@compose-market/sdk 0.6.6 → 0.6.8

package/.speakeasy/a2a.arazzo.yaml

@@ -1,50 +1,66 @@
 arazzo: 1.0.1
 info:
-  title: Compose A2A x402 Workflow
-  summary: Minimal agent-to-agent x402 loop for paid Compose agent calls, trailing receipts, and feedback.
-  version: 0.6.6
+  title: Compose A2A — minimal agent-to-agent loop
+  summary: |
+    Three-step recipe to call any Compose agent from another agent over x402 v2.
+    Step 1 dials blind, step 2 retries with a Permit2 signature against the
+    server's `upto` quote, step 3 submits structured feedback. Authoritative
+    metering decides the actual settled amount inside the envelope.
+  version: 0.6.8
 sourceDescriptions:
-  - name: agentic
-    url: ../specs/agentic.openapi.yaml
+  - name: manowar
+    url: ../specs/manowar.openapi.yaml
     type: openapi
   - name: x402
     url: ../specs/x402.openapi.yaml
     type: openapi
 x-speakeasy-test-server:
   baseUrl: "x-env: COMPOSE_API_TEST_URL; https://api.compose.market"
+
 workflows:
-  - workflowId: a2a_agent_x402_upto_feedback
-    summary: Authorize one bounded agent turn with x402 upto, stream the answer, read the trailing receipt, then submit traceable feedback.
+  - workflowId: a2a_agent_x402_upto
+    summary: |
+      Discover -> sign -> stream -> feedback. The contacting agent never
+      pre-commits a budget; the server returns a finite `upto` envelope in the
+      402, the contacting agent signs Permit2, and the actual settled amount
+      is metered server-side and capped by the envelope.
     inputs:
       type: object
+      required: [agentWallet, userAddress, chainId, message, paymentSignature]
       properties:
         agentWallet:
           type: string
+          description: The deployed Compose agent's wallet address (path param).
           examples: ["x-env: COMPOSE_TEST_AGENT_WALLET"]
         userAddress:
           type: string
+          description: The contacting agent's EVM wallet address.
           examples: ["x-env: COMPOSE_TEST_USER_ADDRESS"]
         chainId:
           type: integer
+          description: CAIP-2 chain id for x402 settlement (eip155:<chainId>).
           examples: [43113]
-        x402MaxAmountWei:
+        message:
           type: string
-          examples: ["1000000"]
+          description: User/agent prompt to send.
+          examples: ["What can you do? Answer briefly."]
         threadId:
           type: string
-          examples: ["a2a-x402-thread"]
-        composeRunId:
-          type: string
-          examples: ["a2a-x402-run"]
-        message:
-          type: string
-          examples: ["Use memory/tools if useful, answer briefly, then ask for feedback."]
+          description: Stable conversation identifier; reuse across turns to keep agent memory continuity.
+          examples: ["a2a-thread-1"]
         paymentSignature:
           type: string
+          description: |
+            Base64 PaymentPayload (x402 v2). The SDK helpers in
+            `@compose-market/sdk` produce this from a Permit2 sig over the
+            `accepts[0]` returned in step 1's PAYMENT-REQUIRED header.
           examples: ["x-env: COMPOSE_TEST_PAYMENT_SIGNATURE"]
+
     steps:
-      - stepId: challenge
-        operationId: $sourceDescriptions.agentic.agents_stream_create
+      # ── Step 1 — Blind dial. Server replies 402 with the `upto` envelope and
+      # the `compose-payment-intent-v1` extension endpoints. No client state.
+      - stepId: discover
+        operationId: $sourceDescriptions.manowar.agents_stream_create
         parameters:
           - name: walletAddress
             in: path
@@ -55,22 +71,23 @@ workflows:
           - name: x-chain-id
             in: header
             value: "$inputs.chainId"
-          - name: x-x402-max-amount-wei
-            in: header
-            value: "$inputs.x402MaxAmountWei"
         requestBody:
           contentType: application/json
           payload:
             message: "$inputs.message"
             threadId: "$inputs.threadId"
-            composeRunId: "$inputs.composeRunId"
             userAddress: "$inputs.userAddress"
         successCriteria:
           - condition: $statusCode == 402
         outputs:
           paymentRequired: $response.header.PAYMENT-REQUIRED
+
+      # ── Step 2 — Repeat the call with the Permit2 signature attached.
+      # Server runs inference, streams SSE, settles the metered amount
+      # (always <= the signed envelope), and writes a `compose.receipt`
+      # SSE event before the stream ends.
       - stepId: paid_stream
-        operationId: $sourceDescriptions.agentic.agents_stream_create
+        operationId: $sourceDescriptions.manowar.agents_stream_create
         parameters:
           - name: walletAddress
             in: path
@@ -81,9 +98,6 @@ workflows:
           - name: x-chain-id
             in: header
             value: "$inputs.chainId"
-          - name: x-x402-max-amount-wei
-            in: header
-            value: "$inputs.x402MaxAmountWei"
           - name: PAYMENT-SIGNATURE
             in: header
             value: "$inputs.paymentSignature"
@@ -92,12 +106,15 @@ workflows:
           payload:
             message: "$inputs.message"
             threadId: "$inputs.threadId"
-            composeRunId: "$inputs.composeRunId"
             userAddress: "$inputs.userAddress"
         successCriteria:
           - condition: $statusCode == 200
         outputs:
           composeRunId: $response.header.x-compose-run-id
+
+      # ── Step 3 — Optional but recommended: leave structured feedback.
+      # This is what trains the public reputation graph and lets your agent
+      # be discoverable by future contacting agents.
       - stepId: feedback
         operationId: $sourceDescriptions.x402.feedback_submit
         parameters:
@@ -115,21 +132,13 @@ workflows:
               id: "$inputs.agentWallet"
             rating: 5
             category: integration
-            message: "A2A x402 paid stream completed."
+            message: "Worked first try."
             context:
-              requestId: "$inputs.composeRunId"
-              composeRunId: "$inputs.composeRunId"
+              composeRunId: $steps.paid_stream.outputs.composeRunId
               chainId: "$inputs.chainId"
-              agentId: "$inputs.agentWallet"
               agentWallet: "$inputs.agentWallet"
               endpoint:
                 method: POST
                 path: "/agent/{agentWallet}/stream"
-            metadata:
-              threadId: "$inputs.threadId"
-              paymentRequiredHeader: $steps.challenge.outputs.paymentRequired
-              stream:
-                answerCompleteOn: done
-                receiptSource: compose.receipt
         successCriteria:
           - condition: $statusCode == 201

