Releases: CoinRithm/coinrithm-agent-trading
Release list
MCP 0.7.14 · TypeScript SDK 0.3.3 · Python SDK 1.8.3
MCP 0.7.14 · TypeScript SDK 0.3.3 · Python SDK 1.8.3
Published to npm and PyPI on 24 September 2026. All four registry downloads match the reviewed SHA-256 manifest; npm integrity also matches. The attached archives are the exact published files.
Changes
- Add an optional prediction-market entry probability floor, checked before fees, with matching TypeScript and Python request fields.
- Reject conflicting futures stop/target changes before sending them to the API.
- Scope permanent model failures to the provider/model that actually failed. Clear unattributed legacy streaks and reset availability failures on successful responses. Mixed fallback rate limits preserve correct failure attribution and metering.
- Add compiled agent definitions, optional local definition checks, strategy sections, candle timing evidence and explicit-zero policy handling.
The MCP npm package includes both the MCP server and the coinrithm-agent runner. API contract version remains 1.7.0. Paper trading only.
Validation
MCP: 1,322 tests. Scheduler: 258 local tests plus CI PostgreSQL integration. Python SDK: 1,422 tests. SDK generation, builds, lint, formatting and compatibility checks passed. Independent review verified the package contents and hashes.
Clean installs of the registry downloads passed the existing Node package, MCP startup and two restart lifecycle checks. Python 1.8.3 imported all 407 modules, passed synchronous/asynchronous mock checks and round-tripped the optional entry-floor field at 0, 20 and 100. These checks use local fixtures without live trading or provider calls.
Delivery
Release source: c374e78. The manifest records the tested source, reviewed archive build and registry verification.
Scheduler deployment 2672 and hosted MCP deployment 2673 run this source. Both had zero restarts at verification. MCP advertises 0.7.14 with 40 tools; health, anonymous market reads and missing-key rejection passed. Casa completed a normal scheduled cycle after deployment with both model-error counters at zero; its definition, model and key were preserved.
Package publication does not restart hosted services. The source README and publishing guide record MCP Registry and documentation delivery separately.
MCP 0.7.13 · TypeScript SDK 0.3.2 · Python SDK 1.8.2
Published package delivery
Published on npm/PyPI on 23 September 2026: MCP/agent runner 0.7.13, TypeScript SDK 0.3.2 and Python SDK 1.8.2. All four registry downloads match the attached reviewed archives byte for byte, and clean installed-package checks passed.
MCP server and agent runner 0.7.13
- Add public whale-wallet rankings and source/wallet movement detail to the 40-tool surface. Identify wallets only when the source provides identity.
- Clarify market-price calibration: complete-book observations 20–28 hours before resolution, event-weighted ECE, and separate
finalPrice/ownCapturetiming lanes. - Preserve venue terms, missing/zero candle-volume coverage, and signed cumulative funding in compact data and agent observations.
- Deliver provider-capacity classification fixes and the supported
classifyProviderFailureengine export. - One npm package includes both
coinrithm-mcpandcoinrithm-agent.
TypeScript 0.3.2 and Python 1.8.2
- Add typed whale-wallet endpoints and the house-only Arena
status=openview. Pending rows retain null score/settlement fields and zero realized PnL; they are not settled-performance evidence. - Preserve optional thesis/advisory fields, venue terms, cumulative funding and candle-volume coverage without inventing missing values.
- Correct spot-cancellation response types and Python error parsing for the existing API behavior.
Documentation and evidence
- API reference follows the current contract. Runnable examples are pinned to TypeScript SDK 0.3.2 and Python SDK 1.8.2. Their existing fixture tests passed in CI, and the updated reference is live.
- Changelog and publishing procedure separate source, hosted service, npm/PyPI, GitHub and MCP Registry delivery.
- Release source:
726d8f0466d0cfd02ab5a9d46aed6ca12a45ff06. Build/review source:9ea286cacceb4d0a0da31c71f4f0b726e8c8b262; their Git trees are identical. - All 26 CI checks passed on the reviewed source. Pages deployment passed.
