github.com/Agent-Field/agentfield

Fixed

• Fix(control-plane): enforce lifecycle_status consistency with agent state (#130)

When agents go offline, the control plane was incorrectly keeping lifecycle_status as "ready" even though health_status correctly showed "inactive". This caused observability webhooks to receive inconsistent data where offline nodes appeared online based on lifecycle_status.

Changes:

• Add defensive lifecycle_status enforcement in persistStatus() to ensure consistency with agent state before writing to storage
• Update health_monitor.go fallback paths to also update lifecycle_status
• Add SystemStateSnapshot event type for periodic agent inventory
• Enhance execution events with full reasoner context and metadata
• Add ListAgents to ObservabilityWebhookStore interface for snapshots

The fix ensures both node_offline events and system_state_snapshot events (every 60s) correctly report lifecycle_status: "offline" for offline agents.

🤖 Generated with Claude Code

Co-authored-by: Claude noreply@anthropic.com (67c67c4)

Other

• Switch hot-reload dev setup from Docker to native Air (#129)

Removes Docker-based dev setup in favor of running Air directly in the host environment. This avoids networking issues between Docker and host (especially on WSL2 where host.docker.internal has limitations).

Changes:

• Remove Dockerfile.dev and docker-compose.dev.yml
• Update dev.sh to run Air natively (auto-installs if missing)
• Update README.md with simplified instructions

Usage remains simple: ./dev.sh

🤖 Generated with Claude Code

Co-authored-by: Claude noreply@anthropic.com (670c0ba)

Other

• Hot reload controlplane local setup (#128) (690d481)

CI

• Ci: disable AI label workflow for fork compatibility

The AI label workflow fails on PRs from forked repositories because GITHUB_TOKEN lacks write permissions. Since many contributions come from forks, disabling the workflow until a proper solution (PAT or GitHub App) is implemented.

🤖 Generated with Claude Code

Co-Authored-By: Claude Opus 4.5 noreply@anthropic.com (6dbc908)

Other

• Add explicit return type to useFilterState hook (#127)

• Add explicit return type to useFilterState hook

• fix(types): use Partial<ExecutionFilters> in UseFilterStateReturn

The convertTagsToApiFormat function returns Partial<ExecutionFilters>, so the return type interface must match to avoid TypeScript errors.

🤖 Generated with Claude Code

Co-Authored-By: Claude Opus 4.5 noreply@anthropic.com

---

Co-authored-by: Abir Abbas abirabbas1998@gmail.com Co-authored-by: Claude Opus 4.5 noreply@anthropic.com (f2168e0)

Added

• Feat(sdk/go, control-plane): Add Vector Memory Ops (#124)

• chore(release): v0.1.26-rc.2

• feat: added vector memory ops

• test(handlers): add unit tests for GetVector and DeleteVector handlers

Add comprehensive test coverage for the new vector memory endpoints:

GetVectorHandler tests:

• TestGetVectorHandler_ReturnsVectorWithMetadata: Full happy path with scope/key/metadata
• TestGetVectorHandler_NotFound: 404 when vector doesn't exist
• TestGetVectorHandler_StorageError: 500 on database failure
• TestGetVectorHandler_DefaultScope: Scope resolution from headers

DeleteVectorHandler tests:

• TestDeleteVectorHandler_RESTfulDelete: DELETE with path parameter
• TestDeleteVectorHandler_BackwardCompatibilityWithBody: POST with JSON body
• TestDeleteVectorHandler_StorageError: 500 on database failure
• TestDeleteVectorHandler_MissingKey: 400 when key is missing

Also updated vectorStorageStub to track GetVector and DeleteVector parameters for assertion verification.

🤖 Generated with Claude Code

Co-Authored-By: Claude Opus 4.5 noreply@anthropic.com

---

Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com> Co-authored-by: Abir Abbas abirabbas1998@gmail.com Co-authored-by: Claude Opus 4.5 noreply@anthropic.com (0dd4e62)

Fixed

• Fix(ci): add issues:write permission for AI label workflow (#126)

The gh pr edit --add-label command requires issues: write permission because labels are managed through the issues API in GitHub, even when applied to pull requests. Without this permission, the workflow fails with: "GraphQL: Resource not accessible by integration (addLabelsToLabelable)"

Added permissions:

• issues: write - Required for adding labels
• contents: read - Explicit permission for checkout

🤖 Generated with Claude Code

Co-authored-by: Claude noreply@anthropic.com (26a71a4)

• Fix(ci): prevent shell injection in AI label workflow (#125)

The PR body was being directly interpolated into a shell variable using ${{ github.event.pull_request.body }}, which caused shell parsing of the content. When PR descriptions contained filenames like CHANGELOG.md on their own lines, the shell would attempt to execute them as commands.

This fix passes the PR body via the env: block instead, which properly escapes the content as an environment variable.

🤖 Generated with Claude Code

Co-authored-by: Claude noreply@anthropic.com (1e2225d)

Added

• Feat(tests): add unit tests for vector memory handler functionality (#123) (9df214d)

• Feat: Add AI-assisted contribution guidelines, task issue template, AI labeling workflow, security, and support policies. (87c9297)

Added

• Feat(observability): add webhook forwarding with dead letter queue (#102)

• feat(observability): add webhook forwarding with dead letter queue

Add observability webhook system for forwarding events to external endpoints:

• Configurable webhook URL with optional HMAC secret and custom headers
• Event batching with configurable size and timeout
• Automatic retry with exponential backoff
• Dead letter queue (DLQ) for failed events with redrive and clear capabilities
• Filter heartbeat events and minor health score fluctuations to reduce noise
• Settings UI page for configuration and DLQ management

🤖 Generated with Claude Code

Co-Authored-By: Claude Opus 4.5 noreply@anthropic.com

• test fixes

• fix(observability): node offline events now properly forwarded to webhook

Fixed a race condition where node offline events were not being forwarded to the observability webhook. The issue was in UpdateAgentStatus which called GetAgentStatus (performing a live health check) to get the "old" status. By the time the health check completed, oldStatus == newStatus, so no events were broadcast.

