boomkin 0.13.0 → 0.13.2
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/CHANGELOG.md +15 -0
- package/README.md +6 -4
- package/catalog/providers.json +46 -81
- package/catalog/skills.json +39 -39
- package/catalog/workflows.json +2 -2
- package/dist/cli.js +518 -464
- package/dist/desk/desk.css +6 -6
- package/docs/AGENT-INDEX.md +1 -1
- package/docs/CHATGPT.md +1 -1
- package/docs/CONNECTIONS.md +4 -4
- package/docs/DESK.md +1 -1
- package/docs/GETTING-STARTED.md +1 -1
- package/docs/PROVIDERS.md +12 -67
- package/docs/UPDATES.md +1 -1
- package/docs/WORKFLOWS.md +63 -63
- package/llms.txt +1 -1
- package/package.json +4 -2
- package/docs/assets/boomkin-hero.svg +0 -14
- package/docs/research/assets/boomkin-comparison-desktop.png +0 -0
- package/docs/research/assets/boomkin-comparison-mobile.png +0 -0
- package/docs/research/minara-2026-10-02.md +0 -170
- package/docs/research/saved-capture-comparison-2026-10-10.md +0 -45
package/llms.txt
CHANGED
|
@@ -2,7 +2,7 @@
|
|
|
2
2
|
|
|
3
3
|
> Your open-source DeFi desk on native Hermes: reviewed skills, bounded research workflows, source evidence and local strategy tests.
|
|
4
4
|
|
|
5
|
-
The
|
|
5
|
+
The source-controlled catalogs own the current workflow, pack, skill and provider inventory. Provider discovery does not connect or install every listed service. CoinGecko is the sole onboarding default. Other harnesses receive portable skills only.
|
|
6
6
|
|
|
7
7
|
## Start here
|
|
8
8
|
|
package/package.json
CHANGED
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
{
|
|
2
2
|
"name": "boomkin",
|
|
3
|
-
"version": "0.13.
|
|
3
|
+
"version": "0.13.2",
|
|
4
4
|
"description": "Your open-source DeFi desk: native Hermes research projects, verifiable evidence and independent financial skills.",
|
|
5
5
|
"type": "module",
|
|
6
6
|
"license": "MIT",
|
|
@@ -8,8 +8,10 @@
|
|
|
8
8
|
"scripts": {
|
|
9
9
|
"boomkin": "bun src/cli.ts",
|
|
10
10
|
"test": "bun test test",
|
|
11
|
+
"test:gate": "bun test ./test/core.test.ts ./test/source.test.ts ./test/integrity.test.ts ./test/state.test.ts ./test/catalog-freshness.test.ts ./test/hermes.test.ts ./test/chatgpt.test.ts ./test/providers.test.ts ./test/mcp-discovery.test.ts ./test/onboarding.test.ts ./test/templates.test.ts ./test/workflows.test.ts ./test/desk.test.ts ./test/docs-code.test.ts ./test/distribution.test.ts ./test/site-copy.test.ts ./test/agent-onboarding.test.ts",
|
|
11
12
|
"typecheck": "tsc --noEmit",
|
|
12
|
-
"check": "bun run typecheck && bun run test",
|
|
13
|
+
"check": "bun run typecheck && bun run test:gate",
|
|
14
|
+
"gate": "bun run check && bun run build && bun run smoke && bun run build:site && bun run check:site",
|
|
13
15
|
"build": "bun scripts/build.ts",
|
|
14
16
|
"smoke": "bun scripts/smoke.ts",
|
|
15
17
|
"build:site": "bun scripts/build-site.ts",
|
|
@@ -1,14 +0,0 @@
|
|
|
1
|
-
<svg xmlns="http://www.w3.org/2000/svg" width="1200" height="440" viewBox="0 0 1200 440" role="img" aria-labelledby="title desc">
|
|
2
|
-
<title id="title">Boomkin — your open-source DeFi desk</title><desc id="desc">Research markets. Inspect positions. Test a thesis.</desc>
|
|
3
|
-
<rect width="1200" height="440" rx="12" fill="#111214"/>
|
|
4
|
-
<path d="M64 96h1072M64 350h1072" stroke="#36383c"/>
|
|
5
|
-
<g font-family="-apple-system, BlinkMacSystemFont, Segoe UI, sans-serif">
|
|
6
|
-
<text x="64" y="64" font-size="27" font-weight="600" letter-spacing="-1" fill="#f6f7f2">boomkin</text>
|
|
7
|
-
<rect x="966" y="39" width="170" height="32" rx="16" fill="#e7f06a"/>
|
|
8
|
-
<text x="1051" y="61" text-anchor="middle" font-size="17" font-weight="700" letter-spacing="1" fill="#111214">GALLEON LABS</text>
|
|
9
|
-
<text x="64" y="155" font-size="19" font-weight="600" letter-spacing="3" fill="#e7f06a">NATIVE HERMES</text>