package/.speakeasy/memory.arazzo.yaml

@@ -1,42 +1,123 @@
 arazzo: 1.0.1
 info:
-  title: Compose Memory Agent-First Workflows
-  summary: Standalone workflow contracts for compact retrieval, turn recording, durable recall, and maintenance.
-  version: 0.6.6
+  title: Compose Memory — agent-first contracts
+  version: 0.6.8
+  summary: |
+    You are an agent. You have memory. Three calls cover 99% of turns.
+
+    PRE-TURN (always):  POST /api/memory/context/assemble
+      → returns { prompt: string|null, items, contextUsage }
+      → inject `prompt` as a system message if non-null. Done.
+
+    POST-TURN (always): POST /api/memory/turns/record  (fire-and-forget)
+      → persists transcript + working buffer + per-turn vector
+      → triggers async fact extraction (graph layer)
+      → idempotent on `turnId`. Pass it.
+
+    REMEMBER (rare):    POST /api/memory/remember
+      → only when the user states a stable fact AND the auto-extractor
+        in record_turn missed it. 99% of facts are auto-extracted.
+
+    Scope contract — memorise these 3 rules and you cannot fail:
+      1. (agentWallet, userAddress) is the durability boundary.
+      2. threadId narrows working+scene only. graph/vectors are
+         cross-thread — recall a fact from any past session.
+      3. mode="local" + haiId partitions a side-tree (HAI-bound). Omit both
+         for default global agent memory.
+
+    Layers (you do NOT pick them; ranker picks for you):
+      working   = rolling thread context        (Mongo sessions)
+      scene     = full transcripts              (Mongo session_transcripts)
+      graph     = durable user facts            (Mongo memory, source:fact)
+      patterns  = mined tool sequences          (Mongo patterns)
+      archives  = cold storage                  (Mongo archives + Pinata)
+      vectors   = hybrid recall                 (Mongo memory, source:session)
+
+    One ranker (Cloudflare BAAI bge-reranker-base + temporal decay + MMR)
+    shortlists across all 6 layers into ≤ 900 chars / ≤ 6 items per turn.
+
+    Token / latency budget defaults (override via params):
+      pre_turn block:  ≤ 900 chars, ≤ 6 items, p50 < 200ms
+      post_turn:       fire-and-forget, p50 < 50ms server-side
+      fact extraction: ~$0.000075/turn (gemini-3.1-flash-lite-preview)
+      voyage embed:    ~$0.00002/turn (voyage-4-large 1024-d)
+      cf rerank:       ~$0.0001/turn (skipped when top-1 dominates)
+
+    Storage summary: Mongo Atlas + Voyage embeddings + Cloudflare reranker.
+    No mem0. No vendor cloud. First-party from end to end.
+
 sourceDescriptions:
   - name: memory
     url: ../specs/memory.openapi.yaml
     type: openapi
+
 x-speakeasy-test-server:
   baseUrl: "x-env: COMPOSE_RUNTIME_TEST_URL; https://runtime.compose.market"
+
 workflows:
-  - workflowId: memory_agent_turn_loop
-    summary: Retrieve compact memory before a turn, persist the turn, then save an explicit durable fact.
+  # ──────────────────────────────────────────────────────────────────
+  # agent_turn_loop — RUN THIS ON EVERY TURN.
+  # ──────────────────────────────────────────────────────────────────
+  - workflowId: agent_turn_loop
+    summary: pre_turn → (agent's model call) → post_turn → optional remember
+    description: |
+      Three steps. Hit them in order.
+        1. pre_turn_context  → get prompt block; inject if non-null.
+        2. post_turn_record  → persist; fire-and-forget; idempotent on turnId.
+        3. remember_fact     → optional; only on explicit user fact.
+
+      Do NOT block your response on step 2 or 3.
     inputs:
       type: object
+      required: [agentWallet, userAddress, threadId, query, userMessage, assistantMessage, modelUsed, totalTokens, turnId]
       properties:
         agentWallet:
           type: string
+          description: agent's on-chain wallet. durability key.
           examples: ["x-env: COMPOSE_TEST_AGENT_WALLET"]
         userAddress:
           type: string
+          description: end-user wallet. durability key.
           examples: ["x-env: COMPOSE_TEST_USER_ADDRESS"]
         threadId:
           type: string
-          examples: ["arazzo-thread"]
+          description: conversation id. narrows working+scene only.
+          examples: ["arazzo-thread-1"]
