, keeps existing PM and
-  funding labels, and writes PM settlement only from the public event
-  verdict (settlement-eligible + "settle" + a provider won/lost outcome
-  result; anything else stays unlabelled). Only
-  bars that close by the horizon are stored, and an existing file for
-  another horizon is reported (`horizon_mismatch`), never kept silently.
-- Bench futures scoring walks only bars that open after asOf and close by
-  the horizon end; a horizon off the bar grid stops at the last full bar
-  before it (`horizonFlooredToBar`), so no price after the horizon is used.
-- Add optional per-agent `signals` thresholds (`rsiOversold`, `rsiOverbought`,
-  `strongMovePct`, `leanMovePct`, `minStrength`) and a `data.indicatorRange`
-  data diet (`1D` 5m bars, `1W` 15m, `1M` 1h, `3M` 4h). They change the setup
-  shortlist AND the gate that decides when the model is called, so personality
-  changes what an agent notices. Absent blocks keep the previous constants and
-  the 1D range exactly. Both are validated fail-closed, strict-linted and
-  `$ref`-able; the prompt names the agent's bar size and RSI levels.
-
-- Every watch entry now carries perpetual open interest when the API serves it
-  (`openInterest`: single-side USD over Bybit and OKX, 1h and 24h change,
-  `asOf`, `stale`), read from the existing `/api/agent/market` call. The prompt
-  explains how to read it with price; it is context, never a trade rule. Older
-  APIs without the block leave entries unchanged.
-
-- Add an agent test bench: `coinrithm-agent record` saves the read inputs of
-  dry-run cycles as cassettes (writes refused, no model call), and
-  `coinrithm-agent bench` replays them through the production runner for
-  several agent variants and repeats, next to skip and market-implied
-  baselines. The deterministic report gives reject codes, action overlap,
-  repeat consistency, seeded paired bootstrap CIs, a calibrated A/A null
-  rate, a chronological holdout and optional label scoring. Quotes are
-  synthesized and counted; see "Bench" in docs/agent-runner.md for the limits.
-
-- Add an optional `universe` block (market boundaries): a market-cap rank
-  band, a minimum 24h volume, stablecoins in or out, curated sectors and chain
-  ecosystems in or out, and the sort that defines an opportunity. Each cycle
-  the runner asks `GET /api/agent/universe` for rows inside those boundaries,
-  resolves the top `resolveTop` (max 10) into full `discovered` watch entries
-  and passes the rest as context. Rows are re-checked client-side, so a server
-  can narrow the set but never widen it; the blocklist still wins and every
-  cap applies. A failed screener degrades to the watchlist only. The block
-  can live in its own file via `$ref`. Without it, `universe_scan` keeps the
-  previous top-gainers scan exactly. Needs backend-v2 with the screener.
-
-- Add three optional, runner-enforced prediction-market policy fields, so
-  numbers that strategies stated only in prose become executable. Absent keeps
-  the previous behaviour exactly; malformed values fail closed.
-  - `risk.pmMinEdgeGapPct` (0..100): the forecast must beat the fee-inclusive
-    cost by that share of the room left to 100 (16 means cost 50 needs 58). It
-    never lowers the global minimum edge, and with it set a `pm_open` without
-    `forecastProbability` is rejected (`pm_forecast_required`,
-    `pm_edge_below_gap_rule`). Mechanical benchmarks are exempt.
-  - `risk.pmMaxEdgePoints` (0..100): overconfidence guard; the forecast may
-    beat the fee-inclusive cost by at most that many points, and a forecast
-    becomes mandatory (`pm_edge_overconfident`). This is an optional action
-    constraint; its effect on future performance requires separate evaluation.
-  - `risk.pmMaxOpenPerEvent` (1..50): open bets per event (source + slug),
-    counting held positions and opens accepted this cycle (`pm_event_cap`).
-  - `risk.pmMaxOpenPerCoinHorizon` (1..50): open bets on the same directly
-    linked coin whose events end within 24 h of each other, across different
-    events, counting held positions and opens accepted this cycle
-    (`pm_coin_horizon_cap`; a value outside 1..50 fails closed). Every linked
-    coin is kept. Reads the additive discover `relatedCoins` and
-    `/positions/pm` `relatedCoins` / `eventEndDate`; a missing coin, or an end
-    that is missing, past or over 5 years out, never blocks. A count cap, not
-    a correlation model, over the open positions the feed returns (at most
-    200, newest first).
-  - `risk.pmMinMinutesToClose` (0..10080): reject a market whose known close is
-    nearer than that to the later of `observation.asOf` and the validation
-    clock after the model call (`pm_closes_too_soon`). An unknown or
-    unparseable close never blocks.
-
-- Keep repeated provider 404/410 availability failures out of the generic model
-  failure kill-switch. BYO and self-hosted agents retain their configured model
-  and provider holds without entering a disable/revive loop. Transient failures
-  and invalid decisions still count toward the existing kill-switch.
-- Keep rejected model text out of retained parser errors: JSON syntax errors
-  omit input excerpts, and schema errors omit rejected enum values and unknown
-  field names. Useful field/type diagnostics remain; validation is unchanged.
-- Reject OpenAI-compatible and Anthropic responses that explicitly end at the
-  output token limit, and Anthropic responses stopped at the context-window
-  limit, even if their partial decision text happens to be valid JSON. No action
-  from that incomplete response reaches execution. Preserve reported token usage
-  and the hosted router's existing bounded malformed-response fallback. A
-  truncated direct-provider retry also retains its malformed classification.
-- Classify rejected actions strings with bounded, fixed diagnostic categories.
-  JSON5 is used only for diagnostics; string-valued actions remain rejected.
-  No raw model output is retained by this diagnostic path.
-
-## 0.7.15 (published 2026-10-02)
-
-The verified release includes source changes through 1 October 2026. Its public
-archive and MCP Registry delivery are recorded in the repository publishing log.
-
-- PM board rows carry event settlement context from discovery's optional
-  `resolution`. Where `scope` is `per_outcome`, each row also carries its own
-  provider-bound `outcomeRule`, preserving exact primary/secondary terms or
-  explicit unknown status. Event-level text does not substitute for another
-  outcome's rule. Missing fields on older APIs remain unknown.
-- Futures agents see `futuresCapacity` in the user prompt: open vs maximum
-  positions, slots left and margin headroom, computed exactly as the runner
-  validates. With no slot or no margin left, the prompt says plainly that no
-  futures_open (new or add) can pass. Production showed agents at 3 of 3
-  positions proposing a new open on most cycles (max_positions 222 and
-  open_margin_exceeds_cap 116 rejections in 6h), each a wasted model call.
-  The validator still enforces both caps.
-- The PM system prompt no longer claims a "genuine information edge" on crypto
-  price markets or makes a PM bet REQUIRED. It says the market already prices
-  the same public data, that short-dated price markets are usually efficient,
-  and that a bet needs a probability gap larger than costs plus a named reason;
-  skipping PM is always valid. The ref anti-hallucination rule is unchanged.
-- Every non-mechanical agent (not only `objective.primary: calibration`) gets
-  the micro-contract churn filter and the deeper 30-row discovery page,
-  including the Bitcoin fallback. Mechanical baselines keep the unmodified
-  12-row universe.
-- PM board rows carry the discover event's optional `referenceProbability` as
-  event-level `consensus` (`prob` 0..1, `venues`, `spreadPts`, `kind`,
-  `outcome`), never mapped to or flipped for a specific outcome. Absent or
-  unusable stays unknown.
-- Non-mechanical PM agents see their own settled forecast record
-  (`pmCalibration` from `GET /api/agent/performance`) when it has at least 20
-  settled forecasts. It is read at most once per 30 minutes per credential,
-  without retries, and omitted on any failure.
-
-- Preflight new futures entries using informational reference eligibility;
-  validate the available position and margin capacity across actions in the
-  same cycle. Adds/exits retain their separate rules, and the API rechecks
-  execution-time eligibility. Preserve reported capacity and spot fill
-  receipts in decision evidence without claiming full historical replay.
-- Describe forecasts as market-aware, not price-blinded. A capacity-limited
-  agent may still exit or adjust protection. Withheld-action diagnostics are
-  attributed to the provider response that produced them.
-- Add `get_news` for scored crypto news, bringing the source MCP surface to
-  41 tools. Arena reads work without a key; discovery defaults to open
-  markets. Default event results remain compact, with `detail: "full"` for
-  the complete event response. Preserve event subtitles and per-outcome
-  price basis in bounded output.
-
-- The hosted HTTP server recovers the pasted-markdown descriptor link
-  (`/mcp` followed by `](`, literal or percent-encoded): GET/HEAD redirect to
-  the service descriptor and POST runs the same MCP handler as `POST /mcp`.
-  Every other unknown path stays 404 and `GET /mcp` stays 405.
-- `pm_data_canonical` documents the canonical `consensus` (or null) and the
-  labelled `consensusHistory`, and names orientation `flipped` as the API does.
-
-- Preserve optional observed-price flags and bounded source provenance in public
-  market summaries, including the exact sides of venue comparisons. A missing
-  observed price remains distinct from an observed zero and unknown provenance.
-  This source change is not included in published npm `0.7.14`.
-
-- Hosted scheduler rollout supports the opt-in `whale_context` runner capability. After the deterministic
-  gate fires for a PM-scoped, non-mechanical cycle, it attaches bounded public
-  tape context and up to two validated wallet movement summaries to the
-  observation. Reads are best effort with explicit availability status and do
-  not widen candidates or change risk/write gates. The observation fingerprint
-  includes this context; the existing partial persisted audit projection does
-  not store raw whale or wallet samples. Published MCP `0.7.14` remains unchanged.
-
-- Preserve full public `walletAddress` values in whale tape and default event
-  summaries for direct wallet-tool follow-up on Polymarket, Limitless and Myriad.
-  Anonymous/null and absent identities remain unchanged. This source fix is not
-  included in the published npm `0.7.14` package.
-- Record explicit zero model calls, tokens, inference cost and action writes for
-  cycles that return before inference due to a kill switch, observed drawdown,
-  failed account reads or an empty market selection. Historical records remain
-  unchanged; failures after a possible provider call retain unknown evidence.
-
-## 0.7.14
-
-- Scope permanent model-error streaks to the attempted provider/model. Discard
-  legacy unattributed streaks and reset availability failures on a successful
-  provider response, including malformed decisions. Old-route failures cannot
-  request an early hold on a replacement model; agent risk limits are unchanged.
-- Attribute a routed permanent-error hold to the attempt that produced the error,
-  even when a later fallback is rate-limited; keep actual-call metering unchanged.
-- Preflight futures protection updates against observed side, mark and liquidation
-  prices, including retained triggers. Reject known invalid end-states with an
-  actionable reason before a write; the API remains authoritative as prices move.
-- Add the optional `risk.pmMinEntryProbabilityPct` policy (0..100 points): the
-  runner rejects `pm_open` when the chosen outcome's raw market probability at
-  entry is below the floor, with a reason that names both numbers. Fees stay
-  in the forecast-edge check, so a 19-point outcome fails a 20-point floor even
-  when fees lift its cost above 20. Absent keeps today's behaviour; a set floor
-  with no quoted probability fails closed. The floor is rendered in the prompt's
-  hard caps.
-- Retain candle timestamps and compact coverage/interval evidence alongside
-  indicators, separately from market-price freshness. Missing timestamps stay
-  unknown; nominal five-minute cadence does not imply current or regular bars.
-  Model input and private decision evidence retain the same context.
-- Load optional `character/entries.md`, `exits.md`, `sizing.md` and `research.md`
-  from local bundles so Studio strategy sections reach the model, manifest and
-  compiled definition. Existing bundles without these files are unchanged.
