- /api/ai500 returned a success/data envelope but the web httpClient
wraps the raw body as data, so the panel read coins as undefined and
showed the empty state; return a flat {coins,count} body like
/api/symbols
- the agent rendered AI500 rankings as a markdown table that the chat
UI flattens into an unreadable line: tool note + system prompt now
mandate numbered lists (one coin per line) and ban tables outright
- provider/nofxos: GetAI500ListCached — 5min TTL cache with stale
fallback on upstream failure; ResolveClient routes through the claw402
gateway when a wallet key is configured (user's claw402 model key ->
CLAW402_WALLET_KEY env -> direct nofxos)
- new GET /api/ai500 endpoint serving the score-sorted board
- new get_ai500_list agent tool + prompt rule: when the user wants coin
picks or creates a strategy without naming coins, consult AI500's
high-scoring entries by default
- web: AI500 sidebar panel (rank, score badge, gain since entry,
5min auto-refresh); clicking an entry asks the agent to analyze it
- system prompt (zh+en): finish-the-work-then-report rule — chain tools in
the current turn, never promise background work; prefill industry-standard
defaults instead of interrogating field by field, batch any unavoidable
questions into one
- manage_strategy tool description: create only needs name, omitted config
merges from the default template, one-shot summary + confirmed=true flow
- manage_trader tool description: resolve exchange/model/strategy bindings
via list tools instead of asking the user for IDs
The agent felt like an artificial idiot because the LLM almost never spoke
for itself: 14+ Go paths injected fmt.Sprintf canned replies, the frontend
filtered out tool-progress events so users saw three dots for 10-20s, the
main prompt told the LLM "be a trading partner" AND "answer only what's
asked", and the planner sliced the toolset by inferred domain so a "BTC
dropped, how much am I losing?" question couldn't see positions and market
at the same time.
- agent/central_brain.go: shouldTrustDeterministicSkillReply now always
returns false. Successful mutations (trader/strategy/model/exchange
create/update/start/stop/delete) flow through reviewTaskCompletion so the
LLM sees the real outcome JSON and writes the user-facing prose. The
trade-confirmation regex path (handleTradeConfirmation) was already
outside this code path and is unaffected.
- agent/agent.go: rewrite the Behavior section of the main system prompt.
Replace the contradictory "answer only what's asked / don't upsell" with
"lead with the direct answer, then optionally one relevant follow-up
only when (a) open risk, (b) missing config, or (c) the next step is
obvious — e.g. created, want me to start it?". Explicitly authorize
chaining ("if the user says create and start, do both this turn") and
ban "please wait / I'll get back to you" language because there is no
background job to come back from.
- agent/tools.go: plannerToolsForText always returns the full 22-tool set
(new __all__ domain). The old per-domain trimming hid manage_trader from
market questions and execute_trade from anything that didn't look like
an explicit trade — cross-domain reasoning was structurally blocked. The
compact-vs-full strategy schema switch is preserved so mutation intents
still see the full config schema.
- web/src/components/agent/{AgentStepPanel,ChatMessages}.tsx: stop
filtering tool: steps. Map raw tool names to friendly labels with emoji
("get_positions" → "📊 检查持仓") in zh/en/id. Users now see what the
agent is doing in real time instead of silence. central_brain routing
chatter still gets dropped.
- agent/planner_tools_test.go: tests updated to assert the new
full-toolset behavior and the compact-vs-full strategy schema switch.
Root cause: when planner fails (402 payment), legacy loop dumps all system
context to LLM which outputs everything. Also final response prompt was too
weak — LLM treated all observations as required output.
Changes:
- Strengthen system prompt: "answer ONLY what user asked", no tables/tutorials
unless requested, no self-intro repeats, no "next step" suggestions
- Add compact observation summary for final response (step summaries only,
no raw JSON blobs)
- Domain-filtered tool selection in legacy loop to prevent over-fetching
- Fix domain routing: "钱包/wallet" → account domain (not model), with
exchange configs included for wallet context
- Widen wallet fast-path: no longer requires "claw402" keyword
- Anti-repetition instructions in planner step selector
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Agent now uses mcp.NewAIClientByProvider() for claw402 provider, ensuring
x402 payment signing works correctly instead of generic HTTP client
- Added ReasoningContent field to Message/LLMResponse structs and wired
serialization/parsing so DeepSeek thinking models work in multi-turn
- Added Beta badge to Agent nav tab in HeaderBar
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Add nil checks for session.TargetRef in all four execute*Action
handlers (Trader/Exchange/Model/Strategy) to prevent panic on
corrupted sessions; bulk-delete and query actions are excluded
- Add ensureHistory() helper and call it in runPlannedAgentWithContextMode
to prevent nil panic when history is not initialized
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
