github.com/Agent-Field/agentfield

Performance

• Perf(dashboard): skip fetching input/result payloads in dashboard queries (#273)

Dashboard API endpoints (summary and enhanced) call QueryExecutionRecords with limits up to 50,000 rows but never use the input_payload or result_payload columns in any of their processing functions. In deployments with large execution payloads (e.g. ~1.1 MB average per row), this caused dashboard API responses of 9-11 seconds due to fetching ~1 GB of TOAST data per request.

Add ExcludePayloads bool to ExecutionFilter. When set, the query substitutes NULL AS input_payload, NULL AS result_payload so the column count stays identical and scanExecution requires no changes. Set ExcludePayloads: true on all six dashboard QueryExecutionRecords call sites.

Co-authored-by: Claude Sonnet 4.6 noreply@anthropic.com (5558f40)

Fixed

• Fix(storage): apply default idle connections when MaxIdleConns is unconfigured (#272)

The guard maxIdle < 0 never triggers for the zero-value of int (0), so SetMaxIdleConns(0) is called — telling database/sql to keep no idle connections. Every query opens and closes a fresh TCP connection to Postgres, adding ~30-150 ms of overhead per request and causing connection churn (~5 new backend PIDs/sec).

Changing to <= 0 ensures the default of 5 idle connections is applied when the setting is omitted, matching the existing maxOpen <= 0 guard on the line above.

Co-authored-by: Claude Opus 4.6 (1M context) noreply@anthropic.com (5b3e08c)

Other

• Revert "feat(harness): parse real cost from opencode JSON output (#269)" (#270)

This reverts commit b1a023dded6f03ea76487fa416aa077f17889429. (7377a72)

Added

• Feat(harness): parse real cost from opencode JSON output (#269)

• fix: capture stderr from Claude Code CLI for error diagnosis

The ClaudeCodeProvider was not passing a stderr callback to ClaudeAgentOptions, so when the claude CLI exited with code 1, the actual error message was lost. Logs only showed "Command failed with exit code 1" with no actionable details.

Now passes a stderr callback that collects output and includes it in both the error log and the RawResult.error_message field.

Co-Authored-By: Claude Opus 4.6 (1M context) noreply@anthropic.com

• fix: set stderr callback on opts object, not agent_options dict

Avoids test assertion failures caused by unexpected 'stderr' key in the agent_options dictionary.

Co-Authored-By: Claude Opus 4.6 (1M context) noreply@anthropic.com

• feat(harness): parse real cost from opencode JSON output

Use -f json flag when invoking opencode CLI and parse cost, prompt_tokens, and completion_tokens from the response. Falls back to estimate_cli_cost() for older opencode versions that don't include metrics in their JSON output.

Depends on: opencode-ai/opencode#TBD

Co-Authored-By: Claude Opus 4.6 (1M context) noreply@anthropic.com

---

Co-authored-by: Claude Opus 4.6 (1M context) noreply@anthropic.com (b1a023d)

Fixed

• Fix: capture stderr from Claude Code CLI for error diagnosis (#268)

• fix: capture stderr from Claude Code CLI for error diagnosis

The ClaudeCodeProvider was not passing a stderr callback to ClaudeAgentOptions, so when the claude CLI exited with code 1, the actual error message was lost. Logs only showed "Command failed with exit code 1" with no actionable details.

Now passes a stderr callback that collects output and includes it in both the error log and the RawResult.error_message field.

Co-Authored-By: Claude Opus 4.6 (1M context) noreply@anthropic.com

• fix: set stderr callback on opts object, not agent_options dict

Avoids test assertion failures caused by unexpected 'stderr' key in the agent_options dictionary.

Co-Authored-By: Claude Opus 4.6 (1M context) noreply@anthropic.com

---

Co-authored-by: Claude Opus 4.6 (1M context) noreply@anthropic.com (afcbeee)

Fixed

• Fix: prevent async executions from getting stuck in running state (#267)

Two issues caused async executions to remain in "running" state forever when the reasoner failed:

1. SDK: asyncio.create_task() return values were not stored, making fire-and-forget tasks eligible for GC before the status callback could be delivered. Now stored in a set with auto-cleanup via done_callback. Also increased callback timeout from 10s to 30s since the shared httpx client's default is too aggressive for concurrent status updates over internal networking.

2. CP: stale execution reaper ran every 1h with a 30m timeout (worst case ~90min to clean up). Reduced to 5m interval with 10m timeout so stuck executions are marked as timed-out within 15 minutes.

Co-authored-by: Claude Opus 4.6 noreply@anthropic.com (950b01c)

Fixed

• Fix(sdk): tune rate limiter defaults for fail-fast behavior (#265)

Reduce exponential backoff aggressiveness to prevent 2+ hour workflow runtimes when using rate-limited providers like OpenRouter. The previous defaults (20 retries, 300s max delay, 300s circuit breaker) caused cascading backoff that compounded across parallel agents.

New defaults: 5 retries, 0.5s base delay, 30s max delay, circuit breaker threshold 5 with 30s timeout. Max theoretical wait per call drops from ~100 minutes to ~2.5 minutes.

Changes:

• Python StatelessRateLimiter: max_retries 20→5, base_delay 1.0→0.5, max_delay 300→30, circuit_breaker_threshold 10→5, timeout 300→30
• TypeScript StatelessRateLimiter: identical parameter changes
• AIConfig: updated Field defaults to match rate limiter
• Added functional tests validating new defaults and max wait bounds

Co-authored-by: Claude Opus 4.6 noreply@anthropic.com (486ff3d)

Added

• Feat(sdk): surface cost_usd and usage from .ai() responses (#264)

• feat(sdk): surface cost_usd and usage from .ai() responses

MultimodalResponse now exposes cost_usd (estimated via litellm) and usage (token counts) extracted from litellm response objects. This enables downstream consumers like pr-af to track .ai() call costs instead of hardcoding them to zero.

Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com

• fix: remove unused pytest import to pass linting

Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com

---

Co-authored-by: Claude Opus 4.6 noreply@anthropic.com (3944cb6)

Documentation

• Docs: add UTM tracking to example project links in README

Replace direct GitHub links for SWE-AF, Deep Research, MongoDB, and sec-af with tracked redirects through agentfield.ai/github/* routes to measure README-driven traffic via Umami analytics. (96cbb77)

Fixed

• Fix: reap stale workflow executions and use updated_at for staleness (#262)

• fix: reap stale workflow executions and use updated_at for staleness detection

The existing MarkStaleExecutions only covered the executions table and used started_at to detect staleness, which missed orphaned workflow executions entirely and could incorrectly timeout legitimately long-running executions. This change:

• Switches staleness detection from started_at to updated_at so only executions with no recent activity are reaped
• Adds MarkStaleWorkflowExecutions to handle the workflow_executions table where orphaned child executions get permanently stuck in running state when their parent fails
• Wires both into the existing ExecutionCleanupService background loop

Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com

• test: add functional tests for stale execution reaper with real SQLite

Tests run against a real database (no mocks) covering:

• Stuck executions reaped while active ones are preserved
• Long-running executions with recent activity NOT incorrectly reaped
• Orphaned workflow children reaped when parent already failed
• Waiting-state executions reaped after inactivity
• Batch limit respected across multiple reaper passes
• End-to-end scenario: parent fails, children stuck in both tables

Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com

• fix: use COALESCE fallback for NULL updated_at in stale reaper queries

• Use COALESCE(updated_at, created_at, started_at) in both MarkStaleExecutions and MarkStaleWorkflowExecutions to handle rows where updated_at was never set
• Add invariant comment documenting that updated_at must be bumped on every meaningful activity for staleness detection to work
• Add tests for NULL updated_at scenario on both execution types

---

Co-authored-by: Claude Opus 4.6 noreply@anthropic.com Co-authored-by: Santosh santosh@agentfield.ai (56410a6)

Fixed

• Fix: update estimate_cli_cost for litellm v1.80+ API (#261)

• fix: update estimate_cli_cost for litellm v1.80+ API

litellm removed prompt_tokens/completion_tokens kwargs from completion_cost() in v1.80. Switch to prompt/completion string params which litellm tokenizes internally.

Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com

• fix: update cost estimation test for litellm v1.80+ API

The test was asserting token_counter calls and prompt_tokens/completion_tokens kwargs which were removed in the implementation fix. Update to match the new prompt/completion string params API.

Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com

---

Co-authored-by: Claude Opus 4.6 noreply@anthropic.com (4583e3c)

Added

• Feat: add token-based cost estimation for CLI harness providers (#260)

OpenCode, Gemini, and Codex providers now estimate LLM cost using litellm's pricing database, so HarnessResult.cost_usd is no longer always None for subprocess-based providers. This enables budget enforcement and cost reporting in downstream consumers like pr-af.

Co-authored-by: Claude Opus 4.6 noreply@anthropic.com (8b94345)

Fixed

• Fix: use status snapshot for node status endpoints to prevent flickering (#259)

GetNodeStatusHandler and BulkNodeStatusHandler were performing live HTTP health checks on every call (1s cache for active agents). With the UI polling every 3s, a single transient network failure in Railway would immediately return "offline", causing agent status to flicker. Now uses GetAgentStatusSnapshot which returns the stored status managed by the background HealthMonitor (which has proper 3-consecutive-failure debouncing and heartbeat gating). The explicit POST .../status/refresh endpoint remains available for on-demand live checks.

Co-authored-by: Claude Opus 4.6 noreply@anthropic.com (584b995)

Fixed

• Fix: wire ApplyEnvOverrides into server startup (#258)

• fix: wire ApplyEnvOverrides into server startup for Railway deployments

The applyEnvOverrides function (handling short env var names like AGENTFIELD_CONNECTOR_ENABLED) was never called from the actual server startup path. main.go uses Viper for config loading, but Viper's AutomaticEnv only matches keys it already knows about from config files. On Railway (no config file), ALL connector env vars were silently ignored, causing connector routes to never be registered.

Export ApplyEnvOverrides and call it after Viper unmarshal so env vars like AGENTFIELD_CONNECTOR_ENABLED, AGENTFIELD_CONNECTOR_TOKEN, and capability flags (AGENTFIELD_CONNECTOR_CAP_*) work on file-less deploys.

Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com

• feat: add connector status routes for list_nodes and get_node_status

The connector's status handler was calling /api/v1/nodes (a regular API endpoint requiring API key auth) instead of connector-scoped routes. Added /api/v1/connector/nodes and /api/v1/connector/nodes/:id/status routes gated by status_read capability, matching the pattern used by other connector domains.

Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com

---

Co-authored-by: Claude Opus 4.6 noreply@anthropic.com (c731519)

Fixed

• Fix: add config_management to connector capability env var map (#257)

The connectorCapEnvMap was missing the config_management capability, so AGENTFIELD_CONNECTOR_CAP_CONFIG_MANAGEMENT env var was silently ignored. This caused connector config routes to not be accessible when configured via environment variables (e.g. Railway deployments).

Co-authored-by: Claude Opus 4.6 noreply@anthropic.com (ff098c1)

Fixed

• Fix: skip global API key check for connector routes (#255)

Connector routes have their own dedicated ConnectorTokenAuth middleware that enforces X-Connector-Token with constant-time comparison. The global APIKeyAuth middleware was incorrectly requiring the API key on these routes too, forcing connectors to know and send the CP's global API key — a credential they should never need.

This adds a prefix skip for /api/v1/connector/ in APIKeyAuth, matching the existing pattern for /health, /ui, and /api/v1/did/ routes.

Also adds comprehensive functional tests for the full connector auth chain:

• ConnectorTokenAuth (valid/invalid/missing token, audit metadata injection)
• ConnectorCapabilityCheck (enabled/disabled/read-only/missing capabilities)
• Integration tests proving connector routes reject requests without a valid connector token, even though they bypass the global API key check

Co-authored-by: Claude Opus 4.6 noreply@anthropic.com (6d969a1)

Added

• Feat: database-backed configuration storage (#254)

• feat: database-backed configuration storage

Add ability to store and manage configuration files in the database instead of (or in addition to) YAML files on disk. This enables remote config management via the connector/SaaS flow.

• Add ConfigStorageModel with versioning and audit fields
• Implement SetConfig/GetConfig/ListConfigs/DeleteConfig in storage layer
• Add config CRUD API endpoints (GET/PUT/DELETE /api/v1/configs/:key)
• Add connector-scoped config routes gated by config_management capability
• Add AGENTFIELD_CONFIG_SOURCE=db flag to load config from database at startup
• Add Goose migration 028_create_config_storage.sql
• Works on both SQLite and PostgreSQL backends

Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com

• feat: add hot-reload endpoint for database-backed configuration

Adds POST /configs/reload endpoint that re-applies database config to the running control plane without requiring a process restart. Only active when AGENTFIELD_CONFIG_SOURCE=db is set.

Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com

• fix: merge DB config fields individually to prevent zeroing out defaults

The ExecutionCleanup struct was being replaced wholesale when only RetentionPeriod was set, zeroing out CleanupInterval and causing a panic (non-positive interval for NewTicker). Now merges each field individually. Also excludes connector config from DB merge since token and capabilities are security-sensitive.

Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com

• fix: address critical security and correctness issues in config storage

• Add 1MB body size limit to SetConfig to prevent DoS via unbounded reads
• Add sync.RWMutex to protect config during hot-reload (prevents data race)
• Replace fragile string error check with errors.Is(err, sql.ErrNoRows)

---

Co-authored-by: Claude Opus 4.6 noreply@anthropic.com Co-authored-by: Santosh santosh@agentfield.ai (7ac9c87)

Fixed

• Fix(install): BSD sed compatibility and env var pipe scoping in install.sh (#251)

Replace GNU-only \s with POSIX [[:space:]]* in get_latest_prerelease_version() sed regex — \s is not recognized by macOS BSD sed, causing the version string to contain raw JSON instead of just the tag name.

Fix documented VERSION/STAGING env var patterns: VAR=val cmd1 | cmd2 scopes VAR to cmd1 only (POSIX shell behavior), so bash never sees it. Corrected to: curl ... | VERSION=X bash

Fixes #250 (96c3ae9)

Fixed

• Fix: allow state transitions from stopping to active/starting

When a node restarts while the control plane still considers it "stopping", heartbeats get rejected causing the node to be stuck offline. Allow stopping → active/starting transitions so nodes can recover from this state.

Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com (d328e60)

Fixed

• Fix(sdk): catch Pydantic ValidationError in structured output parsing

Pydantic v2 ValidationError does not inherit from ValueError, so schema validation failures (e.g. missing required fields) were not caught by the retry logic. This caused LLM responses with incomplete JSON to crash the execution instead of retrying.

Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com (e2330d9)

Other

• Fix squished authorization table layout (#253)

• Fix squished authorization table layout

Widen the grid template columns for Status, Registered, and Actions so they don't overlap. Use flexible sizing for Registered and Actions columns to accommodate varying content widths.

Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com

• Fix authorization table: use fixed widths for right columns

The fr-based columns were consuming nearly all space, squeezing Status/Registered/Actions into a tiny area. Use fixed px widths for the right 3 columns (matching the pattern used by other tables) so they get space allocated first before fr columns divide the rest.

Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com

• Fix Registered column wrapping and Actions column alignment

• Add whitespace-nowrap to Registered time and action buttons so "21 hours ago" and "Approve Reject" stay on one line
• Widen Actions column to 160px to fit both buttons
• Add "Actions" header label so all 6 columns have visible headers

Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com

• Widen Actions column to 200px to fit Approve + Reject buttons

Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com

• Reduce left column fr values to stop hogging space from right columns

All three flexible columns now use 1fr instead of 2fr/1.5fr, giving equal weight and leaving more room for Status, Registered, and Actions.

Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com

---

Co-authored-by: Claude Opus 4.6 noreply@anthropic.com (4d534b9)

Added

• Feat: External Cancel/Pause/Resume Execution (Epic #238) (#246)

• feat(state-machine): add paused execution state and transitions (#239)

• Add ExecutionStatusPaused constant and aliases to pkg/types/status.go
• Add paused state transitions in execution_state_validation.go: running→paused, paused→running, paused→cancelled
• Update SQLite CHECK constraints in local.go
• Add PostgreSQL migration 027_add_paused_execution_status.sql
• Add ExecutionPaused/Resumed/Cancelled event types to event bus
• Add publish helpers for new event types
• Update deriveOverallStatus to handle paused workflows
• Update frontend CanonicalStatus type, theme, badge, hex colors for paused

Part of epic #238 — External Cancel/Pause Execution

• feat(api): add POST /executions/:id/cancel endpoint (#240)

• CancelExecutionHandler: updates execution + workflow execution to cancelled
• Emits ExecutionCancelled event via event bus
• Stores workflow execution event for audit trail
• Supports optional reason field in request body
• Returns previous_status, new status, and cancelled_at timestamp
• Rejects cancel on terminal states (409 Conflict)
• Returns 404 for non-existent executions
• Registers route under agentAPI group
• Comprehensive tests: state transitions, reason handling, edge cases

• feat(api): add POST /executions/:id/pause and /resume endpoints (#241)

• PauseExecutionHandler: transitions running -> paused
• ResumeExecutionHandler: transitions paused -> running
• Both update execution record + workflow execution atomically
• Emit ExecutionPaused/ExecutionResumed events via event bus
• Store workflow execution events for audit trail
• Support optional reason field in request body
• Strict state validation: pause only from running, resume only from paused
• Returns previous_status, new status, and paused_at/resumed_at
• Register routes under agentAPI group
• Comprehensive tests using existing testExecutionStorage helpers

• feat(cli): add af execution cancel|pause|resume commands (#244)

• NewExecutionCommand with cancel, pause, resume subcommands
• Shared executionActionOptions/executionActionConfig for DRY implementation
• Supports --server, --token, --timeout, --json, --reason flags
• Human-readable output by default, raw JSON with --json
• User-friendly error messages for 404/409 status codes
• Registered under RootCmd in root.go

• feat(executor): enforce cancel/pause state in DAG executor (#242)

• callAgent checks execution status before making HTTP call
• Cancelled executions skip agent call and return early
• Paused executions block in waitForResume using event bus pattern
• waitForResume unblocks on ExecutionResumed (return nil) or ExecutionCancelledEvent (return error)
• Race condition guard: checks status before subscribing to event bus
• asyncExecutionJob.process also checks status before callAgent
• Tests: cancel skip, pause+resume flow, pause+cancel flow, async job skip

• feat(ui): add cancel/pause/resume controls to workflow header (#243)

• Add Cancel, Pause, Resume buttons to EnhancedWorkflowHeader
• Cancel uses AlertDialog confirmation (destructive action guard)
• Pause/Resume use ghost buttons with amber/emerald hover states
• Mobile-responsive: icon-only on small screens, icon+label on desktop
• Loading spinners during mutations, all buttons disabled while mutating
• API client functions in executionsApi.ts for cancel/pause/resume
• Routes registered under UI API group in server.go
• NotificationProvider wraps workflow detail page for toast feedback
• New alert-dialog.tsx component (shadcn/Radix pattern)

• fix(storage): add missing 'waiting' status to SQLite and PostgreSQL CHECK constraints

The 'waiting' status (used by HITL approval flow) was a valid canonical status in Go code but was missing from database CHECK constraints. This would cause INSERT/UPDATE failures when executions transition to 'waiting' state.

Fixes both SQLite (local.go) and PostgreSQL (migration 027) constraints.

• fix(ui): add paused status to all CanonicalStatus Record maps

WorkflowNode, HoverDetailPanel, StatusSection, EnhancedWorkflowIdentity, and ExecutionHistoryList all had Record<CanonicalStatus, ...> maps missing the 'paused' entry, causing TypeScript build failures.

• refactor(ui): redesign cancel/pause/resume as icon-only toolbar buttons

Match existing toolbar convention (ghost variant, h-8 w-8, title tooltips). Remove text labels and destructive variant to reduce visual weight. Add separator between execution controls and view controls. Keeps AlertDialog confirmation for cancel safety.

• fix(ui): fix paused priority in deriveOverallStatus and move graph controls to floating toolbar

• Fix deriveOverallStatus() to prioritize paused over running (deliberate user action takes precedence over child execution state)
• Add 2 test cases for paused priority behavior
• Move viewMode (Standard/Performance/Debug) and Focus mode from header to bottom-left floating toolbar in graph view
• Update EnhancedWorkflowDetailPage prop wiring for toolbar migration

• feat(ui): add execution lifecycle controls, live duration, and status filter enhancements

• Redesign CompactExecutionHeader with pause/cancel/resume icon buttons, live elapsed time counter, refresh button, and hover card for secondary details (agent, DID, workflow, input/output sizes)
• Wrap EnhancedExecutionDetailPage with NotificationProvider for toast notifications on pause/cancel/resume actions
• Add live elapsed time display to EnhancedWorkflowHeader for running and paused workflows (replaces N/A with real-time counter)
• Add Paused and Cancelled to STATUS_FILTER_OPTIONS in PageHeader
• Add paused to statusLabels in CompactWorkflowsTable

• fix(ui): single-line headers and fix unicode triangle rendering

• Convert two-line name+subtitle layout to single-line in both workflow and execution detail page headers
• Replace HTML entity ▴ with Unicode ▲ (JSX compat)
• Replace • separators with Unicode middle dot (·)
• Remove unused Clock import from CompactExecutionHeader
• Name appears bold, metadata appears muted on same line

• fix(ui): human-readable durations, visible agent_node_id, LIVE badge → refresh dot

• Replace local formatDuration with shared formatDurationHumanReadable (e.g. 4487.0m → 3d 2h, 74h 43m → 3d 2h)
• Show agent_node_id as distinct mono chip next to reasoner name
• Remove standalone LIVE/IDLE badges from both headers
• Add green pulsing dot on refresh button when execution is live
• Clean up unused Clock import and underscore unused props

• fix(ui): show agent_node_id in workflow header, remove steps/depth clutter

• Extract root agent_node_id from DAG timeline data
• Display as mono chip next to workflow name (same style as execution page)
• Remove 'N steps · depth N' metadata (not useful to users)
• Keep duration with live indicator and run ID for copying
• Update mobile row to match desktop layout

• feat(ui): redesign execution and workflow headers into 2-row layout

Restructure both detail page headers from single-row into a semantically-organized 2-row layout with proper information hierarchy, responsive behavior, and mobile support.

Row 1: status cluster + identity cluster + lifecycle controls Row 2: section navigation tabs (absorbed from separate components) + summary metrics

• Rewrite CompactExecutionHeader with status dot, identity chips, tooltips, controlled cancel AlertDialog, and mobile overflow menu
• Rewrite EnhancedWorkflowHeader with webhook HoverCard, active/failed badges, fullscreen toggle, and graph depth metrics
• Move tab navigation into headers (remove standalone tab components)
• Add mobile 3-row stacked layout with DropdownMenu overflow
• Update both detail pages to pass tab props to headers

• fix(test): resolve data race in execution cleanup test

Use thread-safe syncBuffer for concurrent log writes from cleanup goroutine and Stop() goroutine. The bytes.Buffer is not safe for concurrent writes, causing race detector failures on CI.

• fix: address code review issues in cancel/pause/resume feature

• Add 24h timeout to waitForResume in async execution path to prevent goroutine leaks when resume/cancel events are never delivered
• Use unique subscriber IDs (with nanosecond suffix) to prevent event bus collisions when parallel DAG branches wait on same execution
• Change CancelExecutionHandler to accept ExecutionStore interface instead of storage.StorageProvider for interface segregation
• Fix pause response Reason field to use *string (matching cancel response) so empty reasons are omitted from JSON via omitempty
• Remove debug console.log from executionsApi.ts that leaked execution data to browser console in production

• feat(ui): redesign workflow DAG graph toolbar (#248)

• feat(ui): redesign workflow DAG toolbar with unified GraphToolbar component

Replace scattered graph controls (layout buttons, search, center, fit view, view mode toggle, focus mode) with a single compact icon-based toolbar.