- Exact attached archives and registry-downloaded files pass content/metadata checks, Twine validation and clean installed Windows Node/Python smoke checks, including both binaries and 40 MCP tools. Three independent agent review lanes cleared the documentation and procedure; MCP and SDK archives were independently inspected.
release-manifest.jsonrecords SHA-256 for all four archives, npm integrity, commits and verified registry-download evidence. The published files match the reviewed archives without repacking.
Hosted MCP already reports 0.7.13 at 4b39cd057765f0ab995d0ec4db2cfde665fa4bdd; scheduler runs 184201069cbf888989870e71243ba725fc98b634. Their deployment records are separate. This documentation/package preparation caused no service restart.
Verified versions: MCP and agent runner 0.7.13, TypeScript SDK 0.3.2, and Python SDK 1.8.2. The official MCP Registry independently lists 0.7.13 as active/latest, verified after the exact-tag publication workflow.
Publication-status docs and SDK-example pins were finalized in 1a27f95cc5326e9064aeb9980fc545e0716793fe (PR #33). All 26 reviewed-revision CI checks passed; Pages deployed that exact documentation revision. The release tag and published archives remain on the reviewed package source above.
MCP 0.7.9 and SDK updates - runner reliability
Published packages
This release fixes runner retry timing, request deadlines and periodic prediction-market budget enforcement, and adds mandatory coverage and PostgreSQL checks. It brings the already-published packages together in one GitHub release; the API contract remains 1.7.0.
| Package | Version | Install |
|---|---|---|
| MCP server and agent runner | 0.7.9 | npm install --save-exact @coinrithm/mcp-trading@0.7.9 |
| TypeScript SDK | 0.3.1 | npm install --save-exact @coinrithm/sdk@0.3.1 |
| Python SDK | 1.8.1 | python -m pip install coinrithm-sdk==1.8.1 |
Highlights
- PM hourly budget: periodic evaluations respect the common runner budget. Existing position-management and explicit always-on exemptions remain.
- Bounded API requests: one 30-second deadline covers headers, response bodies and 429 retry waits, with optional cancellation. Uncertain writes are not automatically replayed.
- Correct retry timing: missing or malformed
Retry-Afteruses the existing fallback; explicit zero and valid dates remain supported. - Faithful agent state: atomic state-file replacement, preserved trigger/sizing settings on eject, and completed-action memory that excludes rejected or unexecuted proposals.
- Clearer market data: compact MCP responses preserve quote units, normalized probabilities and settlement provenance. SDK documentation explains that candle volume is a rolling 24-hour observation, not interval turnover.
- Better diagnostics: bounded private decision-input evidence, direct NVIDIA 5xx retry handling, and reason-specific hosted scheduler admission telemetry with existing limits unchanged.
Verification
All five jobs passed in release CI #269.
| Package | Passing tests | Lines | Branches |
|---|---|---|---|
| MCP/runner | 1,164 | 96.63% | 91.79% |
| Scheduler | 211 | 97.89% | 91.56% |
| TypeScript SDK | 4 | 100% | 100% |
| Python SDK | 1,292 | 96.58% | 94.84% |
The scheduler total includes all 11 mandatory PostgreSQL integration tests. Package coverage gates enforce 90% floors, with additional checks for selected critical files. The TypeScript SDK is a small executable wrapper; generated type declarations are not executable coverage. Scope, audit results and reproduction commands.
Registry downloads matched the reviewed archives byte for byte. Fresh npm and PyPI installations passed offline client and compiled runner checks. Hosted MCP and scheduler deployments were verified separately at the package source revision. These checks do not establish trading returns or improved production failure rates.
Source and artifacts
The four attached package archives are the exact artifacts published to npm/PyPI. They were built from d052a7bb7ce791623f4e80e72748b75d50ea6b83; SHA256SUMS covers each archive.