-- Include the exact compiled strategy definition and its digest in local
-  inspection; expose the same snapshot builder to engine consumers.
-- Add `run --expect-definition` to reject a changed baseline before model or
-  account access. The digest is not a full replay record or broker adapter.
-- Hosted validation accepts `limits.maxDailyLossMusd: 0` as no daily loss cap,
-  matching the runner and the hosted API; negative and non-finite values are
-  still rejected.
-- Tactic cap merges treat 0 as unlimited for `maxTradesPerDay` and
-  `maxDailyLossMusd`: a tactic may tighten 0 to a positive cap but can no
-  longer turn a positive cap into 0. Writes and margins keep lower-is-tighter.
-
-Publication and hosted deployment are verified separately. See the repository's
-[release status](https://github.com/CoinRithm/coinrithm-agent-trading#version-clarity)
-for registry availability and the hosted runtime evidence.
-
-## 0.7.13
-
-- Preserve optional per-outcome venue terms in compact prediction-market output,
-  including explicit unavailable values and valid false/zero facts.
-- Include candle volume coverage in agent fundamentals: positive counts identify
-  missing expected venues; unknown coverage stays unknown. Preserve measured zero
-  volume and keep each volume paired with its own coverage.
-- Classify provider capacity failures consistently, including HTTP 503
-  resource exhaustion, in runner results and configured same-model retries.
-- Export `classifyProviderFailure` from the supported engine entry point.
-  Retry counts and trading behavior are unchanged.
-- Surface signed cumulative futures funding and its applied-through timestamp to
-  the agent observation so funding already reflected in balances is not counted twice.
-- Add public whale-wallet context to the 40-tool surface: bounded 7-day or
-  30-day wallet activity and source/wallet movement detail, with provenance and
-  provider-reported context kept distinct from CoinRithm's observed trade flow.
-- Clarify `pm_data_calibration` as market-price calibration: the primary lane
-  uses one complete-book snapshot in the inclusive 20–28 hour pre-resolution
-  window, with event-weighted ECE and separate `finalPrice`/`ownCapture` lanes.
-
-Publication and hosted deployment are verified separately. See the repository's
-[release status](https://github.com/CoinRithm/coinrithm-agent-trading#version-clarity)
-for registry availability and the hosted runtime evidence.
-
-## 0.7.12 — 2026-09-15
-
-Clarify `whoami`, `cancel_spot_order` and `report_pm_opportunity` descriptions,
-removing execution-cost prose unrelated to these operations. Document actual
-authentication, side effects, result fields and retry behavior.
-
-Spot cancellation now advertises its existing idempotent behavior. Opportunity
-reporting no longer advertises unconditional idempotency: duplicate protection
-requires `decisionId` (or `agentTrace.decisionId`) under the same API key; the
-first stored record wins. Reporting remains a write despite requiring only the
-`read` scope, and its evidence remains explicitly self-reported.
-
-Tool names, accepted inputs and execution behavior are unchanged. Published on
-npm and deployed to hosted MCP on 2026-09-15; registry and GitHub downloads match
-the CI archive. A clean registry installation passed startup and tool metadata
-checks. See the [release record](https://github.com/CoinRithm/coinrithm-agent-trading/releases/tag/mcp-trading-v0.7.12).
-Glama's evaluation score is separate from package delivery.
-
-## 0.7.11 — 2026-09-15
-
-Fix opportunity reporting that previously treated resolved API failures as
-successful submissions. Only `ok: true` confirms a report. HTTP errors remain
-unconfirmed; transport failures and exceptions have an unknown delivery outcome.
-
-`CycleResult.opportunity` now contains confirmed reports only. The additive
-`opportunityReport` field preserves the attempted payload, outcome and status,
-without API error bodies or exception details. Existing consumers should use
-this field when they need attempted rather than confirmed evidence.
-
-The reporter retains its one-invocation-per-cycle latch and adds no retries.
-Focused regressions compare successful and failed reporting in skip/act cycles
-and verify unchanged trading results and runner state. SDK versions are unchanged.
-This source entry does not establish registry publication or hosted deployment.
-
-## 0.7.10 — 2026-09-15
-
-Source changes following the published 0.7.9 release. The TypeScript SDK remains
-0.3.1 and Python remains 1.8.1; their runtime source is unchanged.
-
-- Persist file-backed run identity before execution so a first-cycle process
-  crash cannot discard the idempotency identity. Transport uncertainty is still
-  not automatically replayed or recorded as a completed trade.
-- Fix runner startup on Node 18: use the imported Node crypto API instead of
-  depending on a global crypto object. The installed-package matrix reproduced
-  this failure on Linux, Windows and macOS.
-- Add the supported `@coinrithm/mcp-trading/engine` entry point, preserving
-  existing deep imports, and separate observation accounting and opportunity
-  reporting from cycle ordering.
-- Add opt-in, machine-checked crypto return predicates and visible warnings for
-  inactive/reserved configuration. Existing agents are not automatically opted in.
-- Serialize scheduler migrations in a bounded transaction; add an offline
-  credential-rotation helper and interruption/recovery rehearsal.
-- Pin workflow actions and verify release-tool checksums. Add installed-package
-  compatibility and restart smoke checks across operating systems and runtimes.
-
-  0.7.10 was published and deployed on 2026-09-15. Registry and GitHub downloads
-  matched the CI-tested archive; hosted MCP and scheduler deployments finished.
-  See the [release record](https://github.com/CoinRithm/coinrithm-agent-trading/releases/tag/mcp-trading-v0.7.10).
-
-## 0.7.9 - 2026-09-15
-
-Published on npm and verified on 2026-09-15: the registry archive matches the
-reviewed artifact and passes fresh-install checks. See the
-[combined release notes](https://github.com/CoinRithm/coinrithm-agent-trading/releases/tag/mcp-trading-v0.7.9)
-for source provenance, verification and community acknowledgments.
-
-Public market-data fidelity and runner reliability release. No MCP tool was
-renamed or removed, and the API **contract stays 1.7.0**.
-
-**PM evaluation budget.** Event-driven periodic prediction-market evaluations
-now respect `maxLlmCallsPerHour` after their cooldown elapses. Budget skips make
-no provider call and consume no call allowance. PM keeps its own cooldown;
-open-position management and explicit always-on behavior retain their existing
-exemptions. This runner gate is separate from hosted provider-capacity admission.
-
-**Retry-After parsing.** Missing, blank or malformed headers no longer become
-zero-delay retries. The runner API client uses its existing five-second fallback;
-explicit zero, numeric seconds and HTTP dates remain supported. Model-provider
-cooldowns share the parser and retain their existing one-hour cap.
-
-**API request deadlines.** Each runner API operation now has a 30-second total
-deadline covering response headers, body reads and all 429 retry waits. The
-same client serves the hosted scheduler. Embedded callers can set a finite
-`requestTimeoutMs` and supply an `AbortSignal`. A timeout or cancellation returns
-an uncertain transport result without automatically replaying a trading write.
-Timers and listeners are removed when the operation finishes.
-
-**State persistence.** Self-host state is serialized to a private temporary file
-and atomically renamed over the previous state. A failed serialization or rename
-leaves the prior state intact. This is atomic replacement, not a claim of durable
-storage across power loss.
-
-**Agent conversion.** `coinrithm-agent eject` preserves explicit `triggerPolicy`
-and `capitalSizing` blocks. Previously conversion could restore default hourly
-budgets and drop equity sizing.
-
-**Release verification.** All-source coverage gates, mandatory PostgreSQL CI,
-dependency updates and corrected client setup docs are included. See the
-[reliability record](https://github.com/CoinRithm/coinrithm-agent-trading/blob/main/docs/RELIABILITY.md).
-
-**Confirmed-action journal.** Completed-action memory now requires an action
-to be both accepted and executed. Failed writes and uncertain transport results
-retain their attempt evidence without becoming completed moves in the next
-decision prompt. Dry-run proposals remain unexecuted.
-
-**Direct NVIDIA retry.** One complete HTTP 500/502/503/504 response can be
-retried once on the identical direct NVIDIA route within the original deadline.
-Both attempts are retained. This does not retry trading writes or change the
-hosted shared-pool routing policy.
-
-**Capital reconciliation.** Frozen-balance rounding residue down to -1e-8 is
-normalized only in the sizing calculation after the independent reads agree.
-Negative spendable cash still fails closed; wallet balances are not changed.
-
-**Private decision input evidence.** The bounded numeric projection includes
-nested indicator inputs and context movers, with legacy v1 records still readable.
-It does not retain hidden reasoning or raw model output.
-
-**Compact prediction-market evidence.** Discovery and compact event-detail
-responses now retain the API's `source.quoteScale`, `source.methodology` and
-`source.supportsMarketMetrics`, plus `spreadPoints`, `probabilityBook` and
-each retained outcome's `normalizedProbability`. Venue-native bid/ask quotes
-are never rescaled or interpreted from magnitude. Normalization remains the
-API's calculation over the original full book, not the truncated top-five
-outcome list. Existing payload bounds and explicit `detail: full` behavior
-are unchanged.
-
-**Settlement-time provenance.** Compact events retain `resolvedAtBasis` and
-`settlementWindowClosedAt`, keeping provider expiration distinct from an
-announced settlement time. Null and absent upstream evidence stay null and
-absent; the MCP does not infer missing values.
-
-**Candle semantics.** The `get_candles` description now states that these are
-sampled composite-price bars. Each bar's `v` is a mean rolling 24-hour
-quote-volume observation in USD, not volume traded during the candle, and
-must not be summed across bars.
-
-**HTTP completion diagnostics.** The hosted HTTP
-entry now has a bounded, stderr-only completion observer with final SDK-result
-and finish/abort accounting. Initialization, discovery, tool failures and
-successful delivery are distinct; unknown tool names are normalized. Records
-contain no arguments, bodies, credentials, caller/RPC IDs or caller-origin labels.
-Credential presence is not authentication. Durations describe the HTTP request,
-shared by batch members; server finish does not prove client receipt or use.
-Stdio, tools, authentication and dependency versions are unchanged. See
-`DEPLOY.md` for the measurement and retention limits. Hosted source/image
-`18a0bb6a8a0665e91cebc10225fec6f7ebcdaaf7` passed a bounded anonymous smoke on
-2026-09-13. Hosted verification and npm publication are separate release steps.
-
-**Deployment boundaries.** Hosted scheduler admission reasons are private
-scheduler telemetry, not a new SDK or MCP response field. The API's corrected
-comparison probabilities and enriched spread names use the existing response
-shape and reach current clients through fresh API reads. Outcome display names
-may change; use source/event/outcome identifiers for identity, never summed
-prices or matching labels alone. These fixes do not establish trading returns.
-
-## 0.7.8
-
-Runner decision-quality, evidence and paper-capital release. Additive: no MCP
-tool was renamed or removed, and the API **contract stays 1.7.0**. This release
-contains all package changes since published 0.7.7 (`gitHead` `80d0cae`), not
-just the previously listed thesis work.
-
-**Thesis exits.** Every opening action (`futures_open`, `spot_order`,
-`pm_open`) now carries a `thesis`: a one-sentence summary plus an
-`invalidation` with at least one machine-checkable condition (`priceBelow` /
-`priceAbove` for coins, `probabilityBelow` / `probabilityAbove` for prediction
-markets, a `maxHoldMinutes` time stop, and a free-text `catalyst` the model
-re-judges itself). The runner binds the thesis to the position the server
-returns, sanitized side-aware (a rising price never invalidates a long; a
-wrong-side level is dropped rather than re-signed; the time stop is clamped to
-60 minutes .. 30 days), persists it in the run state (`RunState.theses`, the
-same state file / `agent_state` JSON as before, no schema change) and
-re-evaluates it every cycle. A futures position whose price level or time stop
-is breached is closed by the runner before the model is asked anything, logged
-as a `thesis_invalidated` exit with its own idempotency key, after the
-kill-switch and drawdown checks and never instead of them. Prediction-market
-positions have no close endpoint, so a broken PM thesis is surfaced to the
-model instead (do not add, let it settle). The parser is tolerant (a malformed
-thesis never fails the open; a thesis copied onto a close is ignored) and the
-structured-output schema requires it, so schema-enforced hosted models always
-emit one.