|
|
10
|
-
<text x="64" y="234" font-size="68" font-weight="650" letter-spacing="-3" fill="#f6f7f2">Your DeFi desk.</text>
|
|
11
|
-
<text x="64" y="314" font-size="68" font-weight="650" letter-spacing="-3" fill="#e7f06a">Your agent. Your rules.</text>
|
|
12
|
-
<text x="64" y="400" font-size="24" fill="#c5c8c5">Independent skills. Evidence you can inspect.</text>
|
|
13
|
-
<text x="1136" y="400" text-anchor="end" font-size="17" font-weight="600" letter-spacing="2" fill="#e7f06a">RESEARCH / REVIEW / TEST</text>
|
|
14
|
-
</g></svg>
|
|
Binary file
|
|
Binary file
|
|
@@ -1,170 +0,0 @@
|
|
|
1
|
-
# Minara and the open-source Boomkin product
|
|
2
|
-
|
|
3
|
-
Research date: 2 October 2026. This is a source-based product assessment, not a live Minara account test, financial recommendation, performance comparison, or feature-parity claim.
|
|
4
|
-
|
|
5
|
-
## Assessment
|
|
6
|
-
|
|
7
|
-
Boomkin should make a useful DeFi research desk available in a user-owned agent environment: a direct install path, public data that produces a first result, protocol-specific jobs, and portable evidence that another person can inspect. That is a substantial product improvement. Calling a collection of procedures an equivalent to Minara's wallet, strategy engine, marketplace, or Autopilot would obscure the engineering still required.
|
|
8
|
-
|
|
9
|
-
Minara's most transferable strength is continuity between a question, cited observations, a reviewable decision, and a later check on its outcome. Its product also contains hosted infrastructure that is not supplied by its open-source skills or CLI. The research-to-monitoring loop is the best immediate target for Boomkin; managed execution and a strategy engine should be assessed as separate integrations.
|
|
10
|
-
|
|
11
|
-
## Method and source boundaries
|
|
12
|
-
|
|
13
|
-
Reviewed Minara's [documentation index](https://minara.ai/docs/llms.txt), homepage, product pages, 27 detailed documentation pages, open CLI, and open skills. Public Markdown URLs were inconsistent during this review; extensionless GitBook HTML pages were readable. No account was created, model queried, wallet connected, subscription purchased, order submitted, or provider credential accessed.
|
|
14
|
-
|
|
15
|
-
Public visual review found a sparse black-and-white opening with an oversized serif headline, a terminal-like background, and denser financial panels lower down. This report's product behavior findings come from published documentation and source; they do not establish that every marketed interaction works for a new account.
|
|
16
|
-
|
|
17
|
-
Pinned upstream source:
|
|
18
|
-
|
|
19
|
-
- [Minara CLI 25e64ae6](https://github.com/Minara-AI/minara-cli/tree/25e64ae6f14561dddabb51c1edbcd141866c4456), main revision returned by GitHub on this date.
|
|
20
|
-
- [Minara skills 8badf3bc](https://github.com/Minara-AI/minara-skills/tree/8badf3bcaff1592b79128dcea4ea164246909580), skill metadata version 3.0.4.
|
|
21
|
-
- Local baseline: [Boomkin 082a7a52](https://github.com/galleonlabs/boomkin/tree/082a7a521757e1d191248ed7692e5f83e32e07cd), version 0.7.3; [crypto-defi-skills 96c46d1e](https://github.com/galleonlabs/crypto-defi-skills/tree/96c46d1ef427456816b0268ffca8732bd9415559).
|
|
22
|
-
- Previous review: [5 September Minara review in crypto-defi-skills](https://github.com/galleonlabs/crypto-defi-skills/blob/96c46d1ef427456816b0268ffca8732bd9415559/docs/research/minara-review.md). Its findings about interactive recovery, competing account writers, recipient provenance, and uncertain outcomes remain relevant. This review broadens the assessment to the product.
|
|
23
|
-
|
|
24
|
-
The capability matrix below describes this baseline. Changes delivered after the baseline require their own checks and release evidence.
