github.com/Agent-Field/agentfield

⚠️ This is a staging/pre-release version for testing. Not recommended for production use.

AgentField 0.1.134-rc.1

Installation

# Staging binary (use --staging flag)
curl -fsSL https://agentfield.ai/install.sh | bash -s -- --staging

# Python SDK (prerelease - requires --pre flag)
pip install --pre agentfield

# TypeScript SDK
npm install @agentfield/sdk@next

Version Pinning

VERSION=v0.1.134-rc.1 curl -fsSL https://agentfield.ai/install.sh | bash

Manual Download

Download the binary for your platform below, make it executable, and move it to your PATH.

Supported Platforms

• macOS Intel: agentfield-darwin-amd64
• macOS Apple Silicon: agentfield-darwin-arm64
• Linux x86_64: agentfield-linux-amd64
• Linux ARM64: agentfield-linux-arm64

What's Changed

• chore(deps): bump js-yaml from 4.3.0 to 4.3.1 in /desktop in the npm_and_yarn group across 1 directory by @dependabot[bot] in #892
• Fix macOS Go scaffold happy path by @AbirAbbas in #894
• fix(control-plane): make the execution cleanup initial delay testable by @0xHerm in #897
• fix(sdk/python): replace bare asyncio.run() with loop-aware helpers (#620) by @7vignesh in #899
• feat(harness): aforge provider + per-run isolation of the schema output file by @AbirAbbas in #891
• fix(sdk/python): resolve ResultCache cross-loop deadlock (#623) by @7vignesh in #799
• test: add coverage for GinLogger middleware (#557) by @santoshkumarradha in #895
• chore(deps): bump js-yaml to 4.3.1 by @santoshkumarradha in #898
• fix(sdk/python): run destructor cleanup synchronously when no loop is running (#620 follow-up) by @AbirAbbas in #902
• fix(harness): bound CLI output capture (16MB/stream, env-tunable) by @AbirAbbas in #903
• chore(deps): bump nanoid to 3.3.17 by @santoshkumarradha in #900
• fix(go-sdk): allow custom HTTP clients for AI requests by @mikemikimike in #908
• feat(harness): add Grok Build CLI provider by @storyicon in #861
• fix(sdk/python): add lock timeouts + offload blocking requests fallback (#620) by @7vignesh in #904
• feat(sdk/go): triggers package + OnEvent/OnSchedule sugar (#513) by @7vignesh in #906
• fix(sdk/python): harden cross-loop teardown by @mikemikimike in #907
• Seamless agent use: cloud-aware skill, native Codex install, self-healing skill sync by @AbirAbbas in #910
• feat(sdk/go): dispatch envelope unwrap + Context injection (#514) by @7vignesh in #914
• feat(sdk/go): trigger test helpers + fixture library (#515) by @7vignesh in #915
• feat(sdk/go): add Agent.Span for traced in-process sub-executions by @AbirAbbas in #916
• fix: repair broken quickstart flows — Go docs/examples, docker scaffolds, TS template by @AbirAbbas in #918
• fix(ui): only prompt for the admin token when the server enforces one by @AbirAbbas in #920
• feat(examples): triggers-demo-go end-to-end demo + Go skill docs (#516) by @7vignesh in #917
• fix(sdk): preserve multimodal args with user prompt by @mikemikimike in #923
• fix: resolve high-severity npm dependency alerts by @santoshkumarradha in #922
• feat(install): provision the aforge harness binary alongside af (curl, desktop, docker) by @AbirAbbas in #924
• feat(harness): make aforge the default provider across Python, Go, and TypeScript by @AbirAbbas in #905
• fix(sdk/python): exclude litellm 1.97.0 (crashes app.ai() on Python 3.10) by @AbirAbbas in #925
• fix(install): make --no-aforge actually skip the aforge download by @AbirAbbas in #926
• docs: feature Person to Brief AF by @santoshkumarradha in #912
• fix(harness): staging installs put aforge beside af; doctor exits 0 on a survey; doctor knows grok by @AbirAbbas in #927
• fix(harness): attribute default-provider runs to aforge, set HarnessResult.model, pass max_turns to aforge by @AbirAbbas in #928
• fix(sdk/go): attribute default-provider harness runs to aforge, prefer result.Model by @AbirAbbas in #929
• fix(sdk/typescript): pass claude-code options to claude-agent-sdk in camelCase by @AbirAbbas in #931
• fix(sdk/go,sdk/typescript): gemini argv — drop the nonexistent -C flag, map permission modes like Python by @AbirAbbas in #932
• fix(sdk/typescript): opencode surfaces exit-0 stderr failures and signals as errors by @AbirAbbas in #933
• fix(sdk/typescript): convert zod 4 schemas with the native toJSONSchema by @AbirAbbas in #935
• fix(sdk): send X-OpenRouter-Categories on OpenRouter requests by @santoshkumarradha in #930
• fix(sdk/typescript): honour AGENTFIELD_SERVER for the default control-plane URL by @AbirAbbas in #934
• fix(install): disable VCS stamping when building Go agent nodes by @AbirAbbas in #938
• Ship the desktop app with bundled agent nodes, missing-key surfacing, and the offload-by-default skill by @santoshkumarradha in #937
• fix(assets): correct garbled text in README features-strip banner by @AbirAbbas in #940
• Bundle sec-af and cloudsecurity-af with the desktop app by @AbirAbbas in #939
• fix(control-plane): stop counting agent restarts as failed executions by @santoshkumarradha in #941
• fix(ci): stop functional-test runs from reporting to production telemetry by @AbirAbbas in #942
• fix(execute): resolve webhook not triggering on async status callback (#936) by @7vignesh in #943
• fix(execute): propagate agent 4xx in async lane instead of blanket 502 by @7vignesh in #945
• fix(go-sdk): preserve structured AI API errors by @mikemikimike in #948
• fix(execute): make terminal status callbacks idempotent, reject cross-terminal rewrites by @AbirAbbas in #951
• fix(storage): row-lock postgres read-modify-write updates — lost-update race stalled builds by @AbirAbbas in #953