-
-**Fundamentals in the observation.** Each watch entry now carries
-`fundamentals` sourced only from calls the runner already makes: `categories`,
-`marketCapRank` and `marketCapUsd` from the market context; `volume24hUsd` from
-the candles the `indicators` capability already fetches (live-probed
-2026-09-02: each bar's `v` is a rolling 24h volume, so the latest bar is the
-24h figure, never the sum); and up to three `headlines` with `publishedAt`
-timestamps from the one `news` call, attributed through the curated coin-news
-graph. Discovered PM markets carry `endDate` and `liquidityUsd`; open PM
-positions carry their title, side, entry and current probability and
-`openedAt`; open futures positions carry `openedAt`. The system prompt states
-the thesis contract, the runner-enforced exit and how to grade a trade on the
-fundamentals. Not carried, because no agent endpoint serves them: an "about"
-text per coin, a 24h probability change and a cross-venue divergence per PM
-market.
-
-**Fix:** the public movers feed serializes `change24h` / `currentPrice` as
-decimal strings; the universe-scan context rows read them strictly as numbers
-and shipped `undefined` for every mover.
-
-**Opt-in equity-based paper sizing.** A runner can size entries from a
-conservative fraction of its independently attributed paper book instead of a
-fixed stake/margin. The book is accepted only when wallet identity, cash
-partitions, held-position attribution and spot-mark coverage reconcile. Quotes
-then enforce per-entry, per-symbol, deployed-capital and daily-entry limits;
-fee buffers and the API's fee-inclusive quote evidence are included. Any
-missing or inconsistent evidence fails closed. Legacy positions on a different
-book remain visible for management but never inflate the current book's buying
-power.
-
-**Prediction-market decisions use executable economics.** PM opens now reject
-an invalid raw probability and a model forecast that does not clear the quoted
-entry price. Forecast edge is measured against the actual fee/slippage-adjusted
-fill, not the headline market probability. Quote-expiry outcomes are recorded
-separately from risk/balance rejection, and futures risk/reward validation uses
-fee-inclusive entry and stop economics.
-
-**Decision evidence is structured and bounded.** Cycles can expose a sanitized,
-partial private decision-input record: configuration and observation
-fingerprints, daily budget and guard state, plus bounded observation rows with
-explicit omission counts. It is not a prompt, transcript, raw model output or
-hidden reasoning record. The runner also reports quote/validation evidence for
-abstained, forecast-only and quote-expired PM opportunities. Hosted persistence
-and retention remain the caller's responsibility.
-
-**Runtime controls are more faithful.** The model sees the remaining daily
-entry/add budget rather than only static maxima. Entry caps still block new
-risk, while closes and other risk-reducing actions remain available. Direct
-provider HTTP 429 responses are capacity skips rather than model failures, so
-BYO agents do not build a failure streak during ordinary quota pressure.
-Structured-tool decisions remain required where the provider supports that
-contract.
-
-**Scorecard fix.** Maximum drawdown now measures decline from starting equity,
-so an immediate loss is no longer hidden by treating the first post-trade point
-as the high-water mark.
-
-## 0.7.7
-
-Reliability release. Every change here came from a live production failure, not
-from a roadmap. Additive: no tool renamed or removed, and the API **contract
-stays 1.7.0** because nothing on the documented surface changed.
-
-**Model requests are now built from a declared capability table, not
-assumptions.** `providerCapabilities.ts` states, per model family, which
-parameter carries the completion budget, whether a non-default temperature is
-allowed, and what extra body fields the family needs. Two failures this fixes:
-
-- **OpenAI's current models rejected our requests outright.** `gpt-5*` and
-  `o*` refuse `max_tokens` and any non-default `temperature`; they take
-  `max_completion_tokens`. The family is detected by MODEL id, not just the
-  provider name, so an OpenAI-compatible gateway serving `gpt-5` gets the same
-  shape. If you brought your own OpenAI key, this is why it now works.
-- **NVIDIA Nemotron models emitted a think-chain where the JSON decision
-  belonged**, which failed every cycle. The `chat_template_kwargs.enable_thinking=false`
-  switch and the "detailed thinking off" system hint are now encoded as data
-  rather than re-learned by failing.
-
-**New: `probeDecisionContract()`.** An HTTP 200 is not proof a route can run an
-agent. Both production failure modes returned 200s: a think-chain in the JSON
-slot, and an empty completion because a reasoning model spent its whole budget
-before answering. The probe sends a canned mini-observation through the REAL
-decision parser at a >=1024 completion allowance and classifies the result as
-`http`, `empty` or `parse`. Use it before adopting any model id; provider
-catalogs list ids that 404 on invoke.
-
-**Provider trouble no longer disables an agent.** A permanent-looking model
-error (404/410/decommissioned) used to disable the agent after a threshold. On
-2026-08-26 NVIDIA end-of-lifed an entire model line and 35 agents died on that
-path. The runner now reports a hold and keeps retrying each cadence, recovering
-by itself when the provider does. Disables remain for what deserves them:
-revoked credentials, drawdown, kill-switch, user action.
-
-**Failures carry structured metadata.** A failed `decide()` now returns
-`status` and, when the provider sends one, `retryAfterMs` (parsed from
-`Retry-After` in both delta-seconds and HTTP-date form, capped at an hour), so
-a caller can tell a 429 from a 5xx without parsing strings. Error text is
-unchanged.
-
-**`ClientConfig.extraHeaders`.** Headers attached to every request, spread
-before auth so they can never clobber it. Self-host has nothing to put here;
-it exists so CoinRithm's own hosted scheduler can present its attestation
-channel.
-
-**Model names corrected throughout.** The retired Llama 3.x line is gone from
-the README, the runtime defaults and the `quant-reference` example, which is
-relocked onto `nvidia/nemotron-3-nano-30b-a3b`.
-
-## 0.7.6
-
-Agent capability release: universe discovery, first-class behavioral guards,
-and the hosted prose budget made visible. Additive — no tool renamed or
-removed. Contract moves to **1.7.0** (two keyless paths declared).
-
-**New: agents can look beyond their own watchlist.**
-
-- **`get_crypto_movers` tool.** Keyless scan of the tracked coin universe for
-  the biggest 24h gainers or losers. Rows carry `coinId`, `symbol`, `name`,
-  `slug`, `change24hPct`, `priceUsd`.
-- **`universe_scan` capability** for the self-host runner. Each cycle it pulls
-  the top movers, promotes the strongest few into full watch entries marked
-  `discovered: true`, and passes the remainder as compact context. Watchlist
-  and blocklist symbols are excluded up front, so a discovered row can never
-  duplicate a configured pair or bypass the deny list.
-- **Both now carry the coinId through.** The movers row's `ucid` IS the
-  `coinId` that `get_candles` / `get_market_context` / the futures quote path
-  take. It was previously stripped from the tool response and re-derived from
-  the SYMBOL via a resolve round-trip — a wasted call per discovered mover and
-  a real correctness hazard, because symbols collide across listings and the
-  resolver could return a different coin than the one that actually moved.
-
-**New: contract declares the endpoints the tools call.**
-
-- `/api/coins/top-gainers` and `/api/coins/top-losers` are now in
-  `openapi.yaml` (tag `public-crypto-data`), so both SDKs can reach the
-  surface `get_crypto_movers` uses. Probe-verified against prod: bare array,
-  no envelope; `change24h` / `currentPrice` are decimal STRINGS; default
-  `limit` is 3 and out-of-range values return 400 rather than clamping.
-
-**New: personality and boundaries are configurable, and documented.**
-
-- **`character/guards.md`** — first-class hard behavioral guards, merged into
-  the strategy prose as a distinct section rather than buried in the thesis.
-- **`examples/agents/pia-pump-fader`** — a full bundle demonstrating
-  capabilities plus boundary configuration (watchlist/blocklist interaction,
-  the five-point risk gate, re-entry discipline).
-- **`examples/agents/FORKING.md`** — a file-by-file map of what is strategy
-  and what is plumbing, so a fork knows what it is allowed to change.
-- **QUICKSTART** documents capabilities, and a docs-drift tripwire fails the
-  suite when a capability ships undocumented (`universe_scan` shipped
-  invisible in every user surface once; that cannot recur silently).
-
-**Fixed.**
-
-- **Hosted prose budget is validated, not discovered at deploy.**
-  `coinrithm-agent validate --hosted` now checks the 8,000-character merged
-  prose budget and reports the exact overage. A bundle could previously
-  validate clean and still be undeployable. YAML frontmatter is stripped
-  before the count (and before the model sees it — it was being fed in as if
-  it were strategy). `pia-pump-fader` was rebuilt to fit at 7,932.
-- **Permanent failures stop being revived.** A disabled agent whose model is
-  gone or whose key is invalid is no longer resurrected by the scheduler's
-  revive pass; only transient failures are retried.
-- **Fresh scaffolds are no longer bricked** by the capabilities field, and
-  action-confidence tolerance was widened to match what models actually emit.
-- **False market-data licensing assertion corrected** in both READMEs.
-
-⚠ Publishing to npm remains a **manual** step — `publish-mcp.yml` pushes
-`server.json` to the MCP registry only.
-
-## 0.7.5
-
-**Release-hygiene bump. Everything below was already merged but never
-reached npm** — the 0.7.4 tarball was published 2026-07-26T23:01:11Z and
-five commits landed after that instant without a version bump, so the
-repository's 0.7.4 and the published 0.7.4 were different code under one
-version number. This release makes the published artifact match the
-source again.
-
-- **Security.** MCP dependency audit fixes, 6 findings to 0 (`da65e9e`).
-  Anyone on published 0.7.4 is running the pre-audit dependency set.
-- **`pm_data` Gemini exposure** for agents (`3ab04ae`).
-- **Venue methodology and health** exposed as tools (`74cc495`).
-- **Reproducible decision receipts** persisted by the agent runner
-  (`8bebc9e`).
-- **Docs.** Contract version drift corrected and the placeholder SDK
-  README replaced, so the docs stop advertising an install that 404s
-  (`6d61b92`).
-
-No tool was renamed or removed; this is additive plus a dependency
-refresh.
-
-⚠ Publishing this package is a **manual** step — `publish-mcp.yml` only
-pushes `server.json` to the MCP registry, it does not run `npm publish`.
-That asymmetry is exactly how the drift above accumulated unnoticed.
-
-## 0.7.4
-
-Docs-only. No tool behavior change, no API-surface change.
-
-- **Acceptable Use of Market Data.** The README (root and this package) and
-  `openapi.yaml` (`info.termsOfService`, `info.description`, and the
-  `public-pm-data` tag) now reference and summarize CoinRithm's licensing
-  flow-down restriction on Market Data from third-party prediction-market
-  venues: read-only use for paper-trading context and settled-outcome
-  scoring only — no model training/fine-tuning/benchmarking, no
-  redistribution or bulk-extraction, no use to build a competing product.
-  Full terms: <https://www.coinrithm.com/en/terms-of-use>
-
-## 0.7.3
-
-Quality-engine surfaces + independent forecasts. Additive; no breaking change.
-
-- **Quality verdicts in tool responses.** `discover_pm_markets`, `pm_quote`, and
-  the `pm_data_*` tools now surface the persisted truth-engine `quality` object
-  (`decisionEligible`, warning/block reason codes, `policyVersion`, `assessedAt`).