|
|
25
|
-
|
|
26
|
-
## What Minara actually combines
|
|
27
|
-
|
|
28
|
-
### 1. A short account-to-task path
|
|
29
|
-
|
|
30
|
-
Minara's documented onboarding is Google or email verification, then its chat interface. Spot and venue-specific perps wallets are created automatically. The user does not first choose a runtime, assemble skill packs, or configure an RPC. That explains much of the perceived completeness. [Account creation](https://minara.ai/docs/get-started/activate-your-account)
|
|
31
|
-
|
|
32
|
-
The wallet selector determines venue, balance, margin, and order context. Each perps wallet belongs to either Lighter or Hyperliquid, and the deposit guide requires USDC on Arbitrum. Transfers, on-ramp availability, and settlement have their own constraints. A unified interface does not make all assets, venues, and deposits interchangeable. [Deposit and wallet management](https://minara.ai/docs/get-started/deposit)
|
|
33
|
-
|
|
34
|
-
**Boomkin implication:** keep its isolated native Hermes profile and explicit provider choice, but shorten time to a useful public result. Offer a bounded task before any wallet connection, and say which additional inputs enable deeper jobs.
|
|
35
|
-
|
|
36
|
-
### 2. Research outputs with an appropriate level of effort
|
|
37
|
-
|
|
38
|
-
The documented chat modes are ordinary chat, Deep Research, and Data Visualization. Deep Research decomposes a question, queries several sources, cross-validates, and produces a structured cited report; visualization returns tables or charts. The docs describe downloadable PDF/Word reports and file inputs. These are output contracts, not merely longer prompts. [Research modes](https://minara.ai/docs/chat/deep-research)
|
|
39
|
-
|
|
40
|
-
The chat product page emphasizes exact-source links, timestamped report windows, source disagreements, exports, and the ability to make a recurring job from a question. Those claims are useful acceptance criteria for an open product, even where their hosted implementation has not been independently tested here. [Chat product](https://minara.ai/product/chat)
|
|
41
|
-
|
|
42
|
-
**Boomkin implication:** distinguish a quick observation from a research brief. A brief should retain machine-readable evidence, exact query parameters, identity, units, observation time, retrieval time, missing fields, assumptions, and links. A chart must be derived from the same recorded numbers.
|
|
43
|
-
|
|
44
|
-
### 3. Discovery that gives the user a question worth asking
|
|
45
|
-
|
|
46
|
-
The homepage leads with concrete tasks, then shows work steps and a result: market selection, a strategy comparison, or an account overview. Crypto Picks has a narrower documented purpose: short-horizon bullish/bearish watchlists, initial tracking price, current evidence, dynamic targets, and removal history. It is not fundamental valuation. [Homepage](https://minara.ai/home), [Crypto Picks](https://minara.ai/docs/chat/crypto-picks)
|
|
47
|
-
|
|
48
|
-
**Boomkin implication:** make the landing page task-led. Show a source-backed market brief, yield comparison, position inspection, and written thesis review. A synthetic product example must carry that label. A public sample containing current prices must carry its captured timestamp and sources.
|
|
49
|
-
|
|
50
|
-
### 4. A structured Copilot handoff
|
|
51
|
-
|
|
52
|
-
Copilot evaluates multi-timeframe price action, liquidity, funding, and open interest. Its documented result is long, short, or neutral with entry, take-profit, stop-loss, and rationale. The user reviews the plan and chooses whether to place the venue-specific order. The underlying exchanges supply pricing and liquidity; Minara presents the interface and execution layer. [Trading Copilot](https://minara.ai/docs/trade/trading-copilot)
|
|
53
|
-
|
|
54
|
-
Galleon's existing Hyperliquid ticket procedure is already more explicit about network, account mode, DEX, instrument IDs, precision, fees, margin, client order IDs, expiry, reduce-only protection, partial fills, and reconciliation. This should be presented as a concrete strength, without implying that installing the procedure establishes authenticated execution. [Baseline Hyperliquid plan](https://github.com/galleonlabs/crypto-defi-skills/blob/96c46d1ef427456816b0268ffca8732bd9415559/packages/hyperliquid/skills/hyperliquid-plan/SKILL.md)
|
|
55
|
-
|
|
56
|
-
### 5. A real strategy and execution stack
|
|
57
|
-
|
|
58
|
-
Strategy Studio uses Pine Runtime, Minara's TypeScript DSL. It has two distinct OHLC backtest models: time-series strategies make candle-based decisions and use lower-timeframe candles for certain intrabar triggers; cross-sectional strategies rank a universe and rebalance at candle close. The docs say fees, slippage, and perpetual funding are modeled, and identify liquidity, latency, rejection, leverage, overfitting, and cold-start bias as limits. [Strategy Studio](https://minara.ai/docs/trade/strategy-studio), [Backtesting FAQ](https://minara.ai/docs/support/faq/strategy-studio/backtesting)
|
|
59
|
-
|
|
60
|
-
Autopilot runs selected strategies within a specified asset scope and wallet. Its documented rules include canceling existing open orders on start, cross-margin eligibility, reserved risk for eligible existing positions, strategy-specific constraints, protection orders, drawdown stops, and handling manual overrides. These are runtime state and enforcement requirements, not a generic cron prompt. [Autopilot](https://minara.ai/docs/trade/trading-autopilot)
|
|
61
|
-
|
|
62
|
-
**Boomkin implication:** a strategy template may help draft or review an idea. It cannot honestly be marketed as a backtesting engine, paper trading, or autonomous execution without the corresponding executable integration and evidence.