Changed to use GetAgentStatusSnapshot which returns cached/stored status without a live check, preserving the true "old" state for comparison.

Also added observability webhook documentation to ARCHITECTURE.md.

🤖 Generated with Claude Code

Co-Authored-By: Claude Opus 4.5 noreply@anthropic.com

• test(status-manager): add tests for node online/offline event broadcasting

Add comprehensive tests to verify status change events are properly broadcast when nodes transition between active and inactive states:

• TestStatusManagerBroadcastsNodeOfflineEvent: Verifies NodeOffline/ NodeUnifiedStatusChanged events are broadcast when node goes offline
• TestStatusManagerBroadcastsNodeOnlineEvent: Verifies NodeOnline/ NodeUnifiedStatusChanged events are broadcast when node comes online
• TestStatusManagerPreservesOldStatusForEventBroadcast: Verifies the old status is correctly captured before updates, ensuring the fix for the GetAgentStatus race condition doesn't regress

These tests guard against the race condition where UpdateAgentStatus was calling GetAgentStatus (with live health check) instead of GetAgentStatusSnapshot, causing oldStatus == newStatus.

🤖 Generated with Claude Code

Co-Authored-By: Claude Opus 4.5 noreply@anthropic.com

• docs: remove observability webhook documentation

The observability webhook feature remains functional but will not be publicly documented at this time.

🤖 Generated with Claude Code

Co-Authored-By: Claude noreply@anthropic.com

• bugfix on ts-sdk json schema return

• webhook secret fix

• refine webhook events

---

Co-authored-by: Claude Opus 4.5 noreply@anthropic.com (27cf1f7)

Other

• Fix link to AI Backend blog post (fa379c6)

• Update links in README for IAM documentation (3a160de)

• Update README.md (fa62719)

Added

• Feat(dashboard): add comprehensive observability enhancements (#101) (d947542)

Other

• Added note() method for fire-and-forget execution logging (#100) (55fdcf1)

• URL-encode badge endpoint and add cache control (8a4c970)

• Fix shields.io badge: separate badge.json and stats.json files (7d08183)

• Add workflow to update download stats (#87)

• Add workflow to update download stats

Adds a GitHub Actions workflow to automate the collection and updating of download statistics from GitHub releases, PyPI, and NPM.

This workflow:

• Runs every 6 hours or can be triggered manually.
• Fetches download counts from GitHub releases and aggregates them.
• Retrieves total downloads from Pepy.tech for PyPI.
• Collects lifetime download statistics from NPM.
• Calculates a combined total and updates a Gist file.
• The README's download badge is updated to point to this new Gist endpoint for more comprehensive stats.

• Add push trigger for download stats workflow

• Add permissions block to download stats workflow (d400e36)

Chores

• Chore(rag-eval): update default model to GPT-4o (#85)

• Set GPT-4o as default (reliable JSON output)

• Gemini 2.5 Flash as second option

• Move DeepSeek to last (can timeout)

• Remove old Gemini 2.0 Flash

🤖 Generated with Claude Code

Co-authored-by: Claude Opus 4.5 noreply@anthropic.com (f49959a)

Documentation

• Docs: Add cross-reference links to RAG evaluator documentation (#83)

• Add docs link in examples/README.md table for rag_evaluation

• Add documentation callout in rag_evaluation/README.md

🤖 Generated with Claude Code

Co-authored-by: Claude Opus 4.5 noreply@anthropic.com (e51b8d4)

• Docs: Add website deployment guide links (#82)

Reference the full deployment guides at agentfield.ai for Kubernetes and Helm deployment options.

🤖 Generated with Claude Code

Co-authored-by: Claude Opus 4.5 noreply@anthropic.com (59eaf7f)

Other

• Fix parent-child workflow tracking for direct reasoner calls via AgentRouter (#86)

• Fix parent-child workflow tracking for direct reasoner calls via AgentRouter

When reasoners registered via AgentRouter call other reasoners directly (as normal async functions), the parent-child relationship was not being captured in the workflow DAG. This happened because:

1. @router.reasoner() stored the original function but returned it unchanged
2. When include_router() later wrapped functions with tracking, closures in other reasoners still held references to the original unwrapped functions
3. Direct calls bypassed the tracked wrapper entirely

This fix implements lazy-binding in AgentRouter.reasoner():

• The decorator now returns a wrapper that looks up the tracked function at call time via router._tracked_functions
• include_router() registers tracked functions in this lookup table
• Direct reasoner-to-reasoner calls now go through the tracked wrapper, enabling proper parent_execution_id propagation

Changes:

• router.py: Add lazy-binding wrapper in reasoner() decorator
• agent.py: Register tracked functions in router._tracked_functions
• test_router.py: Update tests for new wrapper behavior
• test_workflow_parent_child.py: Add comprehensive tests for parent-child tracking

• Remove unused imports in test_workflow_parent_child.py (342af92)

Other

• Deployments: Docker/Helm/Kustomize quickstarts + demo agents (#81)

• Update Docker deployment and configuration

Refactors the Docker deployment documentation and configuration to improve clarity and flexibility for setting up control planes and agents.

Key changes include:

• Enhancing the README for Docker deployments with more detailed instructions for running agents in Docker, distinguishing between agents on the host and agents within the same Docker Compose network.
• Adding specific guidance on using host.docker.internal for host-based agents and service names for agents within the same network.
• Introducing new Docker Compose services for a demo Go agent and a demo Python agent, enabling them to be run with Docker Compose profiles.
• Updating configuration options in control-plane/internal/config/config.go to include mapstructure tags, improving the flexibility of configuration loading.
• Adding a new test case TestLoadConfig_VCRequirementsFromConfigFile to verify loading VC requirements from a configuration file.
• Modifying the Python hello world example to use an environment variable for the AgentField server URL, making it more adaptable to different deployment scenarios.
• Updating the Dockerized README to include validation steps for execution paths and Verifiable Credentials (VCs).

• Update deployment documentation and manifests

Updates the README files for Docker, Helm, and Kubernetes deployments to improve clarity and provide more streamlined quick-start guides.