-  Markets with critical failures stay visible but cannot drive paper opens or
-  alerts.
-- **`openBlocked` preview on `pm_quote`.** Quotes preview the open-time quality
-  gate (`openBlocked` + `openBlockReasons`), so an agent can skip a market that
-  would 422 before burning the open attempt. The self-host runner
-  (`coinrithm-agent`) does this skip automatically.
-- **Independent forecasts in the runner.** The self-host agent runner elicits the
-  model's OWN probability (judged from the question/resolution criteria/deadline,
-  never anchored to the market price) and submits it as `forecastProbability` on
-  PM opens — feeding the public calibration dataset with proper-scoring-rule
-  forecasts. Clamped to [1,99]; omitted (never faked) when the model does not
-  produce one; `HOUSE_AGENT_FORECAST_ENABLED=false` disables.
-- **`crossPlatform` on event lists** documented in the API contract: sibling
-  venues pricing the same question, on list rows.
-- **ForecastEx venue truth.** Public MCP discovery copy and registry metadata
-  now describe all 11 live venues, including ForecastEx.
-- **Contract synchronization.** Runner templates and example bundles pin the
-  served OpenAPI 1.6.0 contract; canonical scorecard paths are unambiguous.
-
-## 0.7.2
-
-Docs-truth + privacy release. No tool behavior change, no API-surface change.
-
-- **Ten venues in the public listing.** `pm_data_*` tool copy, the README, and
-  `server.json` now name all ten venues (adds Futuur and Myriad). npm `0.7.1` was
-  published before those landed, so the registry listing still advertised "eight
-  venues"; npm versions are immutable, so correcting the public listing required
-  a new release.
-- **`source` parameter description** on `pm_data_events` / `pm_data_event_detail`
-  now enumerates all ten venue slugs. Accepted values are unchanged — this is
-  description text only, which is why it is a patch and not a minor.
-- **Privacy.** Raw model output is no longer persisted, enforcing the package's
-  no-chain-of-thought promise.
-- **New tripwire.** `server.json` (the MCP-registry listing) is now guarded
-  against version and venue-count drift; it had no guard, which is how it went
-  stale in the first place.
-- Refreshed stale Arena-gate example copy.
-
-## 0.7.1
-
-Docs + registry-metadata release; no tool behavior changes.
-
-- **README refresh**: the keyless `pm_data_*` data surface is now front and
-  center — 8 venues (Polymarket, Kalshi, Smarkets, Limitless, Manifold,
-  Metaculus, PredictIt, Rothera — the "seven venues" line predated Rothera),
-  the anonymous hosted-endpoint path, and the `referenceProbability` /
-  `volumeHistory` fields the data tools return.
-- **`server.json`**: hosted endpoint's `Authorization` header marked optional
-  (the `pm_data_*` tools work anonymously — verified live) and the server
-  description now leads with the keyless data surface.
-- Ships the post-0.7.0 commits: `pm_data_event` advertises `volumeHistory`,
-  `pm_data_events` advertises `referenceProbability` on list items, and the
-  hosted MCP root (`GET /`) serves a self-describing JSON landing (with a
-  405 + hint on `GET /mcp`).
-
-## 0.7.0
-
-(Retroactive entry — released 2026-07-05 without a changelog note.)
-
-- **Four keyless `pm_data_*` tools** — CoinRithm's free public cross-venue
-  prediction-market dataset over MCP, no API key required and yours is never
-  attached: `pm_data_overview` (market-wide stats), `pm_data_events`
-  (cross-venue event list), `pm_data_event` (detail incl.
-  `crossSourceMatches` + resolution evidence), `pm_data_whales`
-  (large-trade tape).
-
-## 0.5.0
-
-Agent-runner quality + reliability release. `coinrithm-agent` got materially
-smarter and less noisy; `coinrithm-mcp` is unchanged in shape. Bundles the work
-since 0.4.0.
-
-- **Prediction markets are a first-class venue** in the decide prompt: a short
-  `pmN` ref so small models trade PM reliably, eligible-outcome filtering (only
-  backend-openable outcomes reach the model), and futures-capped agents steered
-  to PM (a separate budget) instead of re-rejecting.
-- **PM anti-churn — now actually effective.** The candidate list is pre-filtered
-  to exclude markets the agent already holds, and the runner + server both block
-  re-betting a held market+outcome (no more one-agent, 25-identical-bets churn).
-  An earlier version read the wrong `/positions/pm` fields and was silently dead;
-  fixed.
-- **Settlement-feedback learning loop.** The agent sees how its own recent bets
-  actually resolved (win/loss/void + realized PnL) as reflective context, so it
-  learns from outcomes across cycles.
-- **Per-trade reasoning stays honest about the market.** A multi-action decision
-  no longer stamps its primary rationale onto a secondary trade about a different
-  market — the trade's public Arena "why" always matches the market it's on.
-- **Futures reliability.** The model is unblinded to per-position mark /
-  liquidation / stop / take-profit prices; take-profit is auto-clamped to a valid
-  R:R target off the stop (kills the `take_profit_not_*_mark` reject waves); and a
-  marking-down PM book now trips the equity-drawdown kill-switch too.
-- **News capability.** Recent high-importance news for the watchlist coins is fed
-  into the decide context as a market-catalyst layer.
-- **Robustness + contract accuracy.** Scheduler/runner hardening, flat-state
-  prompt steers (weak models stop hallucinating closes), manage-enum
-  normalization, and the PM contract now documents real entry friction rather
-  than a disclose-only stance.
-- **Security.** `hono` bumped to 4.12.27 (high-severity advisories: serve-static
-  path traversal, CORS wildcard-with-credentials, body-limit bypass).
-- **Docs.** The npm README leads with value / free / OKF / Studio; stale
-  scheduler and `minDecidedTrades` claims corrected.
-
-## 0.4.0
-
-- **Deterministic scorecard engine (`computeScorecard`).** The reproducible-
-  evaluation engine for `coinrithm.agent.scorecard.v1` — pure math over an
-  agent's realized track record (no network, no model): realized PnL, win rate,
-  expectancy, profit factor, reward-to-risk, Sharpe, Sortino, deflated /
-  probabilistic Sharpe (Bailey & López de Prado — skill vs luck with a multiple-
-  testing penalty), max drawdown, and Brier + ECE calibration for probabilistic
-  calls. Same inputs → identical metrics **and** a sha256 `contentHash` of the
-  canonicalized result, so a scorecard whose hash doesn't reproduce isn't
-  trusted. Metrics are computed AFTER the run from immutable evidence (leakage-
-  separation), so tuning-to-the-metric is structurally impossible. Returns
-  `null` for thin records — never a fabricated number.
-- **Resolver: committable file metadata + functionality pin.** The OKF resolver
-  now carries per-file metadata and pins functionality through resolution, so a
-  bundle's behavior is reproducible from its committed files.
-
-## 0.3.0
-
-- **Agent risk config: coin deny-list (`blocklist`).** `risk.blocklist` lets an
-  agent name symbols it must never open, even if they are on the watchlist —
-  deny wins over allow. Enforced in the runner's decision validator (rejects
-  `futures_open` / `spot_order` on a denied symbol) and surfaced in the system
-  prompt, so the model is told the boundary and the runner re-checks it.
-- **Docs: Open Knowledge Format positioning.** Clarified that a CoinRithm agent
-  is an OKF bundle — a portable directory of markdown + frontmatter that is
-  model-agnostic (run the same definition on any model). Develop and prove it
-  free on paper, then run it anywhere.
-- npm keywords refreshed (`open-knowledge-format`, `okf`, `model-agnostic`,
-  `gemini`) for registry discovery.
-
-## 0.2.0
-
-- Added the **`coinrithm-agent`** self-host runner binary alongside the MCP
-  server: a folder-as-architecture (OKF) agent you bring your own model key to,
-  with caps enforced by the runner (not the model), dry-run by default,
-  paper-only.
-
-## 0.1.x
-
-- Initial `coinrithm-mcp` MCP server: reads, quotes, scoped spot/futures/PM
-  writes, ledger export, and Agent Arena integration over a user-minted API key.
+# Changelog
+
+All notable changes to `@coinrithm/mcp-trading` are documented here. The package
+ships two binaries — `coinrithm-mcp` (the MCP server) and `coinrithm-agent` (the
+self-host agent runner) — versioned together. The CoinRithm **API contract** is
+versioned separately (see `openapi.yaml` `info.version`, currently `1.8.0`).
+
+## 0.7.17 (2026-10-13)
+
+MCP server:
+
+- Opt-in PM paper v2 tools. `COINRITHM_PM_PAPER_V2_TOOLS_ENABLED=true` adds
+  `discover_pm_paper_v2_markets` (`GET /api/agent/pm/v2/discover`, scope
+  `read`), `open_pm_paper_v2` (`POST /api/agent/pm/v2/open`, scope `trade:pm`),
+  `close_pm_paper_v2` (`POST /api/agent/pm/v2/close`, scope `trade:pm`) and
+  `get_pm_paper_v2_positions` (`GET /api/agent/pm/v2/positions`, scope `read`).
+  They serve separately configured house keys only. `maxCashBudget` is an exact
+  decimal total that includes entry fees. `open_pm_paper_v2` requires
+  `forecastProbability` for the bought side (percent above 1 and below 100, or
+  a fraction below 1; exactly 1 is refused, matching the API). A close can
+  return HTTP 202
+  (accepted, still pending); only `executed: true` confirms the exit.
+- Opt-in spot reference tool. `COINRITHM_SPOT_REFERENCE_TOOLS_ENABLED=true` adds
+  `get_spot_reference` (`GET /api/agent/spot/reference/:coinId`, scope `read`):
+  a Binance spot last-ticker reference (USDT) with a stored REST consensus
+  fallback (USD), each with its own status and expiry. It is not an executable
+  quote and does not change market context, quotes or fills.
+- Both flags are off by default. The default surface stays at 41 tools; with
+  both flags on it is 46.
+- API clients add `discoverPmPaperV2`, `openPmPaperV2`, `replayPmPaperV2Open`,
+  `closePmPaperV2`, `pmPaperV2Positions` and `getSpotReference`.
+
+Agent runner and engine:
+
+- Capital-sizing agents: a correct-side futures take-profit that sits inside
+  the agent's own `capitalSizing.minRewardRisk` floor is widened to that floor
+  (fees included) instead of the open being refused as
+  `capital_quote_reward_risk_too_low`. A target at or beyond the floor is never
+  changed; agents without a capital policy are unaffected. Each repair is
+  logged and persisted as `takeProfitRepair` (requested target, applied target
+  and reason: `missing`, `wrong_side` or `below_reward_risk_floor`). Economics
+  that cannot produce a finite target fail closed.
+- `capitalSizing.highConviction` (opt-in, `equity_fraction_v1`): a
+  `futures_open` whose confidence is at or above `minConfidence` is sized and
+  rechecked with its own `perTicketCapitalPct` and `futuresRiskPct` instead of
+  the base ones. Total allocation, cash reserve, open-margin and reward:risk
+  caps still apply; prediction-market and spot sizing never read confidence.
+  Leo and Mia opt in (0.85 / 50% / 5%); the other house agents are unchanged.
+- The runner now shows the agent its last decision's refused actions (action,
+  rejection code and a short reason) in the next prompt, so it fixes the cause
+  instead of resubmitting the same action. The list is kept in the persisted
+  run state, capped at five, replaced by every validated decision and cleared
+  when the model chooses to skip.
+  Explicit server refusals of stop/target updates are included when the mark
+  moves after local validation. Network and server failures remain uncertain
+  outcomes; they are not described as confirmed refusals.