• Add GraphToolbar component with layout/view mode dropdowns, search, focus, and smart center buttons using Phosphor icons with tooltips
• Implement wrapped tree layout in LayoutManager for wide-branching DAGs (wraps 50+ siblings into rows of N columns targeting ~1600px width)
• Remove duplicate mrtree layout (keep Dagre tree as single tree option)
• Extend WorkflowDAGControls with changeLayout() and onLayoutInfoChange
• Simplify VirtualizedDAG by removing layout-related props
• Remove SLOW warning badges, replace with subtle 'slower' text in dropdown
• Unify default layout to tree for all graph sizes

• fix(ui): clean up dead code and unused props from toolbar redesign

• Delete LayoutControls.tsx (fully replaced by GraphToolbar, zero imports remain)
• Remove unused isFullscreen prop from EnhancedWorkflowFlowProps and call site
• Clean blank line artifacts left from removed Panel blocks

• fix(handlers): move state validation inside update callbacks to eliminate TOCTOU race

The cancel/pause/resume handlers previously checked execution state before the atomic update callback, allowing concurrent requests to slip through. Now the state check happens inside the callback where it reads the locked current value, and state-conflict errors are properly mapped to 409 Conflict instead of 500.

Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com

---

Co-authored-by: Abir Abbas abirabbas1998@gmail.com Co-authored-by: Claude Opus 4.6 noreply@anthropic.com (56f7f5c)

Fixed

• Fix(ui): center sidebar nav icons when collapsed (#247)

Remove redundant px-2 from SidebarContent that stacked with SidebarGroup's built-in p-2, causing 32px of horizontal padding inside the 48px collapsed rail. The 32px icon buttons overflowed right, appearing right-justified instead of centered. (6c1eebb)

Fixed

• Fix: include API key in note() request headers (#235)

• fix: include API key in note() request headers

The note() method was sending execution context headers but not the X-API-Key, causing 401 when the control plane has API key auth enabled (production). Works locally because local dev typically has no API key configured.

Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com

• Fix test stub to include _get_auth_headers on client

The test_note_sends_async_request test was failing because the agent stub's client (SimpleNamespace) lacked the _get_auth_headers method added in the note auth fix. The _send_note coroutine calls self.client._get_auth_headers(), which raised AttributeError and silently prevented the HTTP post from executing.

Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com

---

Co-authored-by: Claude Opus 4.6 noreply@anthropic.com (087c2c6)

Other

• Revert "fix: include API key in note() request headers"

This reverts commit 94725ff34008e2fad19d778ec12470c213753168. (8091824)

Added

• Feat: add sec-af autonomous security audit to Built With examples

Adds the AI Security Auditor (sec-af) showcase to the README examples table with a custom editorial image and link to github.com/Agent-Field/sec-af.

• Add assets/examples/ai-security-auditor.png showcase image
• Update README Built With section with new entry, description, and GitHub link (4ee4b0a)

Fixed

• Fix: include API key in note() request headers

The note() method was sending execution context headers but not the X-API-Key, causing 401 when the control plane has API key auth enabled (production). Works locally because local dev typically has no API key configured.

Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com (94725ff)

Documentation

• Docs: add AI tool calling documentation to READMEs (#231)

Document the new native LLM tool-calling feature (PR #228) in the main README and all three SDK READMEs with examples showing auto-discovery, filtered discovery, lazy hydration, guardrails, and observability.

Co-authored-by: Claude Opus 4.6 noreply@anthropic.com (56bf930)

Added

• Feat: native LLM tool-calling support via discover → ai → call pipeline (#228)

• feat: add native LLM tool-calling support via discover → ai → call pipeline (#225)

Add tools= parameter to app.ai() that enables automatic tool-call loops: discover available capabilities, convert to LLM tool schemas, dispatch calls via app.call(), and feed results back until the LLM produces a final response.

Python SDK:

• New tool_calling module with capability-to-schema converters
• Tool-call execution loop with multi-turn support
• Progressive discovery (lazy schema hydration)
• Guardrails (max_turns, max_tool_calls)
• Per-call observability (ToolCallTrace with latency tracking)

Go SDK:

• Tool types (ToolDefinition, ToolCall) on ai.Request/Response
• CapabilitiesToToolDefinitions converter
• ExecuteToolCallLoop on ai.Client
• AIWithTools convenience method on Agent
• WithTools option for ai.Request

Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com

• fix(sdk/python): harden tool-calling DX — rate limiting, stream guard, typed response

• Wrap tool-calling LLM calls with rate limiter and model fallbacks (previously bypassed both, causing naked 429 failures in production)
• Guard stream=True + tools= with clear ValueError
• Type tools parameter with Union instead of Any for IDE discoverability
• Replace monkey-patched _tool_call_trace with typed ToolCallResponse wrapper (exposes .trace, .text, .response with getattr delegation)
• Track hydration_retries in ToolCallTrace for lazy hydration observability
• Add ToolCallResponse tests and update existing tests

Refs: #225, #229

• fix(test): update harness schema test to match #230 prompt wording change

• feat: add TS SDK tool-calling parity, lazy hydration, examples, and E2E-tested fixes

• TypeScript SDK: ToolCalling.ts with full discover/filter/lazy/guardrails pipeline
• TypeScript SDK: lazy hydration uses non-executable selection pass then hydrates
• TypeScript SDK: OpenRouter/Ollama use .chat() API (not Responses API)
• TypeScript SDK: ReasonerContext.aiWithTools() for ctx-level tool calling
• Python SDK: fix invocation_target→call_target conversion in tool dispatch
• Python/TS: tool name sanitization (colons→double underscores) for LLM compat
• Examples: Python + TS orchestrator/worker/test covering all Sam's #225 cases
• E2E tested with GPT-4o-mini and Gemini 2.0 Flash via OpenRouter

Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com

• fix(examples): use ctx.input and app.serve() in TS worker example

SkillHandler receives a single SkillContext arg — input lives on ctx.input, not as a second parameter. Also fix app.run() → app.serve() to match the TS SDK's actual API. Found during E2E manual testing.

Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com

---

Co-authored-by: Claude Opus 4.6 noreply@anthropic.com Co-authored-by: Santosh santosh@agentfield.ai (40638d0)

Fixed

• Fix(harness): add concurrency limiter, stdout fallback, and stronger prompts (#230)

Root cause: unbounded concurrent opencode subprocess calls (20+) overwhelm the LLM API, causing transient failures where output files are never created.

Changes:

• opencode.py: add global asyncio.Semaphore (default 3, configurable via OPENCODE_MAX_CONCURRENT) to throttle concurrent opencode run processes; add 600s timeout; add structured logging for finish/error states
• _schema.py: strengthen output prompt to emphasize Write tool usage; add try_parse_from_text() fallback that extracts JSON from LLM stdout when the output file is missing (fenced blocks, brace matching, cosmetic repair)
• _runner.py: wire stdout fallback after parse_and_validate in both initial and retry paths

Validated: full SEC-AF pipeline (11 hunt strategies, 30 verified findings) completes end-to-end with 0 enricher failures, vs repeated failures before. (2947d5b)

Fixed

• Fix(did): add did:web resolution to document endpoint (#227)

• fix(did): add did:web resolution to /api/v1/did/document/:did endpoint

The GetDIDDocument handler only resolved did:key identities via the in-memory registry. did:web lookups returned "DID not found" even when the agent had a valid did:web document stored in the database.

Add did:web resolution (via didWebService) before falling back to did:key, matching the pattern already used by the ResolveDID handler and the server's serveDIDDocument method.

Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com

• fix(did): normalize URL-decoded %3A in did:web path params

Gin URL-decodes path parameters, turning did:web:localhost%3A8080:agents:foo into did:web:localhost:8080:agents:foo. The database stores the canonical form with %3A, so lookups failed with "DID not found".

Add normalizeDIDWeb() helper that detects decoded port separators and re-encodes them. Applied to both ResolveDID and GetDIDDocument handlers.

Manually verified against running control plane:

• /api/v1/did/document/did:web:... → 200 with W3C DID Document
• /api/v1/did/resolve/did:web:... → 200 with DID resolution result
• did:key paths unchanged (no regression)

Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com

---

Co-authored-by: Claude Opus 4.6 noreply@anthropic.com (cdf4e8b)

Added

• Feat(ui): display both did:key and did:web identities (#226)

• feat(ui): display both did:key and did:web identities with clear distinction

The UI previously only showed did:key identifiers, making did:web identities invisible to users who need them for JWT and external integrations.

Backend: Wire DIDWebService into UI DIDHandler and return did_web in the node DID API response.