         query:
           type: string
-          examples: ["billing preference"]
+          description: user's latest message verbatim. used as recall query.
+          examples: ["what did I tell you about my project?"]
+        userMessage:
+          type: string
+          description: the user side of this turn (often == query).
+          examples: ["I work on AI agents and prefer concise answers."]
+        assistantMessage:
+          type: string
+          description: your reply for this turn.
+          examples: ["Got it — concise, AI agents focus."]
+        modelUsed:
+          type: string
+          description: model id. used for receipts + cost telemetry.
+          examples: ["gpt-4o"]
+        totalTokens:
+          type: integer
+          description: input + output + reasoning tokens for this turn.
+          examples: [1209]
+        turnId:
+          type: string
+          description: |
+            stable id for THIS turn. dedup key for record_turn retries.
+            generate once per turn (e.g. crypto.randomUUID()) and reuse.
+          examples: ["turn_a1b2c3"]
+
     steps:
-      - stepId: get_loop_manifest
-        operationId: $sourceDescriptions.memory.workflows_get
-        parameters:
-          - name: workflowId
-            in: path
-            value: agent_memory_loop
-        successCriteria:
-          - condition: $statusCode == 200
       - stepId: pre_turn_context
+        description: |
+          Hits /api/memory/context/assemble. Returns:
+            prompt: string|null            — inject as system msg verbatim
+            items: AgentMemoryCompactItem[] — for telemetry/debug
+            totals: Record<layer, count>   — pre-shortlist counts
+            contextUsage: { characters, savedCharactersVsRaw, items }
         operationId: $sourceDescriptions.memory.context_assemble
         requestBody:
           contentType: application/json
@@ -45,14 +126,22 @@ workflows:
             userAddress: "$inputs.userAddress"
             threadId: "$inputs.threadId"
             query: "$inputs.query"
-            layers: ["working", "patterns", "vectors"]
-            budget:
-              maxCharacters: 1800
+            limit: 8
+            maxItems: 6
+            maxItemChars: 360
+            budget: { maxCharacters: 900 }
         successCriteria:
           - condition: $statusCode == 200
         outputs:
           contextId: $response.body#/contextId
+          prompt: $response.body#/prompt
+
       - stepId: post_turn_record
+        description: |
+          Hits /api/memory/turns/record. Returns:
+            stored: { transcript, working, vector, graph } — assert all 4 true
+            turnId, sessionId, vectorId
+          Idempotent on turnId. Fire-and-forget on the agent side.
         operationId: $sourceDescriptions.memory.turn_record
         requestBody:
           contentType: application/json
@@ -61,13 +150,20 @@ workflows:
             userAddress: "$inputs.userAddress"
             threadId: "$inputs.threadId"
             contextId: $steps.pre_turn_context.outputs.contextId
-            userMessage: "$inputs.query"
-            assistantMessage: "Recorded."
-            modelUsed: "compose-test"
-            totalTokens: 1
+            turnId: "$inputs.turnId"
+            userMessage: "$inputs.userMessage"
+            assistantMessage: "$inputs.assistantMessage"
+            modelUsed: "$inputs.modelUsed"
+            totalTokens: "$inputs.totalTokens"
         successCriteria:
           - condition: $statusCode == 200
+
       - stepId: remember_fact
+        description: |
+          OPTIONAL. Only call when the user volunteers a stable fact AND
+          you suspect the auto-extractor missed it. Use `type` to
+          categorise so the ranker surfaces the fact correctly later.
+          types: preference|identity|context|skill|relationship|event|other
         operationId: $sourceDescriptions.memory.remember
         requestBody:
           contentType: application/json
@@ -75,56 +171,76 @@ workflows:
             agentWallet: "$inputs.agentWallet"
             userAddress: "$inputs.userAddress"
             threadId: "$inputs.threadId"
-            content: "The user prefers concise billing answers."