New Contributors

• @0xHerm made their first contribution in #897
• @mikemikimike made their first contribution in #908
• @storyicon made their first contribution in #861

Full Changelog: v0.1.126-rc.1...v0.1.134-rc.1

Fixed

• Fix(storage): row-lock read-modify-write updates on postgres (lost-update race) (#953)

UpdateExecutionRecord and the workflow-execution updater are BEGIN -> SELECT -> mutate in Go -> UPDATE every column -> COMMIT. Under postgres READ COMMITTED two overlapping updaters both read the same snapshot and the later COMMIT writes its stale copy back, discarding the earlier one. Observed in production (Railway template, postgres mode): an execution-note write racing the terminal status callback reverted a reasoner's status from succeeded back to running with a NULL result — the SDK's poll loop never saw the reasoner finish and the parent build hung until its 6h timeout. SQLite never shows this because it serializes writers.

Append FOR UPDATE to the RMW SELECTs in postgres mode only (SQLite does not accept the syntax and does not need it): a concurrent updater now blocks until the first transaction commits, then re-reads the committed row. Single-row primary-key locks, so no new deadlock ordering.

The regression test reproduces the exact interleaving deterministically: it fails on the pre-fix code with 'note was lost to the concurrent status write' and passes with the lock (verified against postgres 14; the test skips unless POSTGRES_TEST_URL is set, like the other postgres tests).

Co-authored-by: Claude Fable 5 [email protected] (e55a626)

AgentField 0.1.133

Installation

curl -fsSL https://agentfield.ai/install.sh | bash

Version Pinning

VERSION=v0.1.133 curl -fsSL https://agentfield.ai/install.sh | bash

Manual Download

Download the binary for your platform below, make it executable, and move it to your PATH.

Supported Platforms

• macOS Intel: agentfield-darwin-amd64
• macOS Apple Silicon: agentfield-darwin-arm64
• Linux x86_64: agentfield-linux-amd64
• Linux ARM64: agentfield-linux-arm64

What's Changed

• fix(assets): correct garbled text in README features-strip banner by @AbirAbbas in #940
• Bundle sec-af and cloudsecurity-af with the desktop app by @AbirAbbas in #939
• fix(control-plane): stop counting agent restarts as failed executions by @santoshkumarradha in #941
• fix(ci): stop functional-test runs from reporting to production telemetry by @AbirAbbas in #942
• fix(execute): resolve webhook not triggering on async status callback (#936) by @7vignesh in #943
• fix(execute): propagate agent 4xx in async lane instead of blanket 502 by @7vignesh in #945
• fix(go-sdk): preserve structured AI API errors by @mikemikimike in #948
• fix(execute): make terminal status callbacks idempotent, reject cross-terminal rewrites by @AbirAbbas in #951

Full Changelog: v0.1.132...v0.1.133

⚠️ This is a staging/pre-release version for testing. Not recommended for production use.