Frontend: Show both identity types as clearly separated sections with descriptive labels — "Cryptographic Identity" (did:key) for signing/auth, and "Web Identity" (did:web) for JWT/external use. Each has its own copy button and View Document action.

Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com

• fix(ui): add did:web to identity API and fix unused import

Wire DIDWebService into IdentityHandlers so the DID Explorer page returns did_web alongside did:key. Remove unused Analytics import that was breaking CI builds.

Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com

---

Co-authored-by: Claude Opus 4.6 noreply@anthropic.com (8ffdc28)

Added

• Feat(ui): improve duration display in workflow and execution tables (#223)

• feat(ui): improve duration display in workflow and execution tables

• Add formatDurationHumanReadable() utility for human-readable durations (e.g., '1h 2m' instead of '3748.3s')
• Add LiveElapsedDuration component that ticks every second for running items instead of showing static '-' dash
• Update WorkflowsTable, CompactWorkflowsTable, CompactExecutionsTable, and EnhancedExecutionsTable to use new duration formatting
• Fix 'as any' type assertion in CompactExecutionsTable status prop

Closes #222

• docs: add screenshots for duration display PR (aa02ea1)

Added

• Feat(ai): retry LLM calls on malformed structured output JSON (#224)

When using schema-based structured output, LLMs occasionally return malformed JSON that fails parsing. This adds automatic retry (up to 2 retries) specifically for parse failures, avoiding unnecessary retries for network or API errors.

Co-authored-by: Claude Opus 4.6 noreply@anthropic.com (a462b3a)

Added

• Feat(harness): OpenCode support with schema retry, error preservation, and project_dir routing (#220)

• feat(harness): add schema output diagnosis and enhanced follow-up prompts

Add diagnose_output_failure() that classifies validation failures into specific categories: file missing, empty, invalid JSON, or schema mismatch with field-level diff. Enhance build_followup_prompt() to include schema file references and explicit rewrite instructions for the retry loop.

• feat(harness): add schema validation retry loop with session continuity

Replace single-shot _handle_schema_output() with _handle_schema_with_retry() that retries up to schema_max_retries times (default 2) when JSON validation fails. Each retry:

• Diagnoses the specific failure via diagnose_output_failure()
• Sends a follow-up prompt to the agent with error context
• For Claude: passes resume=session_id to continue the conversation
• For CLI providers: fresh call with the follow-up prompt
• Accumulates cost, turns, and messages across all attempts

This activates the previously dead-code build_followup_prompt() from _schema.py and adds resume_session_id support to the Claude Code provider.

• test(harness): add complex JSON schema debug test script

Standalone script exercising the harness with 5 escalating schema levels:

• simple (2 fields), medium (lists + optionals), complex (13 nested fields), deeply_nested (recursive TreeNode), massive (>4K tokens, file-based path) Tested live with both claude-code and codex providers — all levels pass. Includes manual retry test mode (--retry-test) to exercise the new retry loop.

• feat(harness): add opencode provider with project_dir routing

• Rewrite opencode.py: auto-managed serve+attach pattern to bypass opencode v1.2.10-v1.2.16 'Session not found' bug
• Add project_dir field to HarnessConfig (types.py) so coding agents explore the target repo instead of a temp working directory
• Add output file placement inside project_dir (runner) so sandboxed Write tool can reach the output JSON
• Pass server_url to OpenCodeProvider via factory
• Clean up debug prints from runner and claude provider
• Verified working with openrouter/moonshotai/kimi-k2.5 model

• fix(harness): update opencode provider tests for serve+attach pattern

Tests now pass server_url to skip auto-serve lifecycle in CI where opencode binary is not installed. Asserts updated to match --attach command structure.

• fix(harness): use direct opencode run for auto-approve permissions

opencode run --attach loses auto-approve because the serve process treats attached sessions as interactive, causing permission prompts to hang forever when the model tries to write files.

• fix(harness): align opencode tests with direct run (no --attach)

• fix(harness): crash-safe retry with FailureType classification

• Add FailureType enum (NONE, CRASH, TIMEOUT, API_ERROR, SCHEMA, NO_OUTPUT) to RawResult and HarnessResult for intelligent retry decisions
• Fix returncode masking in run_cli: preserve negative signal values
• Add strip_ansi() to clean ANSI escape codes from stderr
• Crash-aware retry in _handle_schema_with_retry: retryable failures (CRASH, NO_OUTPUT) get full prompt re-send instead of immediate bail
• build_followup_prompt now accepts optional schema param, inlines schema JSON, removed poisonous empty-array hint that caused flat schema failures
• Exponential backoff between schema retries (0.5s base, 5s max)
• Apply same crash classification pattern to opencode, codex, gemini providers
• Update opencode provider test for XDG_DATA_HOME env injection

• fix(harness): remove double-close bug and dead serve code

• write_schema_file: remove try/except that double-closed fd after os.fdopen() already took ownership (caused EBADF on write failure)
• opencode.py: remove ~100 lines of unused serve+attach machinery (_ensure_serve, _cleanup_serve, _find_free_port, class-level singleton state) — execute() uses direct opencode run and never called any of it

Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com

---

Co-authored-by: Abir Abbas abirabbas1998@gmail.com Co-authored-by: Claude Opus 4.6 noreply@anthropic.com (909038b)

• Feat(readme): replace text examples table with visual showcase cards (#216)

Add 3-column visual 'Built With AgentField' section with premium editorial images for Autonomous Engineering Team, Deep Research Engine, and Reactive MongoDB Intelligence. Moved higher in README (after 'What is AgentField?') for better visibility. Removed old text-only Production Examples table. (b9add36)

Chores

• Chore(readme): remove redundant horizontal rules

GitHub already renders ## headings with visual separation. The 11 --- rules created double-spacing that made the README choppy. (48baf65)

Added

• Feat(webhook): support all HITL template response formats (#213)

• feat(webhook): support all HITL template response formats and multi-pause workflows

The webhook approval handler previously only extracted the "decision" field from template responses, causing templates that use "action" (confirm-action, rich-text-editor) or have no explicit decision field (signature-capture) to fail silently.

Changes:

• Extract decision from "action" field as fallback when "decision" is absent
• Default completed webhooks with no decision/action to "approved"
• Add normalizeDecision() to map template-specific values (approve, confirm, reject, deny, abort, cancel) to canonical set (approved/rejected)
• Clear approval request fields on "approved" (not just "request_changes") to support multi-pause workflows where agents issue sequential approvals
• Add localhost:8001 to CORS allowed origins for demo UIs

Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com

• fix: preserve ApprovalRequestID on approved for idempotency and multi-pause

The previous commit cleared ApprovalRequestID on both "approved" and "request_changes" decisions. This broke:

• Idempotent webhook retries (lookup by request ID returned 404)
• Approval-status queries (same lookup failure)
• Callback notifications (in-memory store shared pointer was mutated)

Fix:

• Only clear ApprovalRequestID on "request_changes" (as before)
• On "approved", clear URL fields but preserve the request ID
• Save callback URL before the update closure to avoid shared-pointer aliasing in stores that mutate objects in-place
• Make request-approval handler multi-pause aware: check ApprovalStatus is "pending" (not just ApprovalRequestID existence) so agents can re-request approval after a resolved round

Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com

---

Co-authored-by: Claude Opus 4.6 noreply@anthropic.com (aa15d64)

Added

• Feat(harness): add .harness() method for external coding agent dispatch (#210)

• docs: add harness v2 design document with file-write schema strategy

Design document for .harness() feature — first-class coding agent integration. Covers architecture, provider matrix, universal file-write schema handling with 4-layer recovery, config types, and implementation phases.

