feature/agent-memory-skills-and-unified-model

[cardosofede]

What changed

1. Memory & Skills for AI agents (FEAT-001 → FEAT-003)

• User memory (condor/memory/store.py): persistent file-per-fact store, exposed only via MCP (manage_memory), injected as [USER MEMORY] into /agent chat + trading-agent ticks.
  manage_notes is now a deprecated alias. New /memory command to review/prune.
• Skills (condor/memory/skills.py): markdown playbooks (when-to-apply + steps) retrievable by relevance, can reference routines. New manage_skill MCP tool, injected as [SKILLS].
• Per-assistant stores (FEAT-003): memory + skills are now scoped per assistant (condor chat and each trading agent get their own store) instead of shared per-user. Assistants moved to
  folder form (assistants/condor/AGENT.md); runtime stores gitignored.

2. Unified Agent model (FEAT-004)

• Merged "experts" and "strategies" into one Agent (condor/trading_agent/agent.py). An Agent is consultable (one-shot) and owns strategies that loop. experts.py and routes/experts.py
  removed; endpoints folded into routes/agents.py (/agents/{slug}/consult, /agents/{slug}/strategies/...). Migration script included.

3. Frontend — Hyperliquid wallet connect

• New ConnectHyperliquid flow with EVM/Hyperliquid wallet libs (lib/wallet/), agent+builder-code signup, referral prompt when wallet unfunded, unified balance dedupe.

4. Quality wave (improvements backlog)

• Security: WS auth via subprotocol (not query string), iframe sandboxing for reports, HTML escaping in chart tooltips, masked credential fields, cross-server WS leak fix, encoded URLs.
• Perf: incremental candle updates, memoized chart/executor rows, WS subscription instead of polling, parallel snapshot fetches.
• Refactor: extracted hooks (useBacktest, useResizeDrag), shared ExecutorTable/lib/backtest/lib/formatters, deduped helpers; trimmed dead agent UI components.

5. Infra: pydantic-ai client overhaul, auto-generate JWT secret if absent, recursive AGENT.md file watcher, local-LLM (LM Studio) improvements, new test suites (test_agents,
   test_memory_store, test_skill_store).

---

QA recommendations

Migrations (run first, highest risk)

• Run scripts/migrate_to_agent_model.py and the per-assistant store migration on a copy of prod data. Verify legacy data/notes/*.json and old shared stores migrate correctly and the
  legacy store is archived, not deleted.

Memory & Skills

• In /agent, confirm the agent writes/edits/deletes a memory fact, and that /memory lists and prunes it. Verify edits made by one assistant are not visible to another assistant
  (isolation is the core FEAT-003 guarantee).
• Create a skill via manage_skill, confirm it's injected into the next tick and the agent follows it; test a playbook that references a routine.

Unified Agent / trading agents

• Consult an agent (/agents/{slug}/consult) and run a strategy loop — both paths share the same Agent. Confirm no broken imports (routes/experts.py removal was the original trigger).
  Verify existing brigado agent still ticks after migration.

Hyperliquid connect

• Connect a wallet end-to-end; check funded vs unfunded wallet (referral signup prompt should appear), agent-name editing (16-char cap), and that unified Hyperliquid balances don't
  double-count.

Security regressions (verify the fixes hold)

• WS connects with token in subprotocol (not URL); subscribe to a server you lack access to → must be rejected. Confirm report iframes are sandboxed and chart tooltips escape HTML.

Smoke / regression

• Run uv run pytest (new + existing suites). Exercise dashboard pages touched by the refactor: Executors (new ExecutorTable), Backtesting (extracted hook), Portfolio, Reports, Agent
  detail — watch for empty tables or chart render regressions.
• Confirm bot boots with no WEB_JWT_SECRET set (auto-generation) and AGENT.md edits hot-reload.

[rapcmia]

Commit 36ae26a

• Rebuild docker image and use make setup; make deploy

Retest c1e34c0 dashboard rendering only where output formatting changed:

• Consult answer markdown rendering ✅
  
• Use delegation on agent consult page ✅
  
• I test condor/consult on Gate.io BTC-USDT Spot Test Maker returned a delegate-style response for a different agent ❌
  

  ✅ Tool: condor/consult (16s)
  
  I’m checking the most likely cause of that mismatch, because a consult should not normally come back with a background-delegation message unless the consulted agent itself decided to hand work off or crossed wires with another agent.Most likely, the consult did run, but the consulted Gate.io agent itself responded by calling delegate internally.
  