The changes include:

• Simplifying the Docker Compose setup instructions.
• Refining the Helm chart documentation to recommend PostgreSQL and the Python demo agent by default.
• Streamlining the Kubernetes manifests to suggest the Python demo agent overlay as a recommended starting point.
• Modifying the Python demo agent deployment in Kubernetes to directly install the AgentField SDK from PyPI instead of relying on a pre-built local image. This simplifies the local development workflow for the Python agent.

• Update documentation for deployment examples

Adds instructions for waiting for demo agents to become ready and for building/loading the Go demo agent image with Minikube.

Also includes an example of how to use the API key when authentication is enabled.

Updates the control plane deployment configuration to default AGENTFIELD_CONFIG_FILE to /dev/null.

Adjusts the kustomization file for the postgres demo overlay to use the standard patches key. (b6b0cd3)

Added

• Feat(go-sdk): add ControlPlaneMemoryBackend for distributed memory (#80)

Add a new MemoryBackend implementation that delegates storage to the control plane's /api/v1/memory/* endpoints. This enables distributed, scope-aware memory across agents.

• Implements Set, Get, Delete, and List operations
• Maps SDK scopes (Workflow, Session, User, Global) to API scopes
• User scope maps to "actor" terminology in the API
• Includes comprehensive unit tests

🤖 Generated with Claude Code

Co-authored-by: Claude Opus 4.5 noreply@anthropic.com (6cd445c)

• Feat: add RAG evaluation example with multi-reasoner architecture (#79)

• feat: add RAG evaluation example with multi-reasoner architecture

Adds a comprehensive RAG evaluation agent demonstrating:

• Adversarial debate for faithfulness (prosecutor vs defender + judge)
• Multi-jury consensus for relevance (3 jurors vote on literal/intent/scope)
• Hybrid ML+LLM chain-of-verification for hallucination detection
• Configurable constitutional principles evaluation

Features:

• Docker Compose deployment (control plane + agent + UI)
• Next.js web interface with claim-level breakdown
• Domain-specific presets (medical, legal, financial)
• 3 evaluation modes: quick (4 calls), standard (14), thorough (20+)

🤖 Generated with Claude Code

Co-Authored-By: Claude Opus 4.5 noreply@anthropic.com

• Remove ARCHITECTURE.md, link to website docs instead

The detailed architecture documentation is now on the website at agentfield.dev/examples/complete-agents/rag-evaluator - no need to duplicate content in the repo.

🤖 Generated with Claude Code

Co-Authored-By: Claude Opus 4.5 noreply@anthropic.com

• Add examples README with quick reference table

Index of all examples across Python, TypeScript, and Go with:

• Quick reference table by use case and language
• Detailed per-language tables with key features
• Use case deep dives (RAG progression, multi-agent, serverless)
• Technology stack overview

🤖 Generated with Claude Code

Co-Authored-By: Claude Opus 4.5 noreply@anthropic.com

---

Co-authored-by: Claude Opus 4.5 noreply@anthropic.com (862c41e)

Fixed

• Fix: use executions table for notes storage instead of workflow_executions (#75)

• fix: use executions table for notes storage instead of workflow_executions

The note handlers (AddExecutionNoteHandler, GetExecutionNotesHandler) were querying the workflow_executions table, but execution data is actually stored in the executions table. This caused "execution not found" errors when adding or retrieving notes via app.note().

Changes:

• Add Notes field to types.Execution struct
• Add notes column to ExecutionRecordModel (GORM auto-migrates this)
• Update SQL queries in execution_records.go to include notes column
• Update scanExecution to deserialize notes JSON
• Change ExecutionNoteStorage interface to use GetExecutionRecord and UpdateExecutionRecord instead of GetWorkflowExecution and UpdateWorkflowExecution
• Update AddExecutionNoteHandler to use UpdateExecutionRecord
• Update GetExecutionNotesHandler to use GetExecutionRecord

This fixes both the SDK app.note() functionality and the UI notes panel 404 errors.

🤖 Generated with Claude Code

Co-Authored-By: Claude Opus 4.5 noreply@anthropic.com

• fix: update execution notes tests to use correct storage methods

Tests were using WorkflowExecution type and StoreWorkflowExecution() to set up test data, but the handlers now use Execution type and GetExecutionRecord()/ UpdateExecutionRecord() which query the executionRecords map.

• Change test setup from types.WorkflowExecution to types.Execution
• Change StoreWorkflowExecution() to CreateExecutionRecord()
• Change GetWorkflowExecution() verification to GetExecutionRecord()
• Rename workflowID to runID to match the Execution struct field

🤖 Generated with Claude Code

Co-Authored-By: Claude Opus 4.5 noreply@anthropic.com

---

Co-authored-by: Claude Opus 4.5 noreply@anthropic.com (5dd327e)

Fixed

• Fix: wire up workflow notes SSE endpoint (#74)

The StreamWorkflowNodeNotesHandler existed but was never registered in the routes. This adds the missing route registration for: GET /api/ui/v1/workflows/:workflowId/notes/events

🤖 Generated with Claude Code

Co-authored-by: Claude Opus 4.5 noreply@anthropic.com (c6f31cb)

Added

• Feat(go-sdk): add per-request API key override for AI client (#73)

Add WithAPIKey option to override the client's configured API key on a per-request basis. This brings the Go SDK to parity with the Python SDK, which supports api_key overrides in individual calls.

Changes:

• Add APIKeyOverride field to Request struct (excluded from JSON)
• Add WithAPIKey option function
• Update doRequest and StreamComplete to use override when provided
• Add test for API key override behavior

🤖 Generated with Claude Code

Co-authored-by: Claude Opus 4.5 noreply@anthropic.com (4dd8a70)

Added

• Feat(go-sdk): add Memory and Note APIs for agent state and progress tracking (#71)

Add two major new capabilities to the Go SDK:

Added

• Feat: allow external contributors to run functional tests without API… (#70)

• feat: allow external contributors to run functional tests without API keys

Enable external contributors to run 92% of functional tests (24/26) without requiring access to OpenRouter API keys. This makes it easier for the community to contribute while maintaining full test coverage for maintainers.