|
|
63
|
-
|
|
64
|
-
### 6. Recurring workflows and retained reasoning
|
|
65
|
-
|
|
66
|
-
Minara's visual Workflow surface combines triggers, conditions, AI queries, notifications, trades, and generated code. Deployed workflows cannot be edited in place according to the current guide; credit exhaustion pauses them. Its visual builder is separate from the core agent loop. [Workflow](https://minara.ai/docs/workflow/agentic-workflow)
|
|
67
|
-
|
|
68
|
-
The newer harness docs describe recurring research more carefully: subject, sources, cadence, output, notification rule, observable failure, and quiet runs until useful change. Its thesis tracker records the asset, thesis, invalidation conditions, evidence standard, frequency, and notification rule; the first run uses cited evidence and distinguishes invalidation from ordinary volatility. The docs explicitly do not promise an immutable trade journal. [Automation boundaries](https://minara.ai/docs/agent/build-and-extend/hooks-cron-delegation), [Thesis tracker](https://minara.ai/docs/agent/start-here/use-cases/thesis-tracker)
|
|
69
|
-
|
|
70
|
-
**Boomkin implication:** use native Hermes scheduling and memory. Produce a reviewable monitoring specification before creating an actual job. Record successful scheduling and a successful first run separately. A saved template is not a running monitor.
|
|
71
|
-
|
|
72
|
-
## Architecture and the open-source boundary
|
|
73
|
-
|
|
74
|
-
Minara separates model reasoning from context assembly, tool access, memory, workspace execution, and authorization. Tools are exposed selectively, their results are observations, and memory cannot create financial permission. Its financial docs distinguish research, preview, submission, confirmation, fill, and account state. Those are sound architecture principles to preserve through Hermes. [Model and harness](https://minara.ai/docs/agent/start-here/model-and-harness), [Context and tools](https://minara.ai/docs/agent/core-concepts/context-and-tools), [Financial authorization](https://minara.ai/docs/agent/operate-safely/financial-authorization)
|
|
75
|
-
|
|
76
|
-
The public CLI is an API client. Its [request implementation](https://github.com/Minara-AI/minara-cli/blob/25e64ae6f14561dddabb51c1edbcd141866c4456/src/api/client.ts) uses a configured hosted API, and [chat](https://github.com/Minara-AI/minara-cli/blob/25e64ae6f14561dddabb51c1edbcd141866c4456/src/api/chat.ts) streams authenticated requests to the service. The open [Minara skill](https://github.com/Minara-AI/minara-skills/blob/8badf3bcaff1592b79128dcea4ea164246909580/skills/minara/SKILL.md) routes tasks through that CLI. Source availability for the client does not establish source availability or self-hostability for the data layer, signing service, strategy engine, or marketplace.
|
|
77
|
-
|
|
78
|
-
Minara calls its wallets non-custodial while explicitly stating that it generates and operates the controlling key on the user's behalf using AWS KMS. It describes EIP-7702 delegation, stablecoin gas payment, revocation, and key export. This review does not assess the legal custody classification or independently audit those claims. The technical distinction relevant to Boomkin is that Minara operates a managed signing service; Boomkin's setup does not. [Wallet security](https://minara.ai/docs/technology/wallet-security)
|
|
79
|
-
|
|
80
|
-
The tool-stack docs describe licensed/private data alongside public providers. Some pages claim more than 40 providers and others more than 50. That does not imply that an open replacement can redistribute every feed. Open code, public data, freely available data, and provider access are separate properties. [Tool stack](https://minara.ai/docs/technology/tools-integration), [About](https://minara.ai/docs)
|
|
81
|
-
|
|
82
|
-
DMind has public model artifacts, but provider documentation is not a reason to pin an untested local model. The model-family guide and Hugging Face display differ on some parameter counts. Evaluate the specific artifact's license, chat template, tool calling, requirements, and benchmark behavior before presenting it as a supported model. Hermes's existing model choice remains the appropriate boundary. [Model family](https://minara.ai/docs/technology/dmind/model-family), [Published DMind artifacts](https://huggingface.co/DMindAI)
|
|
83
|
-
|
|
84
|
-
## Capability matrix
|
|
85
|
-
|
|
86
|
-
Status is scoped to the baseline above. “Procedure” means installable guidance and possibly a local helper; it does not mean a connected or enforced service.