  That would explain both symptoms:
  - you used consult
  - but the returned text looked like a delegation confirmation
  - and it mentioned the wrong agent, KuCoin SOL-USDT, which points to bad cross-routing or stale instructions/context inside the Gate.io agent
  
  So the mismatch is probably not “consult turned into delegate” at the platform level. It’s more likely:
  - consult invoked the Gate.io agent correctly
  - the Gate.io agent then chose the wrong action internally
  - it delegated to another agent and surfaced that delegation message back to you
  
  Why the KuCoin name appeared:
  - cross-agent routing mistake in the Gate.io agent’s instructions
  - stale context or copied logic referencing the KuCoin agent
  - a too-broad “delegate long-running tasks” behavior inside that agent
  
  Short version:
  - outer tool: consult
  - inner behavior: mistaken delegate
  - wrong target: likely agent misconfiguration
  

• The agent chat on webui supports delegate ✅
  
  
• Check consult rendering for invalid tool-response errors ✅

  #### Consult page payload
  Use connector fakeexchange for BTC-USDT, fetch the live order book, and return the raw tool error exactly as received.
  

  

---

Retest FEAT-007 shutdown behavior end to end in progress

• End-to-end validation of the Stop and Shutdown workflows completed successfully
• Confirmed Shutdown successfully stopped a live bot, cleared active orders, and kept the spot balance intact.
• Verified Shutdown follows a different flow than the regular Stop command.
• Todo: Retest drawdown thresholds e.g max drawdowns, etc. ❗

---

Retest server pinning behavior introduced in 35d1099 and surfaced in 36ae26a ✅

• The request-side consult guard correctly rejected an unauthorized server selection. When a limited user called consult with server_name: local, the request returned 403 No access. ✅
• create agent with pinned server ✅
• I consulted the pinned-server probe agent after explicitly including the agent name, and it used the pinned server correctly.
• Started a strategy under that same agent, and it also used the pinned server.

  pinned_server_probe | Tick #1
  Agent: pinned_server_probe.pinned_server_loop_2
  Status: Pinned-server verification loop active. No trading.
  Configured pinned server: visible via agent config (runtime server not directly exposed in tool context).
  Action: Hold — no executors created.
  

[rapcmia]

Commit 99a4752

• Setup on source + HAPI (docker)

Retest agents routes and strategy lifecycle, especially /agents/delegations behavior ✅

• Retest on 99a4752 for the earlier route-access setup gap after Condor source was started.
• Logged in through the Condor web-token flow and confirmed auth worked.
• /api/v1/agents/delegations returned the delegations list successfully
• Fake-task checks on the detail and stop routes returned delegation-specific 404 responses, which confirms these routes are not being recognized by the slug

    curl -s http://localhost:8088/api/v1/agents/delegations/test-task \
      -H "Authorization: Bearer XXX" | jq
  
    {
      "detail": "Delegation 'test-task' not found"
    }
  
    curl -s -X POST http://localhost:8088/api/v1/agents/delegations/test-task/stop \
      -H "Authorization: Bearer XXX" | jq
  
    {
      "detail": "Delegation 'test-task' not found"
    }
  

---

Re-test bots page parity between REST and websocket updates, including stopping overlays and controller PnL data. ✅

• Opened the bots page with a live bot/controller already running and saw the bots status, controllers status, and PNL as soon as the page loaded
• Left the page open and watched websocket updates come through without refreshing while i triggered pause/resume, a filled order event and stop
• The stopping overlay appear while shutdown was in progress and disappear after the bot fully stopped and archived

---

Re-test chat permission isolation with two users. ✅

• User A was the admin user and User B was only approved as a normal user, so this retest stayed focused on the condor/consult path and not on broader trading-agent loop behavior.
• User B could still access Condor chat, but when User B tried the direct consult flow the request returned 403 instead of reusing User A's server-backed access or agent context.
• This confirms the two-user isolation on the consult path: approved chat access for User B did not turn into inherited admin consult permission.

---

Re-test risk gating for repeated executor creates in one tick and tracker-failure handling. ✅

| Field | Value |
|---|---|
| Action | BUY LIMIT_MAKER placed |
| Price | $61,583.26 |
| Amount | 0.000116 BTC (~$7.16) |
| Executor ID | `2tWszhjNigpU...` |
| SELL side | Skipped (1-executor limit) |
| Rebalance | Not needed ($6.93 deviation) |

**Note:** The 1-executor open limit prevents two-sided quoting simultaneously. BUY was prioritized since USDT is slightly over-weight (57.4%). On the next tick, if this executor fills or expires, the SELL side can be placed to rebalance.