Ref: #208

• feat(harness): add core types, provider interface, and factory skeleton (#199)

• Add HarnessConfig to types.py (provider required, sensible defaults)
• Add HarnessResult, RawResult, Metrics result types
• Add HarnessProvider protocol (Python) / interface (TypeScript)
• Add build_provider() factory with supported provider validation
• Python: 8 tests passing, TypeScript: 6 tests passing

Closes #199

• feat(harness): add schema handling with universal file-write strategy (#200)

• Universal file-write: always instruct agent to write JSON to {cwd}/.agentfield_output.json
• Prompt suffix generation (inline for small schemas, file-based for large >4K tokens)
• Cosmetic JSON repair: strip markdown fences, trailing commas, truncated brackets
• Full parse+validate pipeline with Layer 1 (direct) + Layer 2 (repair) fallback
• Pydantic v1/v2 + Zod schema support
• Python: 19 tests, TypeScript: 18 tests

Closes #200

• feat(harness): add HarnessRunner with retry and schema orchestration (#201)

• Config resolution: merge HarnessConfig defaults + per-call overrides
• Exponential backoff + jitter retry for transient errors (rate limits, 5xx, timeouts)
• Schema orchestration: prompt suffix injection, Layer 1+2 parse/validate
• Guaranteed temp file cleanup in finally block
• Cost/metrics/session tracking in HarnessResult

Closes #201

• feat(harness): add Claude Code and Codex providers with shared CLI utilities (#202, #203)

• Claude Code provider: Python uses claude_agent_sdk (lazy import), TS uses @anthropic-ai/claude-agent-sdk (dynamic import)
• Codex provider: Python + TS use CLI subprocess with shared async utilities
• Shared CLI module: run_cli, parse_jsonl, extract_final_text for subprocess management
• All providers implement HarnessProvider protocol with execute() method
• 14 Python tests + 12 TypeScript tests, all passing (97 total)

• feat(harness): wire .harness() into Agent class with lazy runner (#204)

• Python: harness_config constructor param, lazy harness_runner property, async harness() method
• TypeScript: harnessConfig in AgentConfig, lazy getHarnessRunner(), async harness() method
• Package exports: HarnessConfig + HarnessResult from agentfield.init
• 8 Python + 6 TypeScript wiring tests, all passing (111 total)

• feat(harness): add Gemini CLI and OpenCode providers (#205, #206)

• Gemini provider: CLI subprocess with -p prompt, --sandbox auto, -m model flags
• OpenCode provider: CLI subprocess with --non-interactive, --model flags
• Factory updated to route all 4 providers: claude-code, codex, gemini, opencode
• Provider exports updated in Python + TypeScript
• 10 Python + 10 TypeScript new tests, all passing (131 total)

• fix(harness): address review feedback — lazy imports, trimmed exports, file permissions

• Remove eager provider imports from Python providers/init.py (lazy loading preserved via factory)
• Trim public API exports in Python harness/init.py and TypeScript harness/index.ts to only public types
• Add 0o600 file permissions for schema/output files in _schema.py and schema.ts
• Fix TypeScript type errors in runner.ts (tsc --noEmit clean)

• feat(harness): add functional tests, fix provider bugs for Claude and OpenCode

• Add 12 Python + 6 TypeScript functional tests invoking real coding agents
• Fix Claude Code permission_mode mapping (auto → bypassPermissions)
• Fix OpenCode CLI command (--non-interactive → run subcommand)
• Mark OpenCode tests as xfail (upstream v1.2.10 headless bug)
• Add harness_live pytest marker, excluded from default runs
• Update unit test expectations for provider command changes

Tested: Codex 4/4 ✅, Claude Code 4/4 ✅, cross-provider ✅ OpenCode: upstream 'Session not found' bug (not our code)

• perf(harness): fix import/allocation regression — lazy imports, WeakMap, async factory

• TS: Replace eager HarnessRunner import with dynamic import() in Agent.ts
• TS: Use WeakMap instead of class property for _harnessRunner (keeps V8 inline)
• TS: Make buildProvider async with per-provider dynamic imports in factory.ts
• Python: Move HarnessRunner import to TYPE_CHECKING + lazy import in property
• Update all 8 test files for async buildProvider/getHarnessRunner changes
• All 131 unit tests passing (69 Python + 62 TypeScript)
• tsc --noEmit clean

• fix(harness): use typing.List for Python 3.8 compat in functional test

Pydantic evaluates annotations at runtime via eval(), so list[str] (PEP 585) fails on Python 3.8 even with 'from future import annotations'. Use typing.List[str] instead.

• feat(harness): add optional 'harness' and 'harness-claude' extras in pyproject.toml

Users can now install the Claude Code SDK dependency declaratively: pip install agentfield[harness] # all harness provider deps pip install agentfield[harness-claude] # just Claude Code SDK

Codex, Gemini, and OpenCode are CLI binaries — no pip packages needed.

---

Co-authored-by: Abir Abbas abirabbas1998@gmail.com Co-authored-by: Claude Opus 4.6 noreply@anthropic.com (ef1fac5)

Fixed

• Fix: allow empty input for parameterless skills/reasoners (#198)

• fix: allow empty input for parameterless skills/reasoners (#196)

Remove binding:"required" constraint on Input field in ExecuteRequest and ExecuteReasonerRequest structs. Gin interprets required on maps as "must be present AND non-empty", which rejects the valid {"input":{}} payload that SDKs send for parameterless calls.

Also remove the explicit len(req.Input)==0 check in prepareExecution and add nil-input guards in the reasoner and skill handlers to match the existing pattern in execute.go.

Closes #196

• test: strengthen empty-input handler coverage

• fix: update empty_input_test.go for ExecuteHandler signature change

Main added an internalToken parameter to ExecuteHandler in PR #197. Update the two test call sites to pass empty string for the new param.

Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com

---

Co-authored-by: Abir Abbas abirabbas1998@gmail.com Co-authored-by: Claude Opus 4.6 noreply@anthropic.com (cbdc23a)

Documentation

• Docs: add human-in-the-loop approval docs to SDK READMEs (#212)

Add approval workflow documentation with code examples to all three SDK READMEs (Python, TypeScript, Go), covering the waiting state feature for pausing agent execution pending human review.

Co-authored-by: Claude Opus 4.6 noreply@anthropic.com (88f24cf)

Testing

• Test(control-plane): add execution cleanup service coverage (#195) (5a227cf)

Added

• Feat: waiting state with approval workflows, VC-based authorization, and multi-version reasoners (#197)

• feat(control-plane): add VC-based authorization foundation

This commit introduces the foundation for the new VC-based authorization system that replaces API key distribution with admin-approved permissions.

Key components added:

• Architecture documentation (docs/VC_AUTHORIZATION_ARCHITECTURE.md)
• Database migrations for permission approvals, DID documents, and protected agents
• Core types for permissions and did:web support
• DIDWebService for did:web generation, storage, and resolution
• PermissionService for permission requests, approvals, and VC issuance

The system enables:

• Agents self-assigning tags (identity declaration)
• Admin approval workflow for protected agent access
• Real-time revocation via did:web
• Control plane as source of truth for approvals

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

• feat(vc-authorization): complete VC-based authorization system implementation

• Add DID authentication middleware with Ed25519 signature verification
• Add permission checking middleware for protected agent enforcement
• Implement admin API handlers for permission management (approve/reject/revoke)
• Add permission request and check API endpoints
• Implement storage layer for DID documents, permission approvals, protected agent rules
• Add comprehensive integration test suite (14 test functions covering all phases)
• Add Admin UI pages: PendingPermissions, PermissionHistory, ProtectedAgents
• Add Go SDK DID authentication support
• Add Python SDK DID authentication support
• Fix CI to enable FTS5 tests (previously all SQLite-dependent tests were skipped)
• Add security documentation for DID authentication
• Add implementation guide documentation

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

• fix(control-plane): fix pre-existing test bugs exposed by FTS5 build tag

TestGetNodeDetailsHandler_Structure expected HTTP 400 for missing route param but Gin returns 404. TestGetNodeStatusHandler_Structure was missing a mock expectation for GetAgentStatus causing a panic.

Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com

• fix(control-plane): fix pre-existing test bugs exposed by FTS5 build tag

The CI workflow change from go test ./... to go test -tags sqlite_fts5 ./... caused previously-skipped tests to execute, revealing 15 pre-existing bugs:

• UI handler tests: Register agents in storage and configure mocks for GetAgentStatus calls; fix assertions to match actual behavior (health check failures mark agents inactive, not error the request)
• VC service tests: Fix GetWorkflowVC lookups to use workflow_vc_id not workflow_id; fix issuer mismatch test to tamper VCDocument JSON instead of metadata field; fix error message assertion for empty VC documents
• VC storage tests: Fix GetWorkflowVC key lookups; fix empty result assertions
• PresenceManager tests: Register agents in storage so markInactive -> UpdateAgentStatus -> GetAgentStatusSnapshot -> GetAgent succeeds; add proper sync.Mutex for callback vars; use require.Eventually instead of time.Sleep; set HardEvictTTL for lease deletion test
• Webhook storage: Fix hardcoded Pending status to use webhook.Status
• Execution records test: Fix LatestStarted assertion (CreateExecutionRecord overwrites updated_at with time.Now())
• Cleanup test: Wire countWorkflowRuns and deleteWorkflowRuns into workflow cleanup path

Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com

• fix(control-plane): fix SSE tests leaking goroutines via incorrect context cancellation

Multiple SSE tests called req.Context().Done() expecting it to cancel the context, but Done() only returns a channel — it doesn't cancel anything. This caused SSE handler goroutines to block forever, leaking and eventually causing a 10-minute test timeout in CI.

Fixed all affected tests to use context.WithCancel + explicit cancel() call, matching the pattern already used by the working SSE tests.

Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com

• ts sdk and bug fix on did web

• feat(examples): add permission test agents and enable VC authorization config

Add two example agents for manually testing the VC authorization system end-to-end: permission-agent-a (caller) and permission-agent-b (protected target). Enable authorization in the default config with seeded protection rules.

Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com

• Fixes

• fix(sdk-python): update test fakes for DID credential wiring in _register_agent_with_did

The previous commit added identity_package access and client credential wiring to _register_agent_with_did but didn't update the test fakes. _FakeDIDManager now provides a realistic identity_package and _FakeAgentFieldClient supports set_did_credentials, so the full registration path is exercised in tests.

Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com

• more improvements

• 6th iteration of fixes

• end to end tested

• feat(sdk): add Go & TS permission test agents, fix DID auth signing

• Add Go permission test agents (caller + protected target with 3 reasoners)
• Add TS permission test agents (caller + tag-protected target with VC generation)
• Fix TS SDK DID auth: pass pre-serialized JSON string to axios to ensure signed bytes match what's sent on the wire
• Fix Python SDK test for async execution manager payload serialization change
• Add go-perm-target protection rule to config
• Gitignore compiled Go agent binaries

Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com

• fix(sdk-ts): update header-forwarding test for pre-serialized JSON body

The execute() method now passes a JSON string instead of an object to axios for DID auth signing consistency. Update test assertion to match.

Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com

• manual testing updates

• fix(vc-auth): fix re-approval deadlock, empty caller_agent_id, and error propagation

• Fix re-approval deadlock: expand auto-request condition to trigger for revoked/rejected statuses, not just empty (permission.go)
• Fix empty caller_agent_id: add DID registry fallback in ResolveAgentIDByDID for did:key resolution (did_service.go, did_web_service.go)
• Fix HTTP 200 for failed executions: return 502 with proper error details when inner agent-to-agent calls fail (execute.go)
• Fix error propagation across all 3 SDKs:
  • Go SDK: add ExecuteError type preserving status code and error_details
  • TS SDK: propagate err.responseData as error_details in all error handlers
  • Python SDK: add ExecuteError class, extract JSON body from 4xx responses instead of losing it via raise_for_status(), propagate error_details in async callback payloads

Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com

• fix go missing func

• address dx changes

• temp

• more fixes

• finalized

• better error prop

• fix: update TS DID auth tests to match nonce-based signing format

Tests expected the old 3-header format ({timestamp}:{bodyHash}) but the implementation correctly uses 4 headers with nonce ({timestamp}:{nonce}:{bodyHash}), matching Go and Python SDKs.

Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com

• fix: add rate limiting to DID auth middleware on execution endpoints

Addresses code scanning alert about missing rate limiting on the authorization route handler. Adds a sliding-window rate limiter (30 requests per IP per 60s) to the local verification middleware.

Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com

• fix: use express-rate-limit for DID auth middleware to satisfy CodeQL

Replace custom Map-based rate limiter with express-rate-limit package, which CodeQL recognizes as a proper rate limiting implementation.

Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com

• fix: remove duplicate countWorkflowRuns method from rebase

Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com

• UI cleanup

• pydantic formatting fix

• connector changes

• implemented multi agents with versioning

• feat(ui): polished authorization page with unified tabs and visual standardization

Replace separate TagApprovalPage and AccessPoliciesPage with a single tabbed AuthorizationPage. Add polished authorization components:

• AccessRulesTab: 48px rows, sorted policies, ALLOW/DENY border colors
• AgentTagsTab: all agents with tag data, sorted, neutral badges
• ApproveWithContextDialog: tag selection with policy impact preview
• PolicyFormDialog: chip-input for tags with known-tag suggestions
• PolicyContextPanel: shows affected policies for selected tags
• RevokeDialog: neutral styling, optional reason
• ChipInput, TooltipTagList: reusable tag UI components

Backend additions:

• GET /api/ui/v1/authorization/agents: returns all agents with tag data
• GET /api/v1/admin/tags: returns all known tags from agents & policies

Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com

• multi versioning connector setup

• add agent to agent direct checks

• bugfixes on connector

• QA fixes

• package lock

• bug fixes on permissions & versioning flow

• fix: add missing DeleteAgentVersion stub and guard postgres migration for fresh DBs

Two CI failures:

1. linux-tests: stubStorage in server_routes_test.go was missing the DeleteAgentVersion method added to the StorageProvider interface by the multi-version work. Add the stub.

2. Functional Tests (postgres): migrateAgentNodesCompositePKPostgres tried to ALTER TABLE agent_nodes before GORM created it on fresh databases. The information_schema.columns query returns count=0 (not an error) when the table doesn't exist, so the function proceeded to run ALTER statements against a nonexistent table. Add an explicit table existence check matching the pattern already used by the SQLite migration path.

Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com

• add postgres testing to dev

• wait flow

• improvements

• bugfix on reasoner path

• reasoner name mismatch fix

• fix skill name mismatch bug

• fix: update test to include approval_expires_at column

The merge brought in a test from main that expected 42 columns in the workflow execution insert query, but the feature branch added approval_expires_at as the 43rd column. Update the test's column list and expected placeholder count to match.

Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com

• fix: remove unused httpx import in test_approval.py

Ruff lint flagged the unused import (F401). The tests use httpx_mock fixture from pytest-httpx, not httpx directly.

Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com

• fix: resolve Python and TypeScript SDK test failures

Python SDK:

• Add pytest-httpx dependency (with Python >=3.10 constraint)
• Register httpx_mock marker for --strict-markers compatibility
• Add importorskip for graceful skip on Python <3.10
• Fix request_approval test calls to match actual API signature

TypeScript SDK:

• Call server.closeAllConnections() before server.close() in afterEach to prevent keep-alive connection timeout in tests

Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com

• fix: update test URL to match reasoner name-based endpoint path

After the reasoner name fix, @agent.reasoner(name="reports_generate") registers at /reasoners/reports_generate (the explicit name), not /reasoners/generate_report (the function name).

Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com

• add examples for waiting state

• fix: resolve Gin route parameter conflict between waiting-state and tag-vc endpoints

The waiting-state feature added routes under /api/v1/agents/:node_id/... which conflicted with the existing tag-vc endpoint using :agentId as the parameter name. Gin requires consistent wildcard names for the same path segment, causing a panic on server startup.

Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com

• fix tests

• fix: correct async endpoint URLs and assertion in waiting state functional tests

Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com

---

Co-authored-by: Claude Opus 4.5 noreply@anthropic.com Co-authored-by: Santosh santosh@agentfield.ai (414f91c)

Added

• Feat: VC-based authorization, sidecar management APIs, and multi-version reasoners (#188)

• feat(control-plane): add VC-based authorization foundation

This commit introduces the foundation for the new VC-based authorization system that replaces API key distribution with admin-approved permissions.