+- Prompts list each open futures position's stop-loss side of the observed mark
+  for `futures_set_sltp`, including whether a stop at the entry price would be
+  refused.
+- Prompts mark watched coins whose perpetual reference is stale or unavailable
+  as blocked for a NEW `futures_open` and give the remaining
+  `futuresEntryChoices`. When no candidate remains, the prompt does not offer
+  `futures_open` for that cycle. Runner validation, existing positions, closes
+  and stop/target updates keep their rules.
+- Futures closes answered with the `hl_paper_v2` execution model: an accepted
+  but unfilled exit (HTTP 202) is recorded as pending, not as executed or
+  failed, and does not advance the reject or failure kill-switch counters.
+  Replayed responses do not consume write budget or repeat journal entries.
+  Legacy responses keep their existing meaning.
+- PM paper v2 houses now fail closed locally on a missing or invalid
+  `forecastProbability`: the decision schema and the structured-output JSON
+  schema require it, so such a decision is rejected (`pm_v2_invalid_decision`)
+  before any checkpoint or API request, instead of reaching the API's
+  `forecast_required` refusal. The prompt states the accepted range. Stored
+  pending requests from earlier versions still replay unchanged. The hosted
+  scheduler picks this up through its `../mcp-trading` dependency.
+- Engine exports for hosted, explicitly enrolled PM paper v2 houses:
+  `runPmPaperCycle`, `readPmHousePolicies`, `selectPmHousePolicy` and
+  `parseDecisionForContract`. The `coinrithm-agent` CLI does not run this
+  cycle, and existing agents keep the legacy decision contract.
+
+## 0.7.16 (published 2026-10-07)
+
+- Preserve independently timed funding-by-venue context in watchlist and
+  discovered entries, including regular and compact prompts. Keep raw rates,
+  interval normalization and unknown source/next-funding times distinct from
+  settlement funding.
+- Bench reports retain up to two safe route attempts with response source,
+  failure class, actions-string diagnostic and admission categories. Legacy
+  missing evidence remains unknown; error text, latency, prompts and raw model
+  output are excluded. Record explicit truncation for nonstandard providers.
+- Carry open interest, positioning, sampled liquidations, macro proxies,
+  chain TVL, stablecoin supply and sampled order-book depth through the runner
+  with their separate freshness, coverage and provenance limits.
+- Keep news source URLs/categories and every related PM coin in context; add
+  the opt-in per-coin-horizon PM cap. Objective guidance stays within the
+  authored strategy and hard caps. Use the supported MCP SDK 1.31 dependency.
+
+- Bench: the PM baselines now preserve the recorded agent's PM board policy
+  (for house agents: depth 30, churn removed), the original opportunity set.
+  Before, they asked for the uncurated 12-row page a live mechanical agent
+  reads, and a cassette recorded without it (house Mia) reported the market
+  as missing. Originally mechanical recordings keep their uncurated board;
+  live mechanical agents are unchanged. New recordings need only the original
+  pass and mark PM baseline readiness after successful PM reads, rather than
+  assuming that an enabled PM venue guarantees recorded evidence.
+- Bench: two more PM baselines, `baseline:base-rate` (uninformative 50%) and
+  `baseline:random` (seeded 20-80%), next to `baseline:skip` and
+  `baseline:market`. They replay the same recorded market pass, so existing
+  corpora need no re-recording; futures/spot-only corpora still get skip only.
+- Add `coinrithm-agent label --corpus <dir> [--horizon-hours 24] [--overwrite]`:
+  writes `labels/<cassetteId>.json` price bars from candles published after
+  each cassette's asOf (reads only). It waits for the horizon to elapse, uses
+  the finest bar size the cassette's age still allows, keeps existing PM and
+  funding labels, and writes PM settlement only from the public event
+  verdict (settlement-eligible + "settle" + a provider won/lost outcome
+  result; anything else stays unlabelled). Only
+  bars that close by the horizon are stored, and an existing file for
+  another horizon is reported (`horizon_mismatch`), never kept silently.
+- Bench futures scoring walks only bars that open after asOf and close by
+  the horizon end; a horizon off the bar grid stops at the last full bar
+  before it (`horizonFlooredToBar`), so no price after the horizon is used.
+- Add optional per-agent `signals` thresholds (`rsiOversold`, `rsiOverbought`,
+  `strongMovePct`, `leanMovePct`, `minStrength`) and a `data.indicatorRange`
+  data diet (`1D` 5m bars, `1W` 15m, `1M` 1h, `3M` 4h). They change the setup
+  shortlist AND the gate that decides when the model is called, so personality
+  changes what an agent notices. Absent blocks keep the previous constants and
+  the 1D range exactly. Both are validated fail-closed, strict-linted and
+  `$ref`-able; the prompt names the agent's bar size and RSI levels.
+
+- Every watch entry now carries perpetual open interest when the API serves it
+  (`openInterest`: single-side USD over Bybit and OKX, 1h and 24h change,
+  `asOf`, `stale`), read from the existing `/api/agent/market` call. The prompt
+  explains how to read it with price; it is context, never a trade rule. Older
+  APIs without the block leave entries unchanged.
+
+- Add an agent test bench: `coinrithm-agent record` saves the read inputs of
+  dry-run cycles as cassettes (writes refused, no model call), and
+  `coinrithm-agent bench` replays them through the production runner for
+  several agent variants and repeats, next to skip and market-implied
+  baselines. The deterministic report gives reject codes, action overlap,
+  repeat consistency, seeded paired bootstrap CIs, a calibrated A/A null
+  rate, a chronological holdout and optional label scoring. Quotes are
+  synthesized and counted; see "Bench" in docs/agent-runner.md for the limits.
+
+- Add an optional `universe` block (market boundaries): a market-cap rank
+  band, a minimum 24h volume, stablecoins in or out, curated sectors and chain
+  ecosystems in or out, and the sort that defines an opportunity. Each cycle
+  the runner asks `GET /api/agent/universe` for rows inside those boundaries,
+  resolves the top `resolveTop` (max 10) into full `discovered` watch entries
+  and passes the rest as context. Rows are re-checked client-side, so a server
+  can narrow the set but never widen it; the blocklist still wins and every
+  cap applies. A failed screener degrades to the watchlist only. The block
+  can live in its own file via `$ref`. Without it, `universe_scan` keeps the
+  previous top-gainers scan exactly. Needs backend-v2 with the screener.
+
+- Add three optional, runner-enforced prediction-market policy fields, so
+  numbers that strategies stated only in prose become executable. Absent keeps
+  the previous behaviour exactly; malformed values fail closed.
+  - `risk.pmMinEdgeGapPct` (0..100): the forecast must beat the fee-inclusive
+    cost by that share of the room left to 100 (16 means cost 50 needs 58). It
+    never lowers the global minimum edge, and with it set a `pm_open` without
+    `forecastProbability` is rejected (`pm_forecast_required`,
+    `pm_edge_below_gap_rule`). Mechanical benchmarks are exempt.
+  - `risk.pmMaxEdgePoints` (0..100): overconfidence guard; the forecast may
+    beat the fee-inclusive cost by at most that many points, and a forecast
+    becomes mandatory (`pm_edge_overconfident`). This is an optional action
+    constraint; its effect on future performance requires separate evaluation.
+  - `risk.pmMaxOpenPerEvent` (1..50): open bets per event (source + slug),
+    counting held positions and opens accepted this cycle (`pm_event_cap`).
+  - `risk.pmMaxOpenPerCoinHorizon` (1..50): open bets on the same directly
+    linked coin whose events end within 24 h of each other, across different
+    events, counting held positions and opens accepted this cycle
+    (`pm_coin_horizon_cap`; a value outside 1..50 fails closed). Every linked
+    coin is kept. Reads the additive discover `relatedCoins` and
+    `/positions/pm` `relatedCoins` / `eventEndDate`; a missing coin, or an end
+    that is missing, past or over 5 years out, never blocks. A count cap, not
+    a correlation model, over the open positions the feed returns (at most
+    200, newest first).
+  - `risk.pmMinMinutesToClose` (0..10080): reject a market whose known close is
+    nearer than that to the later of `observation.asOf` and the validation
+    clock after the model call (`pm_closes_too_soon`). An unknown or
+    unparseable close never blocks.
+
+- Keep repeated provider 404/410 availability failures out of the generic model
+  failure kill-switch. BYO and self-hosted agents retain their configured model
+  and provider holds without entering a disable/revive loop. Transient failures
+  and invalid decisions still count toward the existing kill-switch.
+- Keep rejected model text out of retained parser errors: JSON syntax errors
+  omit input excerpts, and schema errors omit rejected enum values and unknown
+  field names. Useful field/type diagnostics remain; validation is unchanged.
+- Reject OpenAI-compatible and Anthropic responses that explicitly end at the
+  output token limit, and Anthropic responses stopped at the context-window
+  limit, even if their partial decision text happens to be valid JSON. No action
+  from that incomplete response reaches execution. Preserve reported token usage
+  and the hosted router's existing bounded malformed-response fallback. A
+  truncated direct-provider retry also retains its malformed classification.
+- Classify rejected actions strings with bounded, fixed diagnostic categories.
+  JSON5 is used only for diagnostics; string-valued actions remain rejected.
+  No raw model output is retained by this diagnostic path.
+
+## 0.7.15 (published 2026-10-02)
+
+The verified release includes source changes through 1 October 2026. Its public
+archive and MCP Registry delivery are recorded in the repository publishing log.
+
+- PM board rows carry event settlement context from discovery's optional
+  `resolution`. Where `scope` is `per_outcome`, each row also carries its own
+  provider-bound `outcomeRule`, preserving exact primary/secondary terms or
+  explicit unknown status. Event-level text does not substitute for another
+  outcome's rule. Missing fields on older APIs remain unknown.
+- Futures agents see `futuresCapacity` in the user prompt: open vs maximum
+  positions, slots left and margin headroom, computed exactly as the runner
+  validates. With no slot or no margin left, the prompt says plainly that no
+  futures_open (new or add) can pass. Production showed agents at 3 of 3
+  positions proposing a new open on most cycles (max_positions 222 and
+  open_margin_exceeds_cap 116 rejections in 6h), each a wasted model call.
+  The validator still enforces both caps.
+- The PM system prompt no longer claims a "genuine information edge" on crypto
+  price markets or makes a PM bet REQUIRED. It says the market already prices
+  the same public data, that short-dated price markets are usually efficient,
+  and that a bet needs a probability gap larger than costs plus a named reason;
+  skipping PM is always valid. The ref anti-hallucination rule is unchanged.
+- Every non-mechanical agent (not only `objective.primary: calibration`) gets
+  the micro-contract churn filter and the deeper 30-row discovery page,
+  including the Bitcoin fallback. Mechanical baselines keep the unmodified
+  12-row universe.
+- PM board rows carry the discover event's optional `referenceProbability` as
+  event-level `consensus` (`prob` 0..1, `venues`, `spreadPts`, `kind`,
+  `outcome`), never mapped to or flipped for a specific outcome. Absent or
+  unusable stays unknown.
+- Non-mechanical PM agents see their own settled forecast record
+  (`pmCalibration` from `GET /api/agent/performance`) when it has at least 20
+  settled forecasts. It is read at most once per 30 minutes per credential,
+  without retries, and omitted on any failure.
+
+- Preflight new futures entries using informational reference eligibility;
+  validate the available position and margin capacity across actions in the
+  same cycle. Adds/exits retain their separate rules, and the API rechecks
+  execution-time eligibility. Preserve reported capacity and spot fill
+  receipts in decision evidence without claiming full historical replay.
+- Describe forecasts as market-aware, not price-blinded. A capacity-limited
+  agent may still exit or adjust protection. Withheld-action diagnostics are
+  attributed to the provider response that produced them.