Changes:

• Detect forked PRs and automatically skip OpenRouter-dependent tests
• Only 2 tests require OpenRouter (LLM integration tests)
• 24 tests validate all core infrastructure without LLM calls
• Update GitHub Actions workflow to conditionally set PYTEST_ARGS
• Update functional test README with clear documentation

Test coverage for external contributors: ✅ Control plane health and APIs ✅ Agent registration and discovery ✅ Multi-agent communication ✅ Memory system (all scopes) ✅ Workflow orchestration ✅ Go/TypeScript SDK integration ✅ Serverless agents ✅ Verifiable credentials

Skipped for external contributors (maintainers still run these): ⏭️ test_hello_world_with_openrouter ⏭️ test_readme_quick_start_summarize_flow

This change addresses the challenge of running CI for external contributors without exposing repository secrets while maintaining comprehensive test coverage for the core AgentField platform functionality.

• fix: handle push events correctly in functional tests workflow

The workflow was failing on push events (to main/testing branches) because it relied on github.event.pull_request.head.repo.fork which is null for push events. This caused the workflow to incorrectly fall into the else branch and fail when OPENROUTER_API_KEY wasn't set.

Changes:

• Check github.event_name to differentiate between push, pull_request, and workflow_dispatch
• Explicitly handle push and workflow_dispatch events to run all tests with API key
• Preserve fork PR detection to skip OpenRouter tests for external contributors

Now properly handles: ✅ Fork PRs: Skip 2 OpenRouter tests, run 24/26 tests ✅ Internal PRs: Run all 26 tests with API key ✅ Push to main/testing: Run all 26 tests with API key ✅ Manual workflow dispatch: Run all 26 tests with API key

🤖 Generated with Claude Code

Co-Authored-By: Claude Sonnet 4.5 noreply@anthropic.com

• fix: remove shell quoting from PYTEST_ARGS to prevent argument parsing errors

The PYTEST_ARGS variable contained single quotes around '-m "not openrouter" -v' which would be included in the environment variable value. When passed to pytest in the Docker container shell command, this caused the entire string to be treated as a single argument instead of being properly split into separate arguments.

Changed from: '-m "not openrouter" -v' Changed to: -m not openrouter -v

This allows the shell's word splitting to correctly parse the arguments when pytest $$PYTEST_ARGS is evaluated in the docker-compose command.

🤖 Generated with Claude Code

Co-Authored-By: Claude Sonnet 4.5 noreply@anthropic.com

• refactor: separate pytest marker expression from general args for proper quoting

The previous approach of embedding -m not openrouter inside PYTEST_ARGS was fragile because shell word-splitting doesn't guarantee "not openrouter" stays together as a single argument to the -m flag.

This change introduces PYTEST_MARK_EXPR as a dedicated variable for the marker expression, which is then properly quoted when passed to pytest: pytest -m "$PYTEST_MARK_EXPR" $PYTEST_ARGS ...

Benefits:

• Marker expression is guaranteed to be treated as single argument to -m
• Clear separation between marker selection and general pytest args
• More maintainable for future marker additions
• Eliminates shell quoting ambiguity

Changes:

• workflow: Split PYTEST_ARGS into PYTEST_MARK_EXPR + PYTEST_ARGS
• docker-compose: Add PYTEST_MARK_EXPR env var and conditional -m flag
• docker-compose: Only apply -m when PYTEST_MARK_EXPR is non-empty

🤖 Generated with Claude Code

Co-Authored-By: Claude Sonnet 4.5 noreply@anthropic.com

• fix: add proper event type checks before accessing pull_request context

Prevent errors when workflow runs on push events by:

• Check event_name == 'pull_request' before accessing pull_request.head.repo.fork
• Check event_name == 'workflow_dispatch' before accessing event.inputs
• Ensures all conditional expressions only access context properties when they exist

This prevents "Error: Cannot read properties of null (reading 'fork')" errors on push events.

🤖 Generated with Claude Code

Co-Authored-By: Claude Sonnet 4.5 noreply@anthropic.com

---

Co-authored-by: Claude Sonnet 4.5 noreply@anthropic.com (01668aa)

Fixed

• Fix(python-sdk): move conditional imports to module level (#72)

The serve() method had import os and import urllib.parse statements inside conditional blocks. When an explicit port was passed, the first conditional block was skipped, but Python's scoping still saw the later conditional imports, causing an UnboundLocalError when trying to use os.getenv() at line 1140.

Error seen in Docker containers:

UnboundLocalError: cannot access local variable 'os' where it is not
associated with a value

This worked locally because auto_port=True executed the first code path which included import os, but failed in Docker when passing an explicit port value.

Fix: Move all imports to module level where they belong.

🤖 Generated with Claude Code

Co-authored-by: Claude Opus 4.5 noreply@anthropic.com (a0d0538)

Other

• Test pr 68 init fix (#69)

• fix(cli): fix init command input handling issues

• Fix j/k keys not registering during text input
• Fix ctrl+c not cancelling properly
• Fix selected option shifting other items
• Filter special keys from text input
• Add ctrl+u to clear input line
• Add unit tests for init model

🤖 Generated with Claude Code

Co-Authored-By: Claude Opus 4.5 noreply@anthropic.com

• docs: add changelog entry for CLI init fixes

🤖 Generated with Claude Code

Co-Authored-By: Claude Opus 4.5 noreply@anthropic.com

• chore: trigger CI with secrets

• chore: remove manual changelog entry (auto-generated on release)

---

Co-authored-by: fimbulwinter sanandsankalp@gmail.com Co-authored-by: Claude Opus 4.5 noreply@anthropic.com (55d0c61)

Fixed

• Fix: correct parent execution ID for sub-calls in app.call() (#62)

When a reasoner calls a skill via app.call(), the X-Parent-Execution-ID header was incorrectly set to the inherited parent instead of the current execution. This caused workflow graphs to show incorrect parent-child relationships.

The fix overrides X-Parent-Execution-ID to use the current execution's ID after to_headers() is called, ensuring sub-calls are correctly attributed as children of the calling execution.