|
|
87
|
-
|
|
88
|
-
| Capability | Minara documented product | Boomkin / Galleon baseline | Immediate honest improvement |
|
|
89
|
-
| --- | --- | --- | --- |
|
|
90
|
-
| Installation and model setup | Hosted login; open CLI/skill adapters | Native Hermes onboarding, reviewed runtime, isolated profile | Simpler install representation and first task |
|
|
91
|
-
| Market data | Curated public, premium, private feeds | Keyless CoinGecko MCP; public diagnostic; optional providers | Bounded research command with captured evidence |
|
|
92
|
-
| Token identity | Chain/contract/venue context | Dedicated CoinGecko procedure and source methodology | Carry identity into every exported brief |
|
|
93
|
-
| Yield research | Discovery and pool data | DefiLlama yield screen with base/reward distinction and local script | Clean shortlist with explicit exit-diligence gaps |
|
|
94
|
-
| Protocol position work | Broad DeFi tools | Specific Aave, Morpho, Compound, Uniswap, Aerodrome, Lido, Pendle, ERC-4626 jobs | Explain each job's inputs and concrete result |
|
|
95
|
-
| Portfolio analysis | Unified wallet/account surface | Reconciliation procedure; connected reads required | Portable report with coverage and double-count controls |
|
|
96
|
-
| Perps research and planning | Copilot on Lighter/Hyperliquid | Hyperliquid analysis and exact unsigned tickets; generic derivatives procedure | Public market context and explicit handoff |
|
|
97
|
-
| Wallet and financial execution | Managed keys, wallets, gas, authenticated venues | Optional official tools; setup creates no wallet or authority | Expose prerequisites and verification stages |
|
|
98
|
-
| Deep research artifacts | Cited reports, charts, download/share | Evidence guidance; native Hermes can create artifacts | Reproducible brief and machine-readable evidence bundle |
|
|
99
|
-
| Recurring research | Visual workflow and harness schedule | Native Hermes scheduling available; no Boomkin scheduler | Monitoring contract and verified native handoff |
|
|
100
|
-
| Trading memory | Decision context, selective recall | Native Hermes memory; no guaranteed complete trading journal | Written thesis and evidence delta procedure |
|
|
101
|
-
| Backtesting and paper trading | Two historical engines, strategy DSL, paper surface | No equivalent engine in baseline | Clearly scoped upstream engine assessment |
|
|
102
|
-
| Autopilot | Strategy runtime, orders, risk controls, account state | No equivalent managed service | Do not claim it from instructions or cron |
|
|
103
|
-
| Strategy marketplace | Run others' strategies with tracked records | Skill-pack catalog, not a strategy marketplace | Portable reviewed workflows; distinct terminology |
|
|
104
|
-
| Prediction markets | Polymarket research and Hyperliquid outcome surfaces | No dedicated baseline prediction-market procedure | Optional public research using official tooling |
|
|
105
|
-
| Developer API | Hosted paid-key/x402 intelligence and payload endpoints | Independent npm packs and CLI helpers | Stable local artifact contracts; no hosted API claim |
|
|
106
|
-
| Behavioral evaluation | Published benchmark claim | Offline acceptance/routing checks | Measured model runs with fixtures and reproducible grading |
|
|
107
|
-
|
|
108
|
-
## Priorities with observable acceptance criteria
|
|
109
|
-
|
|
110
|
-
### P0: Make public research a working entry point
|
|
111
|
-
|
|
112
|
-
A fresh user should be able to get a bounded, keyless result without funding a wallet. Market and yield requests need explicit inputs, bounded results, source time, retrieval time, units, quality/freshness policy, missing fields, and clear rate-limit errors. Save raw or normalized evidence beside a concise Markdown brief. Treat outages and stale data as visible gaps.
|
|
113
|
-
|
|
114
|
-
**Acceptance:** isolated-install test; representative public-provider read; malformed, empty, stale, and rate-limited fixtures; artifacts agree numerically; success never silently changes the provider or turns missing data into zero.