+- Add `get_news` for scored crypto news, bringing the source MCP surface to
+  41 tools. Arena reads work without a key; discovery defaults to open
+  markets. Default event results remain compact, with `detail: "full"` for
+  the complete event response. Preserve event subtitles and per-outcome
+  price basis in bounded output.
+
+- The hosted HTTP server recovers the pasted-markdown descriptor link
+  (`/mcp` followed by `](`, literal or percent-encoded): GET/HEAD redirect to
+  the service descriptor and POST runs the same MCP handler as `POST /mcp`.
+  Every other unknown path stays 404 and `GET /mcp` stays 405.
+- `pm_data_canonical` documents the canonical `consensus` (or null) and the
+  labelled `consensusHistory`, and names orientation `flipped` as the API does.
+
+- Preserve optional observed-price flags and bounded source provenance in public
+  market summaries, including the exact sides of venue comparisons. A missing
+  observed price remains distinct from an observed zero and unknown provenance.
+  This source change is not included in published npm `0.7.14`.
+
+- Hosted scheduler rollout supports the opt-in `whale_context` runner capability. After the deterministic
+  gate fires for a PM-scoped, non-mechanical cycle, it attaches bounded public
+  tape context and up to two validated wallet movement summaries to the
+  observation. Reads are best effort with explicit availability status and do
+  not widen candidates or change risk/write gates. The observation fingerprint
+  includes this context; the existing partial persisted audit projection does
+  not store raw whale or wallet samples. Published MCP `0.7.14` remains unchanged.
+
+- Preserve full public `walletAddress` values in whale tape and default event
+  summaries for direct wallet-tool follow-up on Polymarket, Limitless and Myriad.
+  Anonymous/null and absent identities remain unchanged. This source fix is not
+  included in the published npm `0.7.14` package.
+- Record explicit zero model calls, tokens, inference cost and action writes for
+  cycles that return before inference due to a kill switch, observed drawdown,
+  failed account reads or an empty market selection. Historical records remain
+  unchanged; failures after a possible provider call retain unknown evidence.
+
+## 0.7.14
+
+- Scope permanent model-error streaks to the attempted provider/model. Discard
+  legacy unattributed streaks and reset availability failures on a successful
+  provider response, including malformed decisions. Old-route failures cannot
+  request an early hold on a replacement model; agent risk limits are unchanged.
+- Attribute a routed permanent-error hold to the attempt that produced the error,
+  even when a later fallback is rate-limited; keep actual-call metering unchanged.
+- Preflight futures protection updates against observed side, mark and liquidation
+  prices, including retained triggers. Reject known invalid end-states with an
+  actionable reason before a write; the API remains authoritative as prices move.
+- Add the optional `risk.pmMinEntryProbabilityPct` policy (0..100 points): the
+  runner rejects `pm_open` when the chosen outcome's raw market probability at
+  entry is below the floor, with a reason that names both numbers. Fees stay
+  in the forecast-edge check, so a 19-point outcome fails a 20-point floor even
+  when fees lift its cost above 20. Absent keeps today's behaviour; a set floor
+  with no quoted probability fails closed. The floor is rendered in the prompt's
+  hard caps.
+- Retain candle timestamps and compact coverage/interval evidence alongside
+  indicators, separately from market-price freshness. Missing timestamps stay
+  unknown; nominal five-minute cadence does not imply current or regular bars.
+  Model input and private decision evidence retain the same context.
+- Load optional `character/entries.md`, `exits.md`, `sizing.md` and `research.md`
+  from local bundles so Studio strategy sections reach the model, manifest and
+  compiled definition. Existing bundles without these files are unchanged.
+- Include the exact compiled strategy definition and its digest in local
+  inspection; expose the same snapshot builder to engine consumers.
+- Add `run --expect-definition` to reject a changed baseline before model or
+  account access. The digest is not a full replay record or broker adapter.
+- Hosted validation accepts `limits.maxDailyLossMusd: 0` as no daily loss cap,
+  matching the runner and the hosted API; negative and non-finite values are
+  still rejected.
+- Tactic cap merges treat 0 as unlimited for `maxTradesPerDay` and
+  `maxDailyLossMusd`: a tactic may tighten 0 to a positive cap but can no
+  longer turn a positive cap into 0. Writes and margins keep lower-is-tighter.
+
+Publication and hosted deployment are verified separately. See the repository's
+[release status](https://github.com/CoinRithm/coinrithm-agent-trading#version-clarity)
+for registry availability and the hosted runtime evidence.
+
+## 0.7.13
+
+- Preserve optional per-outcome venue terms in compact prediction-market output,
+  including explicit unavailable values and valid false/zero facts.
+- Include candle volume coverage in agent fundamentals: positive counts identify
+  missing expected venues; unknown coverage stays unknown. Preserve measured zero
+  volume and keep each volume paired with its own coverage.
+- Classify provider capacity failures consistently, including HTTP 503
+  resource exhaustion, in runner results and configured same-model retries.
+- Export `classifyProviderFailure` from the supported engine entry point.
+  Retry counts and trading behavior are unchanged.
+- Surface signed cumulative futures funding and its applied-through timestamp to
+  the agent observation so funding already reflected in balances is not counted twice.
+- Add public whale-wallet context to the 40-tool surface: bounded 7-day or
+  30-day wallet activity and source/wallet movement detail, with provenance and
+  provider-reported context kept distinct from CoinRithm's observed trade flow.
+- Clarify `pm_data_calibration` as market-price calibration: the primary lane
+  uses one complete-book snapshot in the inclusive 20–28 hour pre-resolution
+  window, with event-weighted ECE and separate `finalPrice`/`ownCapture` lanes.
+
+Publication and hosted deployment are verified separately. See the repository's
+[release status](https://github.com/CoinRithm/coinrithm-agent-trading#version-clarity)
+for registry availability and the hosted runtime evidence.
+
+## 0.7.12 — 2026-09-15
+
+Clarify `whoami`, `cancel_spot_order` and `report_pm_opportunity` descriptions,
+removing execution-cost prose unrelated to these operations. Document actual
+authentication, side effects, result fields and retry behavior.
+
+Spot cancellation now advertises its existing idempotent behavior. Opportunity
+reporting no longer advertises unconditional idempotency: duplicate protection
+requires `decisionId` (or `agentTrace.decisionId`) under the same API key; the
+first stored record wins. Reporting remains a write despite requiring only the
+`read` scope, and its evidence remains explicitly self-reported.
+
+Tool names, accepted inputs and execution behavior are unchanged. Published on
+npm and deployed to hosted MCP on 2026-09-15; registry and GitHub downloads match
+the CI archive. A clean registry installation passed startup and tool metadata
+checks. See the [release record](https://github.com/CoinRithm/coinrithm-agent-trading/releases/tag/mcp-trading-v0.7.12).
+Glama's evaluation score is separate from package delivery.
+
+## 0.7.11 — 2026-09-15
+
+Fix opportunity reporting that previously treated resolved API failures as
+successful submissions. Only `ok: true` confirms a report. HTTP errors remain
+unconfirmed; transport failures and exceptions have an unknown delivery outcome.
+
+`CycleResult.opportunity` now contains confirmed reports only. The additive
+`opportunityReport` field preserves the attempted payload, outcome and status,
+without API error bodies or exception details. Existing consumers should use
+this field when they need attempted rather than confirmed evidence.
+
+The reporter retains its one-invocation-per-cycle latch and adds no retries.
+Focused regressions compare successful and failed reporting in skip/act cycles
+and verify unchanged trading results and runner state. SDK versions are unchanged.
+This source entry does not establish registry publication or hosted deployment.
+
+## 0.7.10 — 2026-09-15
+
+Source changes following the published 0.7.9 release. The TypeScript SDK remains
+0.3.1 and Python remains 1.8.1; their runtime source is unchanged.
+
+- Persist file-backed run identity before execution so a first-cycle process
+  crash cannot discard the idempotency identity. Transport uncertainty is still
+  not automatically replayed or recorded as a completed trade.
+- Fix runner startup on Node 18: use the imported Node crypto API instead of
+  depending on a global crypto object. The installed-package matrix reproduced
+  this failure on Linux, Windows and macOS.
+- Add the supported `@coinrithm/mcp-trading/engine` entry point, preserving
+  existing deep imports, and separate observation accounting and opportunity
+  reporting from cycle ordering.
+- Add opt-in, machine-checked crypto return predicates and visible warnings for
+  inactive/reserved configuration. Existing agents are not automatically opted in.
+- Serialize scheduler migrations in a bounded transaction; add an offline
+  credential-rotation helper and interruption/recovery rehearsal.
+- Pin workflow actions and verify release-tool checksums. Add installed-package
+  compatibility and restart smoke checks across operating systems and runtimes.
+
+  0.7.10 was published and deployed on 2026-09-15. Registry and GitHub downloads
+  matched the CI-tested archive; hosted MCP and scheduler deployments finished.
+  See the [release record](https://github.com/CoinRithm/coinrithm-agent-trading/releases/tag/mcp-trading-v0.7.10).
+
+## 0.7.9 - 2026-09-15
+
+Published on npm and verified on 2026-09-15: the registry archive matches the
+reviewed artifact and passes fresh-install checks. See the
+[combined release notes](https://github.com/CoinRithm/coinrithm-agent-trading/releases/tag/mcp-trading-v0.7.9)
+for source provenance, verification and community acknowledgments.
+
+Public market-data fidelity and runner reliability release. No MCP tool was
+renamed or removed, and the API **contract stays 1.7.0**.
+
+**PM evaluation budget.** Event-driven periodic prediction-market evaluations
+now respect `maxLlmCallsPerHour` after their cooldown elapses. Budget skips make
+no provider call and consume no call allowance. PM keeps its own cooldown;
+open-position management and explicit always-on behavior retain their existing
+exemptions. This runner gate is separate from hosted provider-capacity admission.
+
+**Retry-After parsing.** Missing, blank or malformed headers no longer become
+zero-delay retries. The runner API client uses its existing five-second fallback;
+explicit zero, numeric seconds and HTTP dates remain supported. Model-provider
+cooldowns share the parser and retain their existing one-hour cap.
+
+**API request deadlines.** Each runner API operation now has a 30-second total
+deadline covering response headers, body reads and all 429 retry waits. The
+same client serves the hosted scheduler. Embedded callers can set a finite
+`requestTimeoutMs` and supply an `AbortSignal`. A timeout or cancellation returns
+an uncertain transport result without automatically replaying a trading write.
+Timers and listeners are removed when the operation finishes.
+
+**State persistence.** Self-host state is serialized to a private temporary file
+and atomically renamed over the previous state. A failed serialization or rename
+leaves the prior state intact. This is atomic replacement, not a claim of durable
+storage across power loss.
+
+**Agent conversion.** `coinrithm-agent eject` preserves explicit `triggerPolicy`
+and `capitalSizing` blocks. Previously conversion could restore default hourly
+budgets and drop equity sizing.
+
+**Release verification.** All-source coverage gates, mandatory PostgreSQL CI,
+dependency updates and corrected client setup docs are included. See the
+[reliability record](https://github.com/CoinRithm/coinrithm-agent-trading/blob/main/docs/RELIABILITY.md).
+
+**Confirmed-action journal.** Completed-action memory now requires an action
+to be both accepted and executed. Failed writes and uncertain transport results
+retain their attempt evidence without becoming completed moves in the next
+decision prompt. Dry-run proposals remain unexecuted.
+
+**Direct NVIDIA retry.** One complete HTTP 500/502/503/504 response can be
+retried once on the identical direct NVIDIA route within the original deadline.
+Both attempts are retained. This does not retry trading writes or change the
+hosted shared-pool routing policy.