Co-authored-by: Ivan Viljoen 8543825+ivanvza@users.noreply.github.com (762142e)

Other

• Update README to remove early adopter notice

Removed early adopter section from README. (054fc22)

• Update README.md (dae57c7)

• Update README.md (06e5cee)

• Update README.md (39c2da4)

Other

• Add serverless agent examples and functional tests (#46)

• Add serverless agent examples and functional tests

• Add CLI support for serverless node registration

• Fix serverless execution payload initialization

• Harden serverless functional test to use CLI registration

• Broaden serverless CLI functional coverage

• Persist serverless invocation URLs

• Ensure serverless executions hit /execute

• Fix serverless agent metadata loading

• Derive serverless deployment for stored agents

• Honor serverless metadata during execution

• Backfill serverless invocation URLs on load

• Stabilize serverless agent runtime

• Harden serverless functional harness

• Support serverless agents via reasoners endpoint

• Log serverless reasoner responses for debugging

• Allow custom serverless adapters across SDKs

• Normalize serverless handler responses

• Fix Python serverless adapter typing

• Make serverless adapter typing py3.9-safe

• Fix Python serverless execution context

• Simplify Python serverless calls to sync

• Mark serverless Python agents connected for cross-calls

• Force sync execution path in serverless handler

• Handle serverless execute responses without result key

• Align serverless Python relay args with child signature

• feat: Add workflow performance visualizations, including agent health heatmap and execution scatter plot, and enhance UI mobile responsiveness.

• chore: Remove unused Badge import from ExecutionScatterPlot.tsx and add an empty line to .gitignore. (728e4e0)

• Added docker (74f111b)

• Update README.md (8b580cb)

Fixed

• Fix(sdk/typescript): add DID registration to enable VC generation (#60)

• fix(release): skip example requirements for prereleases

Restore the check to skip updating example requirements for prerelease versions. Even though prereleases are now published to PyPI, pip install excludes them by default per PEP 440. Users running pip install -r requirements.txt would fail without the --pre flag.

Examples should always pin to stable versions so they work out of the box.

🤖 Generated with Claude Code

Co-Authored-By: Claude noreply@anthropic.com

• fix(sdk/typescript): add DID registration to enable VC generation

The TypeScript SDK was not registering with the DID system, causing VC generation to fail with "failed to resolve caller DID: DID not found".

This change adds DID registration to match the Python SDK's behavior:

• Add DIDIdentity types and registerAgent() to DidClient
• Create DidManager class to store identity package after registration
• Integrate DidManager into Agent.ts to auto-register on startup
• Update getDidInterface() to resolve DIDs from stored identity package

When didEnabled is true, the agent now:

1. Registers with /api/v1/nodes/register (existing)
2. Registers with /api/v1/did/register (new)
3. Stores identity package for DID resolution
4. Auto-populates callerDid/targetDid when generating VCs

🤖 Generated with Claude Code

Co-Authored-By: Claude noreply@anthropic.com

• feat(examples): add verifiable credentials TypeScript example

Add a complete VC example demonstrating:

• Basic text processing with explicit VC generation
• AI-powered analysis with VC audit trail
• Data transformation with integrity proof
• Multi-step workflow with chained VCs

🤖 Generated with Claude Code

Co-Authored-By: Claude noreply@anthropic.com

• fix(examples): fix linting errors in VC TypeScript example

• Remove invalid note property from workflow.progress calls
• Simplify AI response handling since schema already returns parsed type

🤖 Generated with Claude Code

Co-Authored-By: Claude noreply@anthropic.com

---

Co-authored-by: Claude noreply@anthropic.com (bd097e1)

• Fix(release): skip example requirements for prereleases (#59)

Restore the check to skip updating example requirements for prerelease versions. Even though prereleases are now published to PyPI, pip install excludes them by default per PEP 440. Users running pip install -r requirements.txt would fail without the --pre flag.

Examples should always pin to stable versions so they work out of the box.

🤖 Generated with Claude Code

Co-authored-by: Claude noreply@anthropic.com (1b7d9b8)

Added

• Feat(release): unify PyPI publishing for all releases (#58)

Publish all Python SDK releases (both prerelease and stable) to PyPI instead of using TestPyPI for prereleases.

Per PEP 440, prerelease versions (e.g., 0.1.20rc1) are excluded by default from pip install - users must explicitly use --pre flag. This simplifies the release process and removes the need for the TEST_PYPI_API_TOKEN secret.

Changes:

• Merge TestPyPI and PyPI publish steps into single PyPI step
• Update release notes to show pip install --pre for staging
• Update install.sh staging output
• Re-enable example requirements updates for prereleases
• Update RELEASE.md documentation

🤖 Generated with Claude Code

Co-authored-by: Claude noreply@anthropic.com (ebf7020)

Fixed

• Fix(release): fix example requirements and prevent future staging bumps (#56)

• fix(examples): revert to stable agentfield version (0.1.19)

The staging release bumped example requirements to 0.1.20-rc.1, but RC versions are published to TestPyPI, not PyPI. This caused Railway deployments to fail because pip couldn't find the package.

Revert to the last stable version (0.1.19) which is available on PyPI.

• fix(release): skip example requirements bump for prerelease versions

Prerelease versions are published to TestPyPI, not PyPI. If we bump example requirements.txt files to require a prerelease version, Railway deployments will fail because pip looks at PyPI by default.