The GitHub tag mcp-trading-v0.7.9 includes subsequent documentation and the community README update in PR #8, with runtime code and package versions unchanged. GitHub's automatically generated source archives represent that tag. Historical hosted records should be matched to their recorded revision, not assumed to match a later package.
Community thanks
- Tibor, for detailed feedback shared by email on runtime behavior, observation semantics, decision evidence and reproducibility, helping clarify the documentation and evidence limits.
- Nils Friedrichs (@nils-friedrichs), for PR #8, replacing the broken discovery badge. The README retains the working Smithery listing alongside LightNow.
Paper trading only: all trades use virtual funds.
MCP 0.7.12 — Clearer tool effects and retry guidance
Clearer MCP tool effects and retry guidance
This patch improves the three definitions highlighted by Glama's expanded tool feedback:
whoamiexplains authentication, identity/permission fields and usage metadata, without unrelated trading-cost prose.cancel_spot_orderexplains cancellation of the unfilled remainder and the ambiguousalreadyClosedresult. Its annotation now advertises the existing idempotent cancellation behavior.report_pm_opportunityexplicitly describes a durable, self-reported evidence write. Its annotation no longer claims unconditional idempotency: duplicate protection requiresdecisionId, or theagentTrace.decisionIdfallback, under the same API key. The first stored record wins.
All 38 tool names, accepted input values and execution behavior are unchanged. The TypeScript SDK remains 0.3.1, Python SDK 1.8.1, and the API contract 1.7.0. Runner and scheduler logic are unchanged. All trading uses paper funds.
Verification
- CI for source
394b3b3: all 25 jobs passed. MCP: 1,215 tests, 96.68% line coverage, 92.21% branch coverage; dependency audit reported zero known vulnerabilities. - Hosted MCP deployment 2505 finished on
394b3b39cd13e31256504ad3704817808c7e97ad. All 52 compiled MCP/runner JavaScript files match the attached CI archive. The public endpoint reports 0.7.12 and serves the corrected definitions; a bounded public-data read returned HTTP 200. Rollback is retained. - These checks establish interface and delivery behavior, not trading returns or a new Glama score.
npm delivery verified: @coinrithm/mcp-trading@0.7.12 is published with the latest tag. Its registry download is byte-for-byte identical to the attached CI archive. A clean Windows/Node 24 install passed stdio initialization and 38-tool discovery with the corrected descriptions and retry annotations; all 52 compiled JavaScript files match. No tool handler, provider call or trade was invoked during that install check.
Archive
coinrithm-mcp-trading-0.7.12.tgz — SHA-256:
bfbc0301b402f23c07ddd27de139e333dc2fcb84157c57743560a4d39f3a98c2
MCP 0.7.11 — accurate opportunity reporting
MCP 0.7.11: accurate opportunity-report outcomes
Failed opportunity reports previously logged success because the reporter did
not inspect resolved API results. This patch confirms a report only when the
API returns ok: true.
Changes
- HTTP errors remain unconfirmed; timeout, cancellation, network failure and
exceptions leave delivery unknown. CycleResult.opportunitycontains confirmed reports only. The additive
opportunityReportfield retains the attempted payload, outcome and status.- Reporting keeps its existing one-method-invocation-per-cycle latch and adds
no retries. Error bodies and exception details are excluded from diagnostics.
Verification
PR #12 adds
regressions for resolved 503/422 failures, status-0 transport failures and thrown
exceptions. Skip and act tests compare other result fields and runner state
with successful-report controls and verify identical execution behavior.
All 25 CI jobs passed
for the attached archive's source tree: 1,212 MCP/runner tests, 96.68% line and
92.21% branch coverage. The reporter has 100% coverage on all four metrics.
The archive passed the existing 12 Node/OS compatibility combinations.