+
+**Capital reconciliation.** Frozen-balance rounding residue down to -1e-8 is
+normalized only in the sizing calculation after the independent reads agree.
+Negative spendable cash still fails closed; wallet balances are not changed.
+
+**Private decision input evidence.** The bounded numeric projection includes
+nested indicator inputs and context movers, with legacy v1 records still readable.
+It does not retain hidden reasoning or raw model output.
+
+**Compact prediction-market evidence.** Discovery and compact event-detail
+responses now retain the API's `source.quoteScale`, `source.methodology` and
+`source.supportsMarketMetrics`, plus `spreadPoints`, `probabilityBook` and
+each retained outcome's `normalizedProbability`. Venue-native bid/ask quotes
+are never rescaled or interpreted from magnitude. Normalization remains the
+API's calculation over the original full book, not the truncated top-five
+outcome list. Existing payload bounds and explicit `detail: full` behavior
+are unchanged.
+
+**Settlement-time provenance.** Compact events retain `resolvedAtBasis` and
+`settlementWindowClosedAt`, keeping provider expiration distinct from an
+announced settlement time. Null and absent upstream evidence stay null and
+absent; the MCP does not infer missing values.
+
+**Candle semantics.** The `get_candles` description now states that these are
+sampled composite-price bars. Each bar's `v` is a mean rolling 24-hour
+quote-volume observation in USD, not volume traded during the candle, and
+must not be summed across bars.
+
+**HTTP completion diagnostics.** The hosted HTTP
+entry now has a bounded, stderr-only completion observer with final SDK-result
+and finish/abort accounting. Initialization, discovery, tool failures and
+successful delivery are distinct; unknown tool names are normalized. Records
+contain no arguments, bodies, credentials, caller/RPC IDs or caller-origin labels.
+Credential presence is not authentication. Durations describe the HTTP request,
+shared by batch members; server finish does not prove client receipt or use.
+Stdio, tools, authentication and dependency versions are unchanged. See
+`DEPLOY.md` for the measurement and retention limits. Hosted source/image
+`18a0bb6a8a0665e91cebc10225fec6f7ebcdaaf7` passed a bounded anonymous smoke on
+2026-09-13. Hosted verification and npm publication are separate release steps.
+
+**Deployment boundaries.** Hosted scheduler admission reasons are private
+scheduler telemetry, not a new SDK or MCP response field. The API's corrected
+comparison probabilities and enriched spread names use the existing response
+shape and reach current clients through fresh API reads. Outcome display names
+may change; use source/event/outcome identifiers for identity, never summed
+prices or matching labels alone. These fixes do not establish trading returns.
+
+## 0.7.8
+
+Runner decision-quality, evidence and paper-capital release. Additive: no MCP
+tool was renamed or removed, and the API **contract stays 1.7.0**. This release
+contains all package changes since published 0.7.7 (`gitHead` `80d0cae`), not
+just the previously listed thesis work.
+
+**Thesis exits.** Every opening action (`futures_open`, `spot_order`,
+`pm_open`) now carries a `thesis`: a one-sentence summary plus an
+`invalidation` with at least one machine-checkable condition (`priceBelow` /
+`priceAbove` for coins, `probabilityBelow` / `probabilityAbove` for prediction
+markets, a `maxHoldMinutes` time stop, and a free-text `catalyst` the model
+re-judges itself). The runner binds the thesis to the position the server
+returns, sanitized side-aware (a rising price never invalidates a long; a
+wrong-side level is dropped rather than re-signed; the time stop is clamped to
+60 minutes .. 30 days), persists it in the run state (`RunState.theses`, the
+same state file / `agent_state` JSON as before, no schema change) and
+re-evaluates it every cycle. A futures position whose price level or time stop
+is breached is closed by the runner before the model is asked anything, logged
+as a `thesis_invalidated` exit with its own idempotency key, after the
+kill-switch and drawdown checks and never instead of them. Prediction-market
+positions have no close endpoint, so a broken PM thesis is surfaced to the
+model instead (do not add, let it settle). The parser is tolerant (a malformed
+thesis never fails the open; a thesis copied onto a close is ignored) and the
+structured-output schema requires it, so schema-enforced hosted models always
+emit one.
+
+**Fundamentals in the observation.** Each watch entry now carries
+`fundamentals` sourced only from calls the runner already makes: `categories`,
+`marketCapRank` and `marketCapUsd` from the market context; `volume24hUsd` from
+the candles the `indicators` capability already fetches (live-probed
+2026-09-02: each bar's `v` is a rolling 24h volume, so the latest bar is the
+24h figure, never the sum); and up to three `headlines` with `publishedAt`
+timestamps from the one `news` call, attributed through the curated coin-news
+graph. Discovered PM markets carry `endDate` and `liquidityUsd`; open PM
+positions carry their title, side, entry and current probability and
+`openedAt`; open futures positions carry `openedAt`. The system prompt states
+the thesis contract, the runner-enforced exit and how to grade a trade on the
+fundamentals. Not carried, because no agent endpoint serves them: an "about"
+text per coin, a 24h probability change and a cross-venue divergence per PM
+market.
+
+**Fix:** the public movers feed serializes `change24h` / `currentPrice` as
+decimal strings; the universe-scan context rows read them strictly as numbers
+and shipped `undefined` for every mover.
+
+**Opt-in equity-based paper sizing.** A runner can size entries from a
+conservative fraction of its independently attributed paper book instead of a
+fixed stake/margin. The book is accepted only when wallet identity, cash
+partitions, held-position attribution and spot-mark coverage reconcile. Quotes
+then enforce per-entry, per-symbol, deployed-capital and daily-entry limits;
+fee buffers and the API's fee-inclusive quote evidence are included. Any
+missing or inconsistent evidence fails closed. Legacy positions on a different
+book remain visible for management but never inflate the current book's buying
+power.
+
+**Prediction-market decisions use executable economics.** PM opens now reject
+an invalid raw probability and a model forecast that does not clear the quoted
+entry price. Forecast edge is measured against the actual fee/slippage-adjusted
+fill, not the headline market probability. Quote-expiry outcomes are recorded
+separately from risk/balance rejection, and futures risk/reward validation uses
+fee-inclusive entry and stop economics.
+
+**Decision evidence is structured and bounded.** Cycles can expose a sanitized,
+partial private decision-input record: configuration and observation
+fingerprints, daily budget and guard state, plus bounded observation rows with
+explicit omission counts. It is not a prompt, transcript, raw model output or
+hidden reasoning record. The runner also reports quote/validation evidence for
+abstained, forecast-only and quote-expired PM opportunities. Hosted persistence
+and retention remain the caller's responsibility.
+
+**Runtime controls are more faithful.** The model sees the remaining daily
+entry/add budget rather than only static maxima. Entry caps still block new
+risk, while closes and other risk-reducing actions remain available. Direct
+provider HTTP 429 responses are capacity skips rather than model failures, so
+BYO agents do not build a failure streak during ordinary quota pressure.
+Structured-tool decisions remain required where the provider supports that
+contract.
+
+**Scorecard fix.** Maximum drawdown now measures decline from starting equity,
+so an immediate loss is no longer hidden by treating the first post-trade point
+as the high-water mark.
+
+## 0.7.7
+
+Reliability release. Every change here came from a live production failure, not
+from a roadmap. Additive: no tool renamed or removed, and the API **contract
+stays 1.7.0** because nothing on the documented surface changed.
+
+**Model requests are now built from a declared capability table, not
+assumptions.** `providerCapabilities.ts` states, per model family, which
+parameter carries the completion budget, whether a non-default temperature is
+allowed, and what extra body fields the family needs. Two failures this fixes:
+
+- **OpenAI's current models rejected our requests outright.** `gpt-5*` and
+  `o*` refuse `max_tokens` and any non-default `temperature`; they take
+  `max_completion_tokens`. The family is detected by MODEL id, not just the
+  provider name, so an OpenAI-compatible gateway serving `gpt-5` gets the same
+  shape. If you brought your own OpenAI key, this is why it now works.
+- **NVIDIA Nemotron models emitted a think-chain where the JSON decision
+  belonged**, which failed every cycle. The `chat_template_kwargs.enable_thinking=false`
+  switch and the "detailed thinking off" system hint are now encoded as data
+  rather than re-learned by failing.
+
+**New: `probeDecisionContract()`.** An HTTP 200 is not proof a route can run an
+agent. Both production failure modes returned 200s: a think-chain in the JSON
+slot, and an empty completion because a reasoning model spent its whole budget
+before answering. The probe sends a canned mini-observation through the REAL
+decision parser at a >=1024 completion allowance and classifies the result as
+`http`, `empty` or `parse`. Use it before adopting any model id; provider
+catalogs list ids that 404 on invoke.
+
+**Provider trouble no longer disables an agent.** A permanent-looking model
+error (404/410/decommissioned) used to disable the agent after a threshold. On
+2026-08-26 NVIDIA end-of-lifed an entire model line and 35 agents died on that
+path. The runner now reports a hold and keeps retrying each cadence, recovering
+by itself when the provider does. Disables remain for what deserves them:
+revoked credentials, drawdown, kill-switch, user action.
+
+**Failures carry structured metadata.** A failed `decide()` now returns
+`status` and, when the provider sends one, `retryAfterMs` (parsed from
+`Retry-After` in both delta-seconds and HTTP-date form, capped at an hour), so
+a caller can tell a 429 from a 5xx without parsing strings. Error text is
+unchanged.
+
+**`ClientConfig.extraHeaders`.** Headers attached to every request, spread
+before auth so they can never clobber it. Self-host has nothing to put here;
+it exists so CoinRithm's own hosted scheduler can present its attestation
+channel.
+
+**Model names corrected throughout.** The retired Llama 3.x line is gone from
+the README, the runtime defaults and the `quant-reference` example, which is
+relocked onto `nvidia/nemotron-3-nano-30b-a3b`.
+
+## 0.7.6
+
+Agent capability release: universe discovery, first-class behavioral guards,
+and the hosted prose budget made visible. Additive — no tool renamed or
+removed. Contract moves to **1.7.0** (two keyless paths declared).
+
+**New: agents can look beyond their own watchlist.**
+
+- **`get_crypto_movers` tool.** Keyless scan of the tracked coin universe for
+  the biggest 24h gainers or losers. Rows carry `coinId`, `symbol`, `name`,
+  `slug`, `change24hPct`, `priceUsd`.
+- **`universe_scan` capability** for the self-host runner. Each cycle it pulls
+  the top movers, promotes the strongest few into full watch entries marked
+  `discovered: true`, and passes the remainder as compact context. Watchlist
+  and blocklist symbols are excluded up front, so a discovered row can never
+  duplicate a configured pair or bypass the deny list.
+- **Both now carry the coinId through.** The movers row's `ucid` IS the
+  `coinId` that `get_candles` / `get_market_context` / the futures quote path
+  take. It was previously stripped from the tool response and re-derived from
+  the SYMBOL via a resolve round-trip — a wasted call per discovered mover and
+  a real correctness hazard, because symbols collide across listings and the
+  resolver could return a different coin than the one that actually moved.
+
+**New: contract declares the endpoints the tools call.**
+
+- `/api/coins/top-gainers` and `/api/coins/top-losers` are now in
+  `openapi.yaml` (tag `public-crypto-data`), so both SDKs can reach the
+  surface `get_crypto_movers` uses. Probe-verified against prod: bare array,
+  no envelope; `change24h` / `currentPrice` are decimal STRINGS; default
+  `limit` is 3 and out-of-range values return 400 rather than clamping.
+
+**New: personality and boundaries are configurable, and documented.**
+
+- **`character/guards.md`** — first-class hard behavioral guards, merged into
+  the strategy prose as a distinct section rather than buried in the thesis.