Now bump_version.py only updates example requirements for stable releases, ensuring deployed examples always use versions available on PyPI. (c86bec5)

Added

• Feat(release): add two-tier staging/production release system (#53)

• feat(release): add two-tier staging/production release system

Implement automatic staging releases and manual production releases:

• Staging: Automatic on push to main (PyPI prerelease, npm @next, staging-* Docker)
• Production: Manual workflow dispatch (PyPI, npm @latest, vX.Y.Z + latest Docker)

Changes:

• Add push trigger with path filters for automatic staging
• Replace release_channel with release_environment input
• Unified PyPI publishing for both staging (prerelease) and production
• Split npm publishing: @next tag (staging) vs @latest (production)
• Conditional Docker tagging: staging-X.Y.Z vs vX.Y.Z + latest
• Add install-staging.sh for testing prerelease binaries
• Update RELEASE.md with two-tier documentation

🤖 Generated with Claude Code

Co-Authored-By: Claude noreply@anthropic.com

• refactor(install): consolidate staging into single install.sh with --staging flag

Instead of separate install.sh and install-staging.sh scripts:

• Single install.sh handles both production and staging
• Use --staging flag or STAGING=1 env var for prerelease installs
• Eliminates code drift between scripts

Usage: Production: curl -fsSL .../install.sh | bash Staging: curl -fsSL .../install.sh | bash -s -- --staging

🤖 Generated with Claude Code

Co-Authored-By: Claude noreply@anthropic.com

---

Co-authored-by: Claude noreply@anthropic.com (3bd748d)

• Feat(sdk/typescript): expand AI provider support to 10 providers

Add 6 new AI providers to the TypeScript SDK:

• Google (Gemini models)
• Mistral AI
• Groq
• xAI (Grok)
• DeepSeek
• Cohere

Also add explicit handling for OpenRouter and Ollama with sensible defaults.

Changes:

• Update AIConfig type with new provider options
• Refactor buildModel() with switch statement for all providers
• Refactor buildEmbeddingModel() with proper embedding support (Google, Mistral, Cohere have native embedding; others throw)
• Add 27 unit tests for provider selection and embedding support
• Install @ai-sdk/google, @ai-sdk/mistral, @ai-sdk/groq, @ai-sdk/xai, @ai-sdk/deepseek, @ai-sdk/cohere packages

🤖 Generated with Claude Code

Co-Authored-By: Claude noreply@anthropic.com (b06b5b5)

Other

• Update versions (a7912f5)

Fixed

• Fix(ui): add API key header to sidebar execution details fetch

The useNodeDetails hook was making a raw fetch() call without including the X-API-Key header, causing 401 errors in staging where API key authentication is enabled. Other API calls in the codebase use fetchWrapper functions that properly inject the key. (f0ec542)

Fixed

• Fix(sdk): inject API key into all HTTP requests

The Python SDK was not including the X-API-Key header in HTTP requests made through AgentFieldClient._async_request(), causing 401 errors when the control plane has authentication enabled.

This fix injects the API key into request headers automatically when:

• The client has an api_key configured
• The header isn't already set (avoids overwriting explicit headers)

Fixes async status updates and memory operations (vector search, etc.) that were failing with 401 Unauthorized.

🤖 Generated with Claude Code

Co-Authored-By: Claude noreply@anthropic.com (97673bc)

Fixed

• Fix(control-plane): remove redundant WebSocket origin check

The WebSocket upgrader's CheckOrigin was rejecting server-to-server connections (like from Python SDK agents) that don't have an Origin header. This caused 403 errors when agents tried to connect to memory events WebSocket endpoint with auth enabled.

The origin check was redundant because:

1. Auth middleware already validates API keys before this handler
2. If auth is enabled, only valid API key holders reach this point
3. If auth is disabled, all connections are allowed anyway

Removes the origin checking logic and simplifies NewMemoryEventsHandler to just take the storage provider.

🤖 Generated with Claude Code

Co-Authored-By: Claude noreply@anthropic.com (44f05c4)

Fixed

• Fix(example): use IPv4 binding for documentation-chatbot

The documentation chatbot was binding to :: (IPv6 all interfaces) which causes Railway internal networking to fail with "connection refused" since Railway routes traffic over IPv4.

Removed explicit host parameter to use the SDK default of 0.0.0.0 which binds to IPv4 all interfaces.

🤖 Generated with Claude Code

Co-Authored-By: Claude noreply@anthropic.com (2c1b205)

• Fix(python-sdk): include API key in memory events WebSocket connections

The MemoryEventClient was not including the X-API-Key header when connecting to the memory events WebSocket endpoint, causing 401 errors when the control plane has authentication enabled.

Changes:

• Add optional api_key parameter to MemoryEventClient constructor
• Include X-API-Key header in WebSocket connect() method
• Include X-API-Key header in history() method (both httpx and requests)
• Pass api_key from Agent to MemoryEventClient in both instantiation sites

🤖 Generated with Claude Code

Co-Authored-By: Claude noreply@anthropic.com (eda95fc)

Other

• Revert "fix(example): use IPv4 binding for documentation-chatbot"

This reverts commit 2c1b2053e37f4fcc968ad0805b71ef89cf9d6d9d. (576a96c)

Fixed

• Fix(python-sdk): update test mocks for api_key parameter

Update test helpers and mocks to accept the new api_key parameter:

• Add api_key field to StubAgent dataclass
• Add api_key parameter to _FakeDIDManager and _FakeVCGenerator
• Add headers parameter to VC generator test mocks

🤖 Generated with Claude Code

Co-Authored-By: Claude noreply@anthropic.com (301e276)

• Fix(python-sdk): add missing API key headers to DID/VC and workflow methods

Comprehensive fix for API key authentication across all SDK HTTP requests:

DID Manager (did_manager.py):

• Added api_key parameter to init
• Added _get_auth_headers() helper method
• Fixed register_agent() to include X-API-Key header
• Fixed resolve_did() to include X-API-Key header

VC Generator (vc_generator.py):

• Added api_key parameter to init
• Added _get_auth_headers() helper method
• Fixed generate_execution_vc() to include X-API-Key header
• Fixed verify_vc() to include X-API-Key header
• Fixed get_workflow_vc_chain() to include X-API-Key header
• Fixed create_workflow_vc() to include X-API-Key header
• Fixed export_vcs() to include X-API-Key header

Agent Field Handler (agent_field_handler.py):

• Fixed _send_heartbeat() to include X-API-Key header

Agent (agent.py):

• Fixed emit_workflow_event() to include X-API-Key header
• Updated _initialize_did_system() to pass api_key to DIDManager and VCGenerator

All HTTP requests to AgentField control plane now properly include authentication headers when API key is configured.

🤖 Generated with Claude Code

Co-Authored-By: Claude noreply@anthropic.com (2517549)

• Fix(python-sdk): add missing API key headers to sync methods

Add authentication headers to register_node(), update_health(), and get_nodes() methods that were missing X-API-Key headers in requests.