+            content: "The user prefers concise answers and works on AI agents."
             type: preference
-            retention: long
-            confidence: 0.9
+            confidence: 0.95
         successCriteria:
           - condition: $statusCode == 200
-  - workflowId: memory_maintenance_loop
-    summary: Discover memory workflows, run maintenance, and inspect reusable learned patterns.
+
+  # ──────────────────────────────────────────────────────────────────
+  # eval_loop — measure recall@K. Use to gate deploys.
+  # ──────────────────────────────────────────────────────────────────
+  - workflowId: eval_loop
+    summary: Run a memory eval and assert recall@K ≥ 0.95.
+    description: |
+      Pass {query, expected} test cases. Server returns recallAtK,
+      precisionAtK, avgContextCharacters. Default layer set covers all 6
+      so you measure end-to-end ranker behaviour.
     inputs:
       type: object
+      required: [agentWallet, userAddress]
       properties:
-        agentWallet:
-          type: string
-          examples: ["x-env: COMPOSE_TEST_AGENT_WALLET"]
+        agentWallet: { type: string, examples: ["x-env: COMPOSE_TEST_AGENT_WALLET"] }
+        userAddress: { type: string, examples: ["x-env: COMPOSE_TEST_USER_ADDRESS"] }
     steps:
-      - stepId: list_workflows
-        operationId: $sourceDescriptions.memory.workflows_list
-        successCriteria:
-          - condition: $statusCode == 200
-      - stepId: list_patterns
-        operationId: $sourceDescriptions.memory.patterns_list
-        parameters:
-          - name: agentWallet
-            in: query
-            value: "$inputs.agentWallet"
-          - name: minSuccessRate
-            in: query
-            value: 0.2
-          - name: limit
-            in: query
-            value: 5
+      - stepId: run_eval
+        operationId: $sourceDescriptions.memory.evals_run
+        requestBody:
+          contentType: application/json
+          payload:
+            agentWallet: "$inputs.agentWallet"
+            userAddress: "$inputs.userAddress"
+            testCases:
+              - { query: "What is my favourite color?", expected: "azure" }
+              - { query: "What do I work on?",          expected: "AI agents" }
+              - { query: "What music do I like?",       expected: "jazz" }
         successCriteria:
           - condition: $statusCode == 200
-      - stepId: run_decay_job
+          - condition: $response.body#/scores/recallAtK >= 0.95
+
+  # ──────────────────────────────────────────────────────────────────
+  # maintenance_loop — Temporal already runs these on schedule.
+  # Hit only for out-of-band forced passes (bulk imports, debugging).
+  # ──────────────────────────────────────────────────────────────────
+  - workflowId: maintenance_loop
+    summary: Force decay refresh + cleanup out-of-band.
+    description: |
+      Daily/weekly Temporal schedules already run these. Use only
+      after bulk imports or for debugging.
+    inputs:
+      type: object
+      required: [agentWallet]
+      properties:
+        agentWallet: { type: string, examples: ["x-env: COMPOSE_TEST_AGENT_WALLET"] }
+        halfLifeDays: { type: integer, default: 30 }
+    steps:
+      - stepId: run_decay_refresh
         operationId: $sourceDescriptions.memory.jobs_create
         requestBody:
           contentType: application/json
           payload:
             type: decay_update
             execution: inline
-            halfLifeDays: 30
+            halfLifeDays: "$inputs.halfLifeDays"
         successCriteria:
           - condition: $statusCode == 200
-        outputs:
-          jobId: $response.body#/jobId
-      - stepId: get_decay_job
-        operationId: $sourceDescriptions.memory.jobs_get
-        parameters:
-          - name: jobId
-            in: path
-            value: $steps.run_decay_job.outputs.jobId
+      - stepId: cleanup_expired
+        operationId: $sourceDescriptions.memory.jobs_create
+        requestBody:
+          contentType: application/json
+          payload:
+            type: cleanup
+            execution: inline
+            olderThanDays: 90
         successCriteria:
           - condition: $statusCode == 200