+- **`examples/agents/pia-pump-fader`** — a full bundle demonstrating
+  capabilities plus boundary configuration (watchlist/blocklist interaction,
+  the five-point risk gate, re-entry discipline).
+- **`examples/agents/FORKING.md`** — a file-by-file map of what is strategy
+  and what is plumbing, so a fork knows what it is allowed to change.
+- **QUICKSTART** documents capabilities, and a docs-drift tripwire fails the
+  suite when a capability ships undocumented (`universe_scan` shipped
+  invisible in every user surface once; that cannot recur silently).
+
+**Fixed.**
+
+- **Hosted prose budget is validated, not discovered at deploy.**
+  `coinrithm-agent validate --hosted` now checks the 8,000-character merged
+  prose budget and reports the exact overage. A bundle could previously
+  validate clean and still be undeployable. YAML frontmatter is stripped
+  before the count (and before the model sees it — it was being fed in as if
+  it were strategy). `pia-pump-fader` was rebuilt to fit at 7,932.
+- **Permanent failures stop being revived.** A disabled agent whose model is
+  gone or whose key is invalid is no longer resurrected by the scheduler's
+  revive pass; only transient failures are retried.
+- **Fresh scaffolds are no longer bricked** by the capabilities field, and
+  action-confidence tolerance was widened to match what models actually emit.
+- **False market-data licensing assertion corrected** in both READMEs.
+
+⚠ Publishing to npm remains a **manual** step — `publish-mcp.yml` pushes
+`server.json` to the MCP registry only.
+
+## 0.7.5
+
+**Release-hygiene bump. Everything below was already merged but never
+reached npm** — the 0.7.4 tarball was published 2026-07-26T23:01:11Z and
+five commits landed after that instant without a version bump, so the
+repository's 0.7.4 and the published 0.7.4 were different code under one
+version number. This release makes the published artifact match the
+source again.
+
+- **Security.** MCP dependency audit fixes, 6 findings to 0 (`da65e9e`).
+  Anyone on published 0.7.4 is running the pre-audit dependency set.
+- **`pm_data` Gemini exposure** for agents (`3ab04ae`).
+- **Venue methodology and health** exposed as tools (`74cc495`).
+- **Reproducible decision receipts** persisted by the agent runner
+  (`8bebc9e`).
+- **Docs.** Contract version drift corrected and the placeholder SDK
+  README replaced, so the docs stop advertising an install that 404s
+  (`6d61b92`).
+
+No tool was renamed or removed; this is additive plus a dependency
+refresh.
+
+⚠ Publishing this package is a **manual** step — `publish-mcp.yml` only
+pushes `server.json` to the MCP registry, it does not run `npm publish`.
+That asymmetry is exactly how the drift above accumulated unnoticed.
+
+## 0.7.4
+
+Docs-only. No tool behavior change, no API-surface change.
+
+- **Acceptable Use of Market Data.** The README (root and this package) and
+  `openapi.yaml` (`info.termsOfService`, `info.description`, and the
+  `public-pm-data` tag) now reference and summarize CoinRithm's licensing
+  flow-down restriction on Market Data from third-party prediction-market
+  venues: read-only use for paper-trading context and settled-outcome
+  scoring only — no model training/fine-tuning/benchmarking, no
+  redistribution or bulk-extraction, no use to build a competing product.
+  Full terms: <https://www.coinrithm.com/en/terms-of-use>
+
+## 0.7.3
+
+Quality-engine surfaces + independent forecasts. Additive; no breaking change.
+
+- **Quality verdicts in tool responses.** `discover_pm_markets`, `pm_quote`, and
+  the `pm_data_*` tools now surface the persisted truth-engine `quality` object
+  (`decisionEligible`, warning/block reason codes, `policyVersion`, `assessedAt`).
+  Markets with critical failures stay visible but cannot drive paper opens or
+  alerts.
+- **`openBlocked` preview on `pm_quote`.** Quotes preview the open-time quality
+  gate (`openBlocked` + `openBlockReasons`), so an agent can skip a market that
+  would 422 before burning the open attempt. The self-host runner
+  (`coinrithm-agent`) does this skip automatically.
+- **Independent forecasts in the runner.** The self-host agent runner elicits the
+  model's OWN probability (judged from the question/resolution criteria/deadline,
+  never anchored to the market price) and submits it as `forecastProbability` on
+  PM opens — feeding the public calibration dataset with proper-scoring-rule
+  forecasts. Clamped to [1,99]; omitted (never faked) when the model does not
+  produce one; `HOUSE_AGENT_FORECAST_ENABLED=false` disables.
+- **`crossPlatform` on event lists** documented in the API contract: sibling
+  venues pricing the same question, on list rows.
+- **ForecastEx venue truth.** Public MCP discovery copy and registry metadata
+  now describe all 11 live venues, including ForecastEx.
+- **Contract synchronization.** Runner templates and example bundles pin the
+  served OpenAPI 1.6.0 contract; canonical scorecard paths are unambiguous.
+
+## 0.7.2
+
+Docs-truth + privacy release. No tool behavior change, no API-surface change.
+
+- **Ten venues in the public listing.** `pm_data_*` tool copy, the README, and
+  `server.json` now name all ten venues (adds Futuur and Myriad). npm `0.7.1` was
+  published before those landed, so the registry listing still advertised "eight
+  venues"; npm versions are immutable, so correcting the public listing required
+  a new release.
+- **`source` parameter description** on `pm_data_events` / `pm_data_event_detail`
+  now enumerates all ten venue slugs. Accepted values are unchanged — this is
+  description text only, which is why it is a patch and not a minor.
+- **Privacy.** Raw model output is no longer persisted, enforcing the package's
+  no-chain-of-thought promise.
+- **New tripwire.** `server.json` (the MCP-registry listing) is now guarded
+  against version and venue-count drift; it had no guard, which is how it went
+  stale in the first place.
+- Refreshed stale Arena-gate example copy.
+
+## 0.7.1
+
+Docs + registry-metadata release; no tool behavior changes.
+
+- **README refresh**: the keyless `pm_data_*` data surface is now front and
+  center — 8 venues (Polymarket, Kalshi, Smarkets, Limitless, Manifold,
+  Metaculus, PredictIt, Rothera — the "seven venues" line predated Rothera),
+  the anonymous hosted-endpoint path, and the `referenceProbability` /
+  `volumeHistory` fields the data tools return.
+- **`server.json`**: hosted endpoint's `Authorization` header marked optional
+  (the `pm_data_*` tools work anonymously — verified live) and the server
+  description now leads with the keyless data surface.
+- Ships the post-0.7.0 commits: `pm_data_event` advertises `volumeHistory`,
+  `pm_data_events` advertises `referenceProbability` on list items, and the
+  hosted MCP root (`GET /`) serves a self-describing JSON landing (with a
+  405 + hint on `GET /mcp`).
+
+## 0.7.0
+
+(Retroactive entry — released 2026-07-05 without a changelog note.)
+
+- **Four keyless `pm_data_*` tools** — CoinRithm's free public cross-venue
+  prediction-market dataset over MCP, no API key required and yours is never
+  attached: `pm_data_overview` (market-wide stats), `pm_data_events`
+  (cross-venue event list), `pm_data_event` (detail incl.
+  `crossSourceMatches` + resolution evidence), `pm_data_whales`
+  (large-trade tape).
+
+## 0.5.0
+
+Agent-runner quality + reliability release. `coinrithm-agent` got materially
+smarter and less noisy; `coinrithm-mcp` is unchanged in shape. Bundles the work
+since 0.4.0.
+
+- **Prediction markets are a first-class venue** in the decide prompt: a short
+  `pmN` ref so small models trade PM reliably, eligible-outcome filtering (only
+  backend-openable outcomes reach the model), and futures-capped agents steered
+  to PM (a separate budget) instead of re-rejecting.
+- **PM anti-churn — now actually effective.** The candidate list is pre-filtered
+  to exclude markets the agent already holds, and the runner + server both block
+  re-betting a held market+outcome (no more one-agent, 25-identical-bets churn).
+  An earlier version read the wrong `/positions/pm` fields and was silently dead;
+  fixed.
+- **Settlement-feedback learning loop.** The agent sees how its own recent bets
+  actually resolved (win/loss/void + realized PnL) as reflective context, so it
+  learns from outcomes across cycles.
+- **Per-trade reasoning stays honest about the market.** A multi-action decision
+  no longer stamps its primary rationale onto a secondary trade about a different
+  market — the trade's public Arena "why" always matches the market it's on.
+- **Futures reliability.** The model is unblinded to per-position mark /
+  liquidation / stop / take-profit prices; take-profit is auto-clamped to a valid
+  R:R target off the stop (kills the `take_profit_not_*_mark` reject waves); and a
+  marking-down PM book now trips the equity-drawdown kill-switch too.
+- **News capability.** Recent high-importance news for the watchlist coins is fed
+  into the decide context as a market-catalyst layer.
+- **Robustness + contract accuracy.** Scheduler/runner hardening, flat-state
+  prompt steers (weak models stop hallucinating closes), manage-enum
+  normalization, and the PM contract now documents real entry friction rather
+  than a disclose-only stance.
+- **Security.** `hono` bumped to 4.12.27 (high-severity advisories: serve-static
+  path traversal, CORS wildcard-with-credentials, body-limit bypass).
+- **Docs.** The npm README leads with value / free / OKF / Studio; stale
+  scheduler and `minDecidedTrades` claims corrected.
+
+## 0.4.0
+
+- **Deterministic scorecard engine (`computeScorecard`).** The reproducible-
+  evaluation engine for `coinrithm.agent.scorecard.v1` — pure math over an
+  agent's realized track record (no network, no model): realized PnL, win rate,
+  expectancy, profit factor, reward-to-risk, Sharpe, Sortino, deflated /
+  probabilistic Sharpe (Bailey & López de Prado — skill vs luck with a multiple-
+  testing penalty), max drawdown, and Brier + ECE calibration for probabilistic
+  calls. Same inputs → identical metrics **and** a sha256 `contentHash` of the
+  canonicalized result, so a scorecard whose hash doesn't reproduce isn't
+  trusted. Metrics are computed AFTER the run from immutable evidence (leakage-
+  separation), so tuning-to-the-metric is structurally impossible. Returns
+  `null` for thin records — never a fabricated number.
+- **Resolver: committable file metadata + functionality pin.** The OKF resolver
+  now carries per-file metadata and pins functionality through resolution, so a
+  bundle's behavior is reproducible from its committed files.
+
+## 0.3.0
+
+- **Agent risk config: coin deny-list (`blocklist`).** `risk.blocklist` lets an
+  agent name symbols it must never open, even if they are on the watchlist —
+  deny wins over allow. Enforced in the runner's decision validator (rejects
+  `futures_open` / `spot_order` on a denied symbol) and surfaced in the system
+  prompt, so the model is told the boundary and the runner re-checks it.
+- **Docs: Open Knowledge Format positioning.** Clarified that a CoinRithm agent
+  is an OKF bundle — a portable directory of markdown + frontmatter that is
+  model-agnostic (run the same definition on any model). Develop and prove it
+  free on paper, then run it anywhere.
+- npm keywords refreshed (`open-knowledge-format`, `okf`, `model-agnostic`,
+  `gemini`) for registry discovery.
+
+## 0.2.0
+
+- Added the **`coinrithm-agent`** self-host runner binary alongside the MCP
+  server: a folder-as-architecture (OKF) agent you bring your own model key to,
+  with caps enforced by the runner (not the model), dry-run by default,
+  paper-only.
+
+## 0.1.x
+
+- Initial `coinrithm-mcp` MCP server: reads, quotes, scoped spot/futures/PM
+  writes, ledger export, and Agent Arena integration over a user-minted API key.