🤖 Generated with Claude Code

Co-Authored-By: Claude noreply@anthropic.com (0c2977d)

Other

• Add Go SDK CallLocal workflow tracking (64c6217)

• Fix Python SDK to include API key in register/heartbeat requests

The SDK's AgentFieldClient stored the api_key but several methods were not including it in their HTTP requests, causing 401 errors when authentication is enabled on the control plane:

• register_agent()
• register_agent_with_status()
• send_enhanced_heartbeat() / send_enhanced_heartbeat_sync()
• notify_graceful_shutdown() / notify_graceful_shutdown_sync()

Also updated documentation-chatbot example to pass AGENTFIELD_API_KEY from environment to the Agent constructor.

🤖 Generated with Claude Code

Co-Authored-By: Claude noreply@anthropic.com (1e6a095)

Added

• Feat: expose api_key at Agent level and fix test lint issues

• Add api_key parameter to Agent class constructor

• Pass api_key to AgentFieldClient for authentication

• Document api_key parameter in Agent docstring

• Fix unused loop variable in ensure_event_loop test fixture

Addresses reviewer feedback that api_key should be exposed at Agent level since end users don't interact directly with AgentFieldClient.

🤖 Generated with Claude Code

Co-Authored-By: Claude noreply@anthropic.com (6567bd0)

• Feat: add API key authentication to control plane and SDKs

This adds optional API key authentication to the AgentField control plane with support in all SDKs (Python, Go, TypeScript).

Other

• Release workflow fix (fde0309)

• Update README.md (c3cfca4)

Chores

• Chore: trigger Railway deployment for PR #39 fix (b4095d2)

Documentation

• Docs(chatbot): add SDK search term relationship

Add search term mapping for SDK/language queries to improve RAG retrieval when users ask about supported languages or SDKs.

🤖 Generated with Claude Code

Co-Authored-By: Claude noreply@anthropic.com (87a4d90)

• Docs(chatbot): add TypeScript SDK to supported languages

Update product context to include TypeScript alongside Python and Go:

• CLI commands now mention all three language options
• Getting started section references TypeScript
• API Reference includes TypeScript SDK

This fixes the RAG chatbot returning only Python/Go when asked about supported languages.

🤖 Generated with Claude Code

Co-Authored-By: Claude noreply@anthropic.com (9510d74)

Fixed

• Fix(vector-store): fix PostgreSQL DeleteByPrefix and update namespace defaults

• Fix DeleteByPrefix to use PostgreSQL || operator for LIKE pattern (the previous approach with prefix+"%" in Go wasn't working correctly with parameter binding)

• Change default namespace from "documentation" to "website-docs" to match the frontend chat API expectations

• Add scope: "global" to clear_namespace API call to ensure proper scope matching

🤖 Generated with Claude Code

Co-Authored-By: Claude noreply@anthropic.com (cbfdf7b)

• Fix(docs-chatbot): use correct start command

Change start command from python -m agentfield.run (doesn't exist) to python main.py (the actual entry point).

🤖 Generated with Claude Code

Co-Authored-By: Claude noreply@anthropic.com (b71507c)

• Fix(docs-chatbot): override install phase for PyPI wait

The previous fix used buildCommand which runs AFTER pip install. This fix overrides the install phase itself:

• Add nixpacks.toml with [phases.install] to run install.sh
• Update railway.json to point to nixpacks.toml
• Update install.sh to create venv before waiting for PyPI

The issue was that buildCommand runs after the default install phase, so pip had already failed before our script ran.

🤖 Generated with Claude Code

Co-Authored-By: Claude noreply@anthropic.com (f8bf14b)

• Fix(docs-chatbot): use railway.json for Railpack PyPI wait

Railway now uses Railpack instead of Nixpacks. Update config:

• Replace nixpacks.toml with railway.json
• Force NIXPACKS builder with custom buildCommand
• Fix install.sh version check using pip --dry-run

🤖 Generated with Claude Code

Co-Authored-By: Claude noreply@anthropic.com (8c22356)

Fixed

• Fix(docs-chatbot): handle PyPI race condition in Railway deploys

Add install script that waits for agentfield package to be available on PyPI before installing. This fixes the race condition where Railway deployment triggers before the release workflow finishes uploading to PyPI.

• Add install.sh with retry logic (30 attempts, 10s intervals)
• Add nixpacks.toml to use custom install script

🤖 Generated with Claude Code

Co-Authored-By: Claude noreply@anthropic.com (e45f41d)

Added

• Feat: add delete-namespace endpoint for RAG reindexing

Adds a new DELETE /api/v1/memory/vector/namespace endpoint that allows clearing all vectors with a given namespace prefix. This enables the documentation chatbot to wipe and reindex its RAG data when docs change.

Changes:

• Add DeleteVectorsByPrefix to StorageProvider interface
• Implement DeleteByPrefix for SQLite and Postgres vector stores
• Add DeleteNamespaceVectorsHandler endpoint
• Add clear_namespace skill to documentation chatbot
• Update MemoryStorage interface with new method

🤖 Generated with Claude Code

Co-Authored-By: Claude noreply@anthropic.com (bc1f41e)

• Feat(sdk-python): expose execution context via app.ctx property

Add a ctx property to the Agent class that provides direct access to the current execution context during reasoner/skill execution. This enables a more ergonomic API:

Before: from agentfield.execution_context import get_current_context ctx = get_current_context() workflow_id = ctx.workflow_id

After: workflow_id = app.ctx.workflow_id

The property returns None when accessed outside of an active execution (e.g., at module level or after a request completes), matching the behavior of app.memory. This prevents accidental use of stale or placeholder context data.

Also fixes integration test fixtures to support the current monorepo structure where control-plane lives at repo root instead of apps/platform/agentfield.