|
|
115
|
-
|
|
116
|
-
### P0: Ship the public website as a product guide
|
|
117
|
-
|
|
118
|
-
Lead with one clear promise, one install action, and a representative work product. Use a structured research example to teach evidence and next steps, then show actual jobs and independent packs. Give documentation a deliberate hierarchy: start, research, jobs, connections, artifacts, updates, contribution. Source content from the catalog where possible so counts and identifiers do not drift.
|
|
119
|
-
|
|
120
|
-
**Acceptance:** mobile and desktop render checks; working install-copy control; keyboard navigation; no overflow; readable contrast; functioning links; current catalog counts; public demo clearly labeled; deployment revision and serving page verified.
|
|
121
|
-
|
|
122
|
-
### P1: Make research and thesis monitoring portable procedures
|
|
123
|
-
|
|
124
|
-
A research brief needs a decision, horizon, universe, source plan, observations, contrary evidence, uncertainty, and the smallest next check. A thesis tracker needs written invalidation criteria, evidence quality, cadence, saved state location, and a meaningful-change notification rule. Provider failure is a failed check, not an invalidated thesis.
|
|
125
|
-
|
|
126
|
-
**Acceptance:** scenarios for stale sources, shared upstreams, conflicting observations, ordinary volatility, true invalidation, missing data, and a recurrence that makes no unauthorized trade or notification. Validate the pinned Hermes scheduler contract before writing exact invocation instructions.
|
|
127
|
-
|
|
128
|
-
### P1: Publish evaluation evidence that describes what was tested
|
|
129
|
-
|
|
130
|
-
The existing corpus checks validate packaging, references, routing fixtures, and deterministic helpers. They should not become a marketing model score. Add agent runs only with explicit model/runtime revisions, fixed fixtures, recorded outputs, decision-level grading, and honest scope.
|
|
131
|
-
|
|
132
|
-
**Acceptance:** an independent reader can rerun the cases and distinguish structural tests, model behavior, public-provider smoke, authenticated capabilities, and financial settlement.
|
|
133
|
-
|
|
134
|
-
### P2: Choose optional maintained engines for the true runtime gaps
|
|
135
|
-
|
|
136
|
-
Backtesting and automated strategies deserve an upstream integration decision. Evaluate a maintained engine against historical coverage, costs, account semantics, precision, execution recovery, license, and reproducible paper behavior. Avoid adding a homemade engine merely to decorate the website.
|
|
137
|
-
|
|
138
|
-
**Acceptance:** a pinned executable contract, safe backtest/paper run, independent artifact review, and clear engine/venue limits before making product claims.
|
|
139
|
-
|
|
140
|
-
## Maintained upstream candidates
|
|
141
|
-
|
|
142
|
-
| Need | Primary source | Fit and limitation |
|
|
143
|
-
| --- | --- | --- |
|
|
144
|
-
| Runtime, models, memory, MCP, recurrence | [Hermes Agent](https://github.com/NousResearch/hermes-agent), [native scheduling](https://hermes-agent.nousresearch.com/docs/user-guide/features/cron/) | Already Boomkin's runtime. Current docs can exceed the reviewed pin; inspect the installed contract. |
|
|
145
|
-
| Keyless token and DEX research | [CoinGecko official MCP](https://docs.coingecko.com/ai-integration/mcp-server) | Free shared-rate endpoint; full tools and higher limits can require a plan. Counts and access should be discovered live. |
|
|
146
|
-
| Public yield data and methodology | [DefiLlama yield server](https://github.com/DefiLlama/yield-server) | Open adapters and yield definitions; data is screening input, not a net-return or withdrawal guarantee. |
|
|
147
|
-
| Venue data and execution | [Hyperliquid developer API](https://hyperliquid.gitbook.io/hyperliquid-docs/for-developers/api) | Reuse official API/SDK; venue/account-mode and tool authority still require explicit support. |
|
|
148
|
-
| Prediction-market research | [Polymarket official agent skills](https://github.com/Polymarket/agent-skills/blob/main/market-data.md), [official docs](https://docs.polymarket.com/) | Official read paths for discovery, market metadata, order books, and positions. Access to data does not authorize trading or remove market restrictions. |
|
|
149
|
-
| Strategy research and backtests | [Freqtrade backtesting](https://docs.freqtrade.io/en/stable/backtesting/) | Maintained engine with documented simulation assumptions. Not automatic compatibility with Hyperliquid account behavior or arbitrary DeFi positions. |
|
|
150
|
-
| Modular automated strategies | [Hummingbot strategies](https://hummingbot.org/strategies/), [controllers](https://hummingbot.org/strategies/v2-strategies/controllers/) | Separates market-data providers, controllers, and executors. Connector, deployment, and paper/backtest fit require assessment. |
|
|
151
|
-
| Optional finance-specialized local models | [DMind model artifacts](https://huggingface.co/DMindAI) | Public weights are available; individual artifact license, device fit, and real tool behavior require separate validation. |
|
|
152
|
-
|
|
153
|
-
These are candidates and source references, not newly installed dependencies or supported Boomkin integrations.