package/.speakeasy/tests.arazzo.yaml

@@ -1,8 +1,8 @@
 arazzo: 1.0.1
 info:
   title: Compose SDK Agent-First Workflows
-  summary: E2E workflow contracts for the Compose SDK, inference, x402, and agentic loops.
-  version: 0.6.6
+  summary: E2E workflow contracts for the Compose SDK, inference, x402, and manowar loops.
+  version: 0.6.8
 sourceDescriptions:
   - name: x402
     url: ../specs/x402.openapi.yaml
@@ -10,8 +10,8 @@ sourceDescriptions:
   - name: inference
     url: ../specs/inference.openapi.yaml
     type: openapi
-  - name: agentic
-    url: ../specs/agentic.openapi.yaml
+  - name: manowar
+    url: ../specs/manowar.openapi.yaml
     type: openapi
   - name: memory
     url: ../specs/memory.openapi.yaml
@@ -339,7 +339,7 @@ workflows:
           - condition: $statusCode == 200
           - condition: $response.header.Content-Type != "application/json"
   - workflowId: runtime_agent_workflow_loop
-    summary: Run agent and workflow streams against the Compose agentic service.
+    summary: Run agent and workflow streams against the Compose manowar service.
     x-speakeasy-test-server:
       baseUrl: "x-env: COMPOSE_RUNTIME_TEST_URL; https://runtime.compose.market"
     inputs:
@@ -365,7 +365,7 @@ workflows:
           examples: ["arazzo-thread"]
     steps:
       - stepId: stream_agent
-        operationId: $sourceDescriptions.agentic.agents_stream_create
+        operationId: $sourceDescriptions.manowar.agents_stream_create
         parameters:
           - name: Authorization
             in: header
@@ -391,7 +391,7 @@ workflows:
         successCriteria:
           - condition: $statusCode == 200
       - stepId: stream_workflow
-        operationId: $sourceDescriptions.agentic.workflows_chat_stream
+        operationId: $sourceDescriptions.manowar.workflows_chat_stream
         parameters:
           - name: Authorization
             in: header