🤖 Generated with Claude Code

Co-Authored-By: Claude noreply@anthropic.com (e01dcea)

• Feat(ts-sdk): add DID client and memory helpers (4b74998)

• Feat(ts-sdk): add heartbeat and local call coverage (cf228ec)

• Feat(ts-sdk): scaffold typescript sdk core (09dcc62)

Chores

• Chore: ignore env files (3937821)

• Chore(ts-sdk): align heartbeat and memory clients, improve example env loading (fee2a7e)

• Chore(ts-sdk): load env config for simulation example (9715ac5)

• Chore(ts-sdk): remove AI stubs from simulation example (7b94190)

• Chore(ts-sdk): make simulation example runnable via build (9a87374)

• Chore(ts-sdk): fix typings, add heartbeat config, lock deps (f9af207)

Fixed

• Fix: revert conftest changes to prevent CI failures

The integration tests should skip gracefully in CI when the control plane cannot be built. Reverting conftest changes that caused the tests to attempt building when they should skip.

🤖 Generated with Claude Code

Co-Authored-By: Claude noreply@anthropic.com (f86794c)

• Fix: remove unused import to pass linting

🤖 Generated with Claude Code

Co-Authored-By: Claude noreply@anthropic.com (5a975fa)

• Fix flaky tests (bfb86cb)

• Fix(ts-sdk): normalize router IDs to align with control plane (7c36c8b)

• Fix(ts-sdk): register full reasoner definitions (e5cc44d)

Other

• Ts sdk (ce3b965)

• Recover agent state on restart and speed up node status (7fa12ca)

• Remove unused configuration variables

Audit of agentfield.yaml revealed many config options that were defined but never actually read or used by the codebase. This creates confusion for users who set these values expecting them to have an effect.

Removed from YAML config:

• agentfield: mode, max_concurrent_requests, request_timeout, circuit_breaker_threshold (none were wired to any implementation)
• execution_queue: worker_count, request_timeout, lease_duration, max_attempts, failure_backoff, max_failure_backoff, poll_interval, result_preview_bytes, queue_soft_limit, waiter_map_limit
• ui: backend_url
• storage.local: cache_size, retention_days, auto_vacuum
• storage: config field
• agents section entirely (discovery/scaling never implemented)

Removed from Go structs:

• AgentsConfig, DiscoveryConfig, ScalingConfig
• CoreFeatures, EnterpriseFeatures
• DataDirectoriesConfig
• Unused fields from AgentFieldConfig, ExecutionQueueConfig, LocalStorageConfig, StorageConfig, UIConfig

The remaining config options are all actively used:

• agentfield.port, execution_cleanup.*, execution_queue webhook settings
• ui.enabled/mode/dev_port
• api.cors.*
• storage.mode/local.database_path/local.kv_store_path/vector.*
• features.did.* (all DID/VC settings)

🤖 Generated with Claude Code

Co-Authored-By: Claude noreply@anthropic.com (ee6e6e0)

• Adds more links to documentation

Adds several new links to the README.md file that direct users to more detailed documentation pages. These links cover production-ready features, comparisons with agent frameworks, the full feature set, and the core architecture. (d5a9922)

• Update documentation links

Updates several external links within the README to point to the correct documentation paths.

This ensures that users can navigate to the relevant guides and information seamlessly. (ac6f777)

• Updated arch (4ed9806)

• Improve README Quick Start guide

Updates the README's quick start section to provide a more comprehensive and user-friendly guide.

This revision clarifies the installation process, introduces a dedicated step for agent creation with a default configuration option using af init --defaults, and specifies the necessary command-line instructions for each terminal in the control plane + agent node architecture.

It also refines the example API call to use a more descriptive agent endpoint (my-agent.demo_echo) and adds examples for Go and TypeScript, as well as detailing how to use interactive mode for agent initialization. (4e897f0)

• Refactor README for clarity and expanded content

Updates the README to provide a more detailed explanation of AgentField's purpose and features.

Key changes include:

• Enhanced "What is AgentField?" section to emphasize its role as backend infrastructure for autonomous AI.

• Improved "Quick Start" section with clearer steps and usage examples.

• Expanded "Build Agents in Any Language" section to showcase Python, Go, TypeScript, and REST API examples.

• Introduced new sections like "The Production Gap" and "Identity & Trust" to highlight AgentField's unique value proposition.

• Refined "Who is this for?" and "Is AgentField for you?" sections for better audience targeting.

• Updated navigation links and visual elements for improved readability and user experience. (f05cd95)

• Typescript schema based formatting improvements (fcda991)

• Typescript release and init (218326b)

• Functional tests (99b6f9e)

• Add TS SDK CI and functional TS agent coverage (857191d)

• Add MCP integration (5bc36d7)

• Separate example freom sdk (909dc8c)

• Memory & Discovery (84ff093)

• TS SDK simulation flow working (5cab496)

• Add .env to git ignore (172e8a9)

• Update README.md (4e0b2e6)

• Fix MemoryEventClient init for sync contexts (1d246ec)

• Fix memory event client concurrency and compatibility (2d28571)

• Improve LLM prompt formatting and citations

Refactors the system and user prompts for the documentation chatbot to improve clarity and LLM performance. This includes:

• Restructuring and clarifying the prompt instructions for citations, providing explicit guidance on how to use and format them.

• Enhancing the citation key map format to be more descriptive and user-friendly for the LLM.

• Explicitly stating that the citations array in the response should be left empty by the LLM, as it will be injected by the system.

• Updating the Citation schema to correctly reflect that the key should not include brackets.

• Adding a specific "REFINEMENT MODE" instruction to the refined prompt to guide the LLM's behavior in a second retrieval attempt.

• Minor cleanup and adjustments to prompt text for better readability. (56246ad)

• Update dependencies for improved compatibility

Updates several npm package dependencies, including browserslist, caniuse-lite, and electron-to-chromium, to their latest versions. This ensures better compatibility and incorporates recent improvements and bug fixes from these packages. (c72278c)

• Implement automatic agent method delegation

Improves the AgentRouter by implementing getattr to automatically delegate any unknown attribute or method access to the attached agent. This eliminates the need for explicit delegation methods for agent functionalities like ai(), call(), memory, note(), and discover().

This change simplifies the AgentRouter's interface and makes it more transparently proxy agent methods. Added tests to verify the automatic delegation for various agent methods and property access, as well as error handling when no agent is attached. (26c9288)

Testing

• Tests hanging fix (dd2eb8d)