|
|
154
|
-
|
|
155
|
-
## Presentation rules
|
|
156
|
-
|
|
157
|
-
- Present concrete jobs and outputs before pack inventory.
|
|
158
|
-
- Demonstrate the question, observed work, evidence, and next decision in one coherent example.
|
|
159
|
-
- Put identity, time, source, and risk beside the metric they qualify.
|
|
160
|
-
- Use a restrained financial-workspace visual language; do not reuse Minara's assets, copy, reported returns, customer logos, or testimonials.
|
|
161
|
-
- Label illustrative data as illustrative. Never use simulated profit, speed, “always on,” or provider counts as proof.
|
|
162
|
-
- Separate runtime support, installable procedures, connected read access, authenticated execution, and terminal settlement.
|
|
163
|
-
- Explain optional provider costs and model requirements at the point they matter.
|
|
164
|
-
- Maintain an honest capability boundary on the site. A local open-source product can be compelling on its own merits.
|
|
165
|
-
|
|
166
|
-
## Decisions this research supports
|
|
167
|
-
|
|
168
|
-
Use Hermes for the operating layer. Keep the independent pack architecture. Add a useful public research path and reusable evidence artifacts. Make the website demonstrate those outcomes. Add a portable thesis/research procedure where it improves actual task quality. Treat strategy engines, hosted services, managed signing, and execution enforcement as separate engineering contracts.
|
|
169
|
-
|
|
170
|
-
The reference bar is the clarity and completeness of Minara's user journey. The open-source implementation should earn its own capability claims through reproducible results.
|
|
@@ -1,45 +0,0 @@
|
|
|
1
|
-
# Saved capture comparison: decision record
|
|
2
|
-
|
|
3
|
-
Research date: 2026-10-10. Baseline: `60fdb2f` on `galleonlabs/boomkin` main. No open pull requests or issues were listed at selection time. This is a draft implementation for Andrew's review; no deployment or runtime activation is part of this work.
|
|
4
|
-
|
|
5
|
-
## Product evidence and choice
|
|
6
|
-
|
|
7
|
-
Boomkin 0.11.0 already delivers three local public-data jobs, wallet accounting, frozen strategy validation, templates, exports and native Hermes follow-ups. PR #60 added the local desk. Reimplementing the wallet audit or another catalog/template expansion would duplicate recent work. The current desk retains independent captures but makes the human switch between whole reports to work out what changed. Its follow-up API also chooses the latest completed run even when the human is viewing an older run.
|
|
8
|
-
|
|
9
|
-
Selected upgrade: a saved-capture comparison with an explicit earlier baseline, a portable structured export, and native follow-ups tied to the displayed capture. It adds value to the existing refresh/history loop without requiring an account, paid API, new protocol adapter or scheduler. It preserves a small human interface and documents the exact machine contract.
|
|
10
|
-
|
|
11
|
-
Other opportunities considered: clickable evidence inspection would improve reading but does not answer the returning user's main question; automatic monitoring would add scheduling, provider and delivery authority beyond this task; broader asset/thesis intelligence would require new research sources and clear provider cost choices. These remain outside this PR.
|
|
12
|
-
|
|
13
|
-
## Current primary sources
|
|
14
|
-
|
|
15
|
-
- [Senpi](https://senpi.ai/) presents wallet-focused quant-desk jobs, trading scores and cost-leak investigation. Its public product surface supports a task-first interface, not copying its trading execution or marketing performance claims. Boomkin already has the wallet entry point; comparing subsequent captures is the missing continuation.
|
|
16
|
-
- [Elfa provenance](https://www.elfa.ai/provenance) describes timestamped, source-traceable market context. [Auto capabilities](https://docs.elfa.ai/auto/capabilities/) explicitly identifies state across evaluations, source freshness and replay/deduplication as part of monitoring. Inference for Boomkin: establish a reliable manual comparison contract before adding any monitoring integration. No Elfa service is connected in this change.