package/CHANGELOG.md

@@ -1,5 +1,36 @@
 # Changelog
 
+## 0.6.8
+
+### Highlights
+
+- **First-party 6-layer memory framework.** Mem0 cloud removed entirely. Graph-layer durable facts now live in the same Mongo `memory` collection as conversational vectors (with `source: "fact"`, `metadata.layer: "graph"`), extracted by `gemini-3.1-flash-lite-preview` via our own inference gateway. Voyage `voyage-4-large` (1024-d) is the sole embedder; Cloudflare `bge-reranker-base` is the cross-encoder. Recall p50 dropped from ~905ms (mem0 round-trip) to <120ms.
+- **Cross-thread durable recall now actually works.** The vectors layer was thread-scoped by mistake — fixed; `(agentWallet, userAddress)` is the durability boundary, threadId narrows working+scene only.
+- **Agent memory tools simplified.** `save_memory` + `search_memory` + `search_all_memory` collapsed into `memory_recall` and `memory_remember`. The cross-layer ranker handles which layers contribute; agents do NOT pick layers.
+- **Post-turn persistence is fire-and-forget.** Fact extraction never blocks the agent response.
+- **SDK ergonomic shortcuts:**
+  - `sdk.memory.recall(query, { agentWallet, userAddress })` — string-first pre-turn block
+  - `sdk.memory.save(content, { agentWallet, userAddress, type? })` — string-first durable fact
+- **`MemoryShorthandOptions`** type added for the helpers above.
+- **`AgentMemoryRecordTurnResponse.stored`** now exposes `graph` alongside `transcript`/`working`/`vector` so callers can assert all four bookend layers landed.
+- **Memory Arazzo workflow rewritten** as a 10/10 onboarding doc — covers the canonical `agent_turn_loop`, `maintenance_loop`, `pattern_promotion_loop`, and `eval_loop`. No mem0 references; explicit scope rules; explicit cost / latency budget; per-step descriptions explain when and why.
+
+### Bug fixes
+
+- The vectors layer filter no longer includes `threadId` — durable cross-thread recall now works.
+- API gateway no longer aborts the prepared x402 payment when the SSE `[DONE]` terminator passes through `parseSseEventBlock`.
+- `stopped` and `error` SSE events now carry usage payload so the gateway settles partial work on user-initiated abort.
+
+## 0.6.7
+
+### Highlights
+
+- Agent runtime: rich streaming events. New `tool-args-delta` (per-token tool argument streaming, mirrors Codex `response.function_call_arguments.delta` / Claude Code `input_json_delta`) and `stopped` (user-initiated abort).
+- Agent runtime: `sdk.agent.stop({ agentWallet, runId, threadId? })` aborts an in-flight stream while preserving the LangGraph checkpoint — a subsequent stream call with the same `threadId` resumes the thread's conversation, CoT, and memory.
+- Identity binding restored. Agents always know their on-chain identity (name, description, wallet, skills) via per-turn typed `AgentIdentity` block hydrated once from Pinata IPFS — no more "I'm a helpful assistant" drift.
+- Multi-step tool loop fixed. Removed the kill-switch that unbound tools after one successful batch; the model now decides exit by emitting no tool_calls (LangGraph Deep Agents / Manus / Codex / Claude Code SOTA pattern).
+- Memory tools restored: `save_memory`, `search_memory`, `search_all_memory` re-bound, all routed through the unified `runAgentMemoryLoop` (six-layer memory: working / scene / graph / patterns / archives / vectors).
+
 ## Unreleased
 
 ### Highlights
@@ -34,7 +65,7 @@ Root-cause SDK generation, modality catalog, and local build version-sync pass.
 - `npm test` - 52 passing tests.
 - `npm run typecheck`
 - `npm run build`
-- `node --input-type=module -e '...'` import check for `@compose-market/sdk`, `/x402`, `/inference`, `/inference/modality`, and `/agentic`.
+- `node --input-type=module -e '...'` import check for `@compose-market/sdk`, `/x402`, `/inference`, `/inference/modality`, and `/manowar`.
 - `npm pack --dry-run --cache /tmp/compose-npm-cache`
 