Key components added:

• Architecture documentation (docs/VC_AUTHORIZATION_ARCHITECTURE.md)
• Database migrations for permission approvals, DID documents, and protected agents
• Core types for permissions and did:web support
• DIDWebService for did:web generation, storage, and resolution
• PermissionService for permission requests, approvals, and VC issuance

The system enables:

• Agents self-assigning tags (identity declaration)
• Admin approval workflow for protected agent access
• Real-time revocation via did:web
• Control plane as source of truth for approvals

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

• feat(vc-authorization): complete VC-based authorization system implementation

• Add DID authentication middleware with Ed25519 signature verification
• Add permission checking middleware for protected agent enforcement
• Implement admin API handlers for permission management (approve/reject/revoke)
• Add permission request and check API endpoints
• Implement storage layer for DID documents, permission approvals, protected agent rules
• Add comprehensive integration test suite (14 test functions covering all phases)
• Add Admin UI pages: PendingPermissions, PermissionHistory, ProtectedAgents
• Add Go SDK DID authentication support
• Add Python SDK DID authentication support
• Fix CI to enable FTS5 tests (previously all SQLite-dependent tests were skipped)
• Add security documentation for DID authentication
• Add implementation guide documentation

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

• fix(control-plane): fix pre-existing test bugs exposed by FTS5 build tag

TestGetNodeDetailsHandler_Structure expected HTTP 400 for missing route param but Gin returns 404. TestGetNodeStatusHandler_Structure was missing a mock expectation for GetAgentStatus causing a panic.

Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com

• fix(control-plane): fix pre-existing test bugs exposed by FTS5 build tag

The CI workflow change from go test ./... to go test -tags sqlite_fts5 ./... caused previously-skipped tests to execute, revealing 15 pre-existing bugs:

• UI handler tests: Register agents in storage and configure mocks for GetAgentStatus calls; fix assertions to match actual behavior (health check failures mark agents inactive, not error the request)
• VC service tests: Fix GetWorkflowVC lookups to use workflow_vc_id not workflow_id; fix issuer mismatch test to tamper VCDocument JSON instead of metadata field; fix error message assertion for empty VC documents
• VC storage tests: Fix GetWorkflowVC key lookups; fix empty result assertions
• PresenceManager tests: Register agents in storage so markInactive -> UpdateAgentStatus -> GetAgentStatusSnapshot -> GetAgent succeeds; add proper sync.Mutex for callback vars; use require.Eventually instead of time.Sleep; set HardEvictTTL for lease deletion test
• Webhook storage: Fix hardcoded Pending status to use webhook.Status
• Execution records test: Fix LatestStarted assertion (CreateExecutionRecord overwrites updated_at with time.Now())
• Cleanup test: Wire countWorkflowRuns and deleteWorkflowRuns into workflow cleanup path

Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com

• fix(control-plane): fix SSE tests leaking goroutines via incorrect context cancellation

Multiple SSE tests called req.Context().Done() expecting it to cancel the context, but Done() only returns a channel — it doesn't cancel anything. This caused SSE handler goroutines to block forever, leaking and eventually causing a 10-minute test timeout in CI.

Fixed all affected tests to use context.WithCancel + explicit cancel() call, matching the pattern already used by the working SSE tests.

Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com

• ts sdk and bug fix on did web

• feat(examples): add permission test agents and enable VC authorization config

Add two example agents for manually testing the VC authorization system end-to-end: permission-agent-a (caller) and permission-agent-b (protected target). Enable authorization in the default config with seeded protection rules.

Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com

• Fixes

• fix(sdk-python): update test fakes for DID credential wiring in _register_agent_with_did

The previous commit added identity_package access and client credential wiring to _register_agent_with_did but didn't update the test fakes. _FakeDIDManager now provides a realistic identity_package and _FakeAgentFieldClient supports set_did_credentials, so the full registration path is exercised in tests.

Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com

• more improvements

• 6th iteration of fixes

• end to end tested

• feat(sdk): add Go & TS permission test agents, fix DID auth signing

• Add Go permission test agents (caller + protected target with 3 reasoners)
• Add TS permission test agents (caller + tag-protected target with VC generation)
• Fix TS SDK DID auth: pass pre-serialized JSON string to axios to ensure signed bytes match what's sent on the wire
• Fix Python SDK test for async execution manager payload serialization change
• Add go-perm-target protection rule to config
• Gitignore compiled Go agent binaries

Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com

• fix(sdk-ts): update header-forwarding test for pre-serialized JSON body

The execute() method now passes a JSON string instead of an object to axios for DID auth signing consistency. Update test assertion to match.

Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com

• manual testing updates

• fix(vc-auth): fix re-approval deadlock, empty caller_agent_id, and error propagation

• Fix re-approval deadlock: expand auto-request condition to trigger for revoked/rejected statuses, not just empty (permission.go)
• Fix empty caller_agent_id: add DID registry fallback in ResolveAgentIDByDID for did:key resolution (did_service.go, did_web_service.go)
• Fix HTTP 200 for failed executions: return 502 with proper error details when inner agent-to-agent calls fail (execute.go)
• Fix error propagation across all 3 SDKs:
  • Go SDK: add ExecuteError type preserving status code and error_details
  • TS SDK: propagate err.responseData as error_details in all error handlers
  • Python SDK: add ExecuteError class, extract JSON body from 4xx responses instead of losing it via raise_for_status(), propagate error_details in async callback payloads

Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com

• fix go missing func

• address dx changes

• temp

• more fixes

• finalized

• better error prop

• fix: update TS DID auth tests to match nonce-based signing format

Tests expected the old 3-header format ({timestamp}:{bodyHash}) but the implementation correctly uses 4 headers with nonce ({timestamp}:{nonce}:{bodyHash}), matching Go and Python SDKs.

Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com

• fix: add rate limiting to DID auth middleware on execution endpoints

Addresses code scanning alert about missing rate limiting on the authorization route handler. Adds a sliding-window rate limiter (30 requests per IP per 60s) to the local verification middleware.

Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com

• fix: use express-rate-limit for DID auth middleware to satisfy CodeQL

Replace custom Map-based rate limiter with express-rate-limit package, which CodeQL recognizes as a proper rate limiting implementation.

Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com

• fix: remove duplicate countWorkflowRuns method from rebase

Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com

• UI cleanup

• pydantic formatting fix

• connector changes

• implemented multi agents with versioning

• feat(ui): polished authorization page with unified tabs and visual standardization

Replace separate TagApprovalPage and AccessPoliciesPage with a single tabbed AuthorizationPage. Add polished authorization components:

• AccessRulesTab: 48px rows, sorted policies, ALLOW/DENY border colors
• AgentTagsTab: all agents with tag data, sorted, neutral badges
• ApproveWithContextDialog: tag selection with policy impact preview
• PolicyFormDialog: chip-input for tags with known-tag suggestions
• PolicyContextPanel: shows affected policies for selected tags
• RevokeDialog: neutral styling, optional reason
• ChipInput, TooltipTagList: reusable tag UI components

Backend additions:

• GET /api/ui/v1/authorization/agents: returns all agents with tag data
• GET /api/v1/admin/tags: returns all known tags from agents & policies

Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com

• multi versioning connector setup

• add agent to agent direct checks

• bugfixes on connector

• QA fixes

• package lock

• bug fixes on permissions & versioning flow

• fix: add missing DeleteAgentVersion stub and guard postgres migration for fresh DBs

Two CI failures:

1. linux-tests: stubStorage in server_routes_test.go was missing the DeleteAgentVersion method added to the StorageProvider interface by the multi-version work. Add the stub.

2. Functional Tests (postgres): migrateAgentNodesCompositePKPostgres tried to ALTER TABLE agent_nodes before GORM created it on fresh databases. The information_schema.columns query returns count=0 (not an error) when the table doesn't exist, so the function proceeded to run ALTER statements against a nonexistent table. Add an explicit table existence check matching the pattern already used by the SQLite migration path.

Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com

• add postgres testing to dev

• docs: add changelog and env vars for connector, versioning, and authorization

Document the feat/connector release including multi-versioning, VC-based authorization, and connector subsystem in CHANGELOG.md. Add authorization and connector environment variable sections to ENVIRONMENT_VARIABLES.md.

Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com

---

Co-authored-by: Claude Opus 4.5 noreply@anthropic.com Co-authored-by: Santosh santosh@agentfield.ai (917b49b)

Fixed

• Fix(release): add [skip ci] to version bump commit to prevent infinite loop (#194)

The release workflow pushes a version bump commit to main, which triggers another release workflow run, creating an infinite loop. Adding [skip ci] to the commit message prevents the pushed commit from triggering any workflows.

Co-authored-by: Claude Opus 4.6 noreply@anthropic.com (ff0a88f)