|
|
17
|
-
- [CoinGecko Simple Price](https://docs.coingecko.com/reference/simple-price) documents exact IDs and `include_last_updated_at` for checking freshness. A second HTTP capture can contain the same cached price observation. Comparison therefore matches provider and identity and checks source timestamps rather than interpreting retrieval time as a fresh price move.
|
|
18
|
-
- [Hyperliquid Info endpoint](https://hyperliquid.gitbook.io/hyperliquid-docs/for-developers/api/info-endpoint) documents at most 2,000 fills per time-query response and access to only the 10,000 most recent fills. It distinguishes spot and HIP-3 fills. A refresh of the same frozen historical window can change retained coverage; a changed activity subtotal is not portfolio profit earned since the last capture.
|
|
19
|
-
- The source-controlled [strategy engine provenance](../../site/vendor/provenance.json), [desk adapter](../../src/desk.ts) and [project evidence contract](../PROJECTS.md) are authoritative for the installed behavior. The worker inspected public helper files at the existing catalog's exact immutable revisions. No skills version, dependency contract or other repository is changed.
|
|
20
|
-
|
|
21
|
-
Competitor pages are first-party descriptions, not independent verification of their capabilities. The implementation decision rests on Boomkin's captured code behavior and the provider contracts above.
|
|
22
|
-
|
|
23
|
-
## Comparison boundaries
|
|
24
|
-
|
|
25
|
-
Comparisons use two completed public captures in the same project, oldest first, with unchanged saved inputs and helper revision. Each new derived result has a saved SHA-256 digest; existing captures remain readable and exportable, but captures from before that digest contract cannot produce numeric comparisons. Both source manifests, frozen briefs and derived result bytes must pass local consistency checks. These checks do not prove source truth or economic validity.
|
|
26
|
-
|
|
27
|
-
Market marks are matched by provider and exact identity. Missing/stale observations stay unavailable; cached or backward timestamps do not yield movement claims. Wallet decimals retain signed exact arithmetic and partial-coverage caveats. Wallet history totals compare the same requested window, whereas gross exposure describes sequential current-state reads. Strategy comparisons preserve period boundaries, frozen specification and data/engine hashes; shifted data or methodology withholds numeric deltas instead of implying an improvement in the rule.
|
|
28
|
-
|
|
29
|
-
The UI defaults to the nearest earlier completed capture and lets the user choose another. Source additions, missing artifacts and changed bytes are separate from economic change. There is no model call or external fetch in comparison. An explicit follow-up sends the selected run ID and preserves parent project/run identity; older API callers that omit a run ID retain latest-completed behavior.
|
|
30
|
-
|
|
31
|
-
## Validation plan
|
|
32
|
-
|
|
33
|
-
Use fixture captures to cover complete/partial data, signed decimals, cached/reversed timestamps, changed periods/hashes, missing legacy digests, changed sources/results, wrong run identity and exact-run follow-up provenance. Run repository-native typecheck/tests/build, static documentation checks, and desktop/mobile browser inspection against a disposable local fixture profile. An independent reviewer inspects the actual diff before a draft PR is created. Record actual results in the PR and the workspace delivery artifact.
|
|
34
|
-
|
|
35
|
-
## Review and rendered evidence (2026-10-11)
|
|
36
|
-
|
|
37
|
-
An independent agent reviewed the actual implementation diff and the exact pinned public helper contracts. It found two substantive issues: corrupt manifest shapes could throw instead of returning an unavailable comparison; the wallet helper's zero defaults for entirely failed history reads could become misleading deltas. Both were fixed with API and unit regressions. A follow-up review also corrected the unavailable-state source disclosure. Final independent review reported no remaining high-confidence findings, with 37 focused tests and 512 assertions passing, plus typecheck and whitespace checks. This is agent review evidence, not a GitHub approval.
|
|
38
|
-
|
|
39
|
-
Browser checks used a disposable loopback fixture profile, synthetic market/wallet responses and the already-vendored dated Bitcoin history. No model, credential, transaction or live market claim was involved. Verified desktop (1440px) and mobile (390px), all three comparison jobs, baseline switching, first-capture empty state, partial provider values, comparison export, retry after a simulated network error, and navigation while a comparison response was delayed. Mobile document width remained 390px and the comparison table fit its 346px content width, including the change column.
|
|
40
|
-
|
|
41
|
-

|
|
42
|
-
|
|
43
|
-

|
|
44
|
-
|
|
45
|
-
Full test/build/CI results are recorded in the draft PR. The shared Mac's measured load average exceeded 300 during validation; the default five-second test limit timed out unrelated baseline tests. Extended-timeout runs are reported explicitly rather than represented as a default-check pass. No runtime was activated or release deployed.
|