• Gate.io BTC-USDT retest on gateio_btc_usdt_spot_test_maker / gateio_btc_usdt_spot_test_loop session 12 allowed the first create and then stopped the second create path in the same tick.
• Condor route evidence matched the report: session 12 journal recorded Placed BUY LIMIT_MAKER 0.000116 BTC @ $61,583.26 with executor 2tWszhjNigpURhu4YdPjRex4LtLn5HhfVAMjVBARmYCp, and snapshot #1 showed SELL side | Skipped (1-executor limit).
• Since the max_executor is set to 1, it only placed sell order

---

Manually validate the new Market Making Expert consult/delegate flow. ✅

• Test API route

  curl -s http://localhost:8088/api/v1/agents/market_making_expert/consult \
      -H "Authorization: Bearer XXX" \
      -H "Content-Type: application/json" \
      -d '
    {
      "task": "QA validation only: explain in 2 short bullets when Market Making Expert should be consulted vs delegated. Do not deploy anything.",
      "context": "Testing the new consult/delegate flow for Condor source.",
      "chat_id": 0
    }' | jq
  
  {
    "agent": "market_making_expert",
    "answer": "- **Consult** when you need advisory output inline — regime assessment, spread calibration, inventory review, or config tuning — without touching any live bots.\n- **Delegate** when you want a full bot deployment run
    autonomously in the background — the agent reads `pmm_mister_deploy` and executes the end-to-end flow (regime → config → deploy) without mid-flow user confirmation."
  }
  
  ---
  
  curl -s http://localhost:8088/api/v1/agents/market_making_expert/delegate \
    -H "Authorization: Bearer XXX" \
    -H "Content-Type: application/json" \
    -d '
  {
    "task": "QA validation only: return a short background confirmation message for delegate flow. Do not deploy, trade, or edit files.",
    "chat_id": 0,
    "timeout_s": 120
  }' | jq
  
  {
    "task_id": "market_making_expert-delegate-e08d80b6",
    "status": "running"
  }
  
  ---
  
  curl -s http://localhost:8088/api/v1/agents/delegations/market_making_expert-delegate-e08d80b6 \
    -H "Authorization: Bearer XXX" | jq
  
  {
    "task_id": "market_making_expert-delegate-e08d80b6",
    "agent": "market_making_expert",
    "user_id": 1474408604,
    "chat_id": 0,
    "server_name": null,
    "task": "QA validation only: return a short background confirmation message for delegate flow. Do not deploy, trade, or edit files.",
    "status": "done",
    "result": "Background delegate confirmed. Market Making Expert agent is active and ready. No deployment, trading, or file changes initiated — standing by for task assignment.",
    "error": ""
  }
  

• Test using telegram
  
  

[rapcmia]

• Memory separation worked across Condor chat, trading-agent flows, and Routine Builder style routine access.
• The unified agent migration worked: Telegram trading-agent loops still ran, and Condor consult reached the correct unified agent path.
• Trading-agent session launch now carries the correct user identity through CONDOR_USER_ID.
• Condor web login worked correctly, including one-time token use, redirect to the right page, and token invalidation after reuse/logout.
• Protected Condor routes enforced auth correctly, and authenticated agent/delegation routes resolved as expected.
• Access isolation worked: a normal approved user could not reuse an admin user's server-backed consult access.
• server_required=true behavior improved: consult was blocked for no-server users, and launch was also stopped in practice when required checks could not complete.
• Normal consult still worked for the matching non-server-required setup.
• Bots page data stayed consistent between REST and websocket updates, including status, PnL, and stopping overlays.
• Delegation flow worked end to end: delegation routes resolved correctly, background tasks completed, and results could be fetched separately.
• Consult and delegation responses rendered clearly in the UI, including markdown tables, headings, bullets, and readable error output.
• The new Market Making Expert flow worked as intended: consult returned inline advice, while delegate ran in the background and finished separately.
• Risk gating blocked repeated executor creation in the same tick after the first approved create.
• FEAT-007 shutdown behavior worked end to end and stayed distinct from normal Stop.
• Dry-run startup no longer showed the earlier obvious journal-write failure.
• Strategy detail correctly showed saved Dry run counts.
• The live hummingbot-api gateway docs no longer showed a passphrase field.
• The Routines page worked in both grid and table views, including search, sorting, and linked report navigation.
• Server pinning worked correctly at runtime once the pinned agent/strategy path was tested directly.