The TypeScript SDK remains 0.3.1, Python SDK 1.8.1, and API contract
1.7.0. MCP still exposes 38 tools. Trading uses virtual funds.
Install and archive
npm install --save-exact @coinrithm/mcp-trading@0.7.113f3dd869e9436fda7c15d864fc123f5f071c4e85f984840e02b35b3454481a2f coinrithm-mcp-trading-0.7.11.tgz
Source commit: 7e8ebe8b75ffb328b34316afc5ef01501895b848.
Hosted MCP deployment 2502 and scheduler deployment 2503 finished on
that source. All 52 compiled MCP/runner JavaScript files in each running image
match the attached archive. Health and an anonymous public MCP read passed;
prior images are retained for rollback. All 58 agent configuration fingerprints
are unchanged.
Controlled probes of the compiled modules in each container confirmed success,
503/422 errors, a real client timeout with a simulated transport, and an
exception. They made no provider calls or trades. No opportunity-report events
were observed in the bounded scheduler-log sample from 14:37:59 to 14:39:54 UTC.
npm publication is verified: the registry download matches the SHA-256 above.
A clean registry installation on Windows/Node 24.11.0 passed startup, engine
imports, locking, persistence, 38-tool discovery and the compiled reporting
success/error/timeout/exception probes. All 52 installed compiled JavaScript
files match the archive. These checks made no provider calls or trades.
MCP 0.7.10 — runner recovery and compatibility
MCP 0.7.10: runner recovery and compatibility
This patch preserves a file-backed runner's identity before its first possible
write, so a process crash can recover the same intent key. It also fixes the
Node 18 startup failure found by the new installed-package compatibility tests.
Changes
- Supported
@coinrithm/mcp-trading/engineimport, with existing deep imports
preserved. Observation reconciliation and opportunity reporting now have
separate ownership in the runner. - Optional crypto entry predicates reject otherwise valid entries when the
configured return condition is false or its evidence is missing/stale.
Existing agents are not automatically opted in. validateandinspectwarn about inactive abstention settings and reserved
web search capability.- Hosted scheduler migrations use a bounded database transaction and advisory
lock. An offline credential-rotation tool has interruption and recovery tests. - Required pull requests and six CI checks protect
main. Actions use pinned
commit SHAs; downloaded release tools are pinned and checksum-verified.
Verification
All 25 CI jobs passed
for source 9dc6e6ed470b47222d97e370fc04427fc83e0a92.
| Scope | Passing tests | Lines | Branches |
|---|---|---|---|
| MCP and runner | 1,201 | 96.67% | 92.19% |
| Scheduler | 222, including 13 PostgreSQL integration tests | 98.02% | 91.71% |
Installed-package checks cover Linux, Windows and macOS on Node 18, 20, 22 and
24. Python wheel checks cover Linux 3.10–3.14 and Windows/macOS 3.12. Separate
processes and a loopback execution fixture exercise response loss, process
termination, restart, state recovery and duplicate-intent rejection.
These are controlled checks. Recovery still depends on the execution service
honoring idempotency keys; this is not an exactly-once or profitability guarantee.
Delivery
Hosted MCP deployment 2499 and scheduler deployment 2501 finished on
the source above. Both running images' 52 compiled MCP/runner JavaScript files
match the release archive. Health and public MCP checks passed; prior images
are retained for rollback. No production credential rotation was performed.
The TypeScript SDK remains 0.3.1, Python SDK 1.8.1, and API contract
1.7.0. Their runtime source is unchanged. MCP still exposes 38 tools;
trading uses virtual funds.
npm install --save-exact @coinrithm/mcp-trading@0.7.10The attached archive is the one exercised by the successful compatibility CI.
SHA-256:
710071aee8272d7dce802c1033d67f1c914796b64652cc1489dbd373de64dcc4 coinrithm-mcp-trading-0.7.10.tgz
Changes ·
Runner guide ·
Scheduler rotation runbook
The npm registry archive was downloaded and matched the SHA-256 above. A clean
registry installation passed startup, tool discovery, persistence, both restart
scenarios and compiled deadline/retry/budget checks on Windows, Node 24.11.0.
These delivery checks are separate from CI and the hosted checks.