 ---
@@ -48,14 +79,14 @@ OpenAPI/Arazzo contract freeze plus production payment-safety pass.
 - Added canonical Speakeasy-ready contracts:
   - `/Users/jabyl/Downloads/compose-market/packages/sdk/specs/x402.openapi.yaml` for x402 settlement, Compose Keys, sessions, payments, and facilitator operations.
   - `/Users/jabyl/Downloads/compose-market/packages/sdk/specs/inference.openapi.yaml` for model discovery, inference, and realtime inference streams.
-  - `/Users/jabyl/Downloads/compose-market/packages/sdk/specs/agentic.openapi.yaml` for agents, workflows, memory, workspace search, tools, MCP, and mesh execution.
+  - `/Users/jabyl/Downloads/compose-market/packages/sdk/specs/manowar.openapi.yaml` for agents, workflows, memory, workspace search, tools, MCP, and mesh execution.
   - `/Users/jabyl/Downloads/compose-market/packages/sdk/.speakeasy/tests.arazzo.yaml` for model discovery, Compose Key session lifecycle, Compose Key inference, raw x402 challenge/retry, streaming inference frames, and runtime agent/workflow loops.
 - Added `.speakeasy/workflow.yaml` and `.speakeasy/gen.yaml` so the SDK repo is ready for Speakeasy generation once the CLI is authenticated.
 - Generated and exported Speakeasy TypeScript clients:
   - `@compose-market/sdk/x402`, `/x402/keys`, `/x402/session`, and `/x402/payments` for x402 settlement, Compose Keys, session state, payments, and facilitator operations.
   - `@compose-market/sdk/inference` for model discovery, inference, and realtime streams.
-  - `@compose-market/sdk/agentic` plus `/agentic/agent`, `/agentic/workflow`, `/agentic/memory`, and `/agentic/tools`.
-  - Generated schema and operation types under `/inference/schemas`, `/inference/operations`, `/agentic/schemas`, and `/agentic/operations`.
+  - `@compose-market/sdk/manowar` plus `/manowar/agent`, `/manowar/workflow`, `/manowar/memory`, and `/manowar/tools`.
+  - Generated schema and operation types under `/inference/schemas`, `/inference/operations`, `/manowar/schemas`, and `/manowar/operations`.
 - Root `npm run build` now compiles the hand-written orchestration SDK plus both generated clients before publish.
 - Hardened money types: Compose Key atomic amounts are strings end to end; `budgetUsd` is now an exact decimal string and `budgetWei` is a positive integer string.
 - Added explicit inference payment modes:
@@ -75,7 +106,7 @@ OpenAPI/Arazzo contract freeze plus production payment-safety pass.
 - `speakeasy lint arazzo` passes for `.speakeasy/tests.arazzo.yaml`.
 - `speakeasy run -y --skip-upload-spec --skip-testing --output console --target x402` passes and generated `generated/x402`.
 - `speakeasy run -y --skip-upload-spec --skip-testing --output console --target inference` passes and generated `generated/inference`.
-- `speakeasy run -y --skip-upload-spec --skip-testing --output console --target agentic` passes and generated `generated/agentic`.
+- `speakeasy run -y --skip-upload-spec --skip-testing --output console --target manowar` passes and generated `generated/manowar`.
 
 ### Speakeasy generation notes
 

package/README.md

@@ -77,16 +77,16 @@ The published package also exports generated clients for direct, contract-first
 ```ts
 import { ComposeMarket as X402 } from "@compose-market/sdk/x402";
 import { ComposeMarket as Inference } from "@compose-market/sdk/inference";
-import { ComposeMarket as Agentic } from "@compose-market/sdk/agentic";
+import { ComposeMarket as Agentic } from "@compose-market/sdk/manowar";
 ```
 
 - `@compose-market/sdk/x402` exposes generated contracts for x402 settlement, reusable Compose Keys, session state, payments, and facilitator operations.
 - `@compose-market/sdk/x402/keys`, `@compose-market/sdk/x402/session`, `@compose-market/sdk/x402/payments`, and `@compose-market/sdk/x402/feedback` expose generated payment/session/feedback resources directly.
 - `@compose-market/sdk/inference` exposes generated contracts for model discovery, inference, and realtime inference streams.
 - `@compose-market/sdk/inference/modality` exposes the generated modality catalog resource directly.
-- `@compose-market/sdk/agentic` exposes generated contracts for agents, workflows, memory, workspace search, tools, MCP, and mesh execution.
-- `@compose-market/sdk/agentic/agent`, `@compose-market/sdk/agentic/workflow`, `@compose-market/sdk/agentic/memory`, and `@compose-market/sdk/agentic/tools` expose generated agentic resources directly.
-- `@compose-market/sdk/inference/schemas`, `@compose-market/sdk/inference/operations`, `@compose-market/sdk/agentic/schemas`, and `@compose-market/sdk/agentic/operations` expose generated schema and operation types.
+- `@compose-market/sdk/manowar` exposes generated contracts for agents, workflows, memory, workspace search, tools, MCP, and mesh execution.
+- `@compose-market/sdk/manowar/agent`, `@compose-market/sdk/manowar/workflow`, `@compose-market/sdk/manowar/memory`, and `@compose-market/sdk/manowar/tools` expose generated manowar resources directly.
+- `@compose-market/sdk/inference/schemas`, `@compose-market/sdk/inference/operations`, `@compose-market/sdk/manowar/schemas`, and `@compose-market/sdk/manowar/operations` expose generated schema and operation types.
 
 The default `ComposeSDK` remains the higher-level orchestration surface: Compose Key first, raw x402 challenge/sign/retry fallback, typed receipts, streaming aggregation, storage, feedback, and event bus.
 
@@ -230,9 +230,9 @@ Not bundled:
 
 ## Contract generation
 
-The canonical OpenAPI contracts live in `specs/x402.openapi.yaml`, `specs/inference.openapi.yaml`, and `specs/agentic.openapi.yaml`. Speakeasy configuration lives in `.speakeasy/`, including the Arazzo workflow tests for model selection, Compose Key sessions, Compose Key inference, raw x402 challenge/retry, streaming frames, and agentic agent/workflow loops.
+The canonical OpenAPI contracts live in `specs/x402.openapi.yaml`, `specs/inference.openapi.yaml`, and `specs/manowar.openapi.yaml`. Speakeasy configuration lives in `.speakeasy/`, including the Arazzo workflow tests for model selection, Compose Key sessions, Compose Key inference, raw x402 challenge/retry, streaming frames, and manowar agent/workflow loops.
 
-Generated clients live under `generated/x402`, `generated/inference`, and `generated/agentic`. The root build compiles the hand-written orchestration client and all generated clients before publish.
+Generated clients live under `generated/x402`, `generated/inference`, and `generated/manowar`. The root build compiles the hand-written orchestration client and all generated clients before publish.
 
 Speakeasy generation keeps `text/event-stream` in the contracts and uses envelope responses so payment, receipt, and request headers remain part of the typed generated surface. The ergonomic streaming aggregation helpers live on the default `ComposeSDK` orchestration client.
 
@@ -244,7 +244,7 @@ Speakeasy generation keeps `text/event-stream` in the contracts and uses envelop
 - `specs/*.openapi.yaml` `info.version`.
 - `.speakeasy/tests.arazzo.yaml` `info.version`.
 - `.speakeasy/gen.yaml` TypeScript generation version.
-- Generated package metadata under `generated/x402`, `generated/inference`, and `generated/agentic`.
+- Generated package metadata under `generated/x402`, `generated/inference`, and `generated/manowar`.
 - `src/version.ts`.
 
 ## License

package/dist/resources/agent.d.ts

@@ -36,6 +36,18 @@ export declare class AgentResource {
     private readonly ctx;
     constructor(ctx: AgentResourceContext);
     stream(params: AgentStreamCreateParams, options?: ComposeCallOptions): ComposeStreamIterator<AgentRuntimeEvent, AgentStreamFinalResult>;
+    /**
+     * Abort an in-flight stream for (agentWallet, runId). Conversation/CoT/memory
+     * for the thread are preserved by the LangGraph checkpoint and can be resumed
+     * by issuing a new stream call with the same threadId.
+     */
+    stop(params: {
+        agentWallet: string;
+        runId: string;
+        threadId?: string;
+    }): Promise<{
+        stopped: boolean;
+    }>;
 }
 export type { AgentStreamCreateParams, AgentStreamFinalResult, AgentRuntimeEvent, SessionBudgetSnapshot, SessionInvalidReason };
 //# sourceMappingURL=agent.d.ts.map