AgentField 0.1.133-rc.5

Installation

# Staging binary (use --staging flag)
curl -fsSL https://agentfield.ai/install.sh | bash -s -- --staging

# Python SDK (prerelease - requires --pre flag)
pip install --pre agentfield

# TypeScript SDK
npm install @agentfield/sdk@next

Version Pinning

VERSION=v0.1.133-rc.5 curl -fsSL https://agentfield.ai/install.sh | bash

Manual Download

Download the binary for your platform below, make it executable, and move it to your PATH.

Supported Platforms

• macOS Intel: agentfield-darwin-amd64
• macOS Apple Silicon: agentfield-darwin-arm64
• Linux x86_64: agentfield-linux-amd64
• Linux ARM64: agentfield-linux-arm64

What's Changed

• chore(deps): bump js-yaml from 4.3.0 to 4.3.1 in /desktop in the npm_and_yarn group across 1 directory by @dependabot[bot] in #892
• Fix macOS Go scaffold happy path by @AbirAbbas in #894
• fix(control-plane): make the execution cleanup initial delay testable by @0xHerm in #897
• fix(sdk/python): replace bare asyncio.run() with loop-aware helpers (#620) by @7vignesh in #899
• feat(harness): aforge provider + per-run isolation of the schema output file by @AbirAbbas in #891
• fix(sdk/python): resolve ResultCache cross-loop deadlock (#623) by @7vignesh in #799
• test: add coverage for GinLogger middleware (#557) by @santoshkumarradha in #895
• chore(deps): bump js-yaml to 4.3.1 by @santoshkumarradha in #898
• fix(sdk/python): run destructor cleanup synchronously when no loop is running (#620 follow-up) by @AbirAbbas in #902
• fix(harness): bound CLI output capture (16MB/stream, env-tunable) by @AbirAbbas in #903
• chore(deps): bump nanoid to 3.3.17 by @santoshkumarradha in #900
• fix(go-sdk): allow custom HTTP clients for AI requests by @mikemikimike in #908
• feat(harness): add Grok Build CLI provider by @storyicon in #861
• fix(sdk/python): add lock timeouts + offload blocking requests fallback (#620) by @7vignesh in #904
• feat(sdk/go): triggers package + OnEvent/OnSchedule sugar (#513) by @7vignesh in #906
• fix(sdk/python): harden cross-loop teardown by @mikemikimike in #907
• Seamless agent use: cloud-aware skill, native Codex install, self-healing skill sync by @AbirAbbas in #910
• feat(sdk/go): dispatch envelope unwrap + Context injection (#514) by @7vignesh in #914
• feat(sdk/go): trigger test helpers + fixture library (#515) by @7vignesh in #915
• feat(sdk/go): add Agent.Span for traced in-process sub-executions by @AbirAbbas in #916
• fix: repair broken quickstart flows — Go docs/examples, docker scaffolds, TS template by @AbirAbbas in #918
• fix(ui): only prompt for the admin token when the server enforces one by @AbirAbbas in #920
• feat(examples): triggers-demo-go end-to-end demo + Go skill docs (#516) by @7vignesh in #917
• fix(sdk): preserve multimodal args with user prompt by @mikemikimike in #923
• fix: resolve high-severity npm dependency alerts by @santoshkumarradha in #922
• feat(install): provision the aforge harness binary alongside af (curl, desktop, docker) by @AbirAbbas in #924
• feat(harness): make aforge the default provider across Python, Go, and TypeScript by @AbirAbbas in #905
• fix(sdk/python): exclude litellm 1.97.0 (crashes app.ai() on Python 3.10) by @AbirAbbas in #925
• fix(install): make --no-aforge actually skip the aforge download by @AbirAbbas in #926
• docs: feature Person to Brief AF by @santoshkumarradha in #912
• fix(harness): staging installs put aforge beside af; doctor exits 0 on a survey; doctor knows grok by @AbirAbbas in #927
• fix(harness): attribute default-provider runs to aforge, set HarnessResult.model, pass max_turns to aforge by @AbirAbbas in #928
• fix(sdk/go): attribute default-provider harness runs to aforge, prefer result.Model by @AbirAbbas in #929
• fix(sdk/typescript): pass claude-code options to claude-agent-sdk in camelCase by @AbirAbbas in #931
• fix(sdk/go,sdk/typescript): gemini argv — drop the nonexistent -C flag, map permission modes like Python by @AbirAbbas in #932
• fix(sdk/typescript): opencode surfaces exit-0 stderr failures and signals as errors by @AbirAbbas in #933
• fix(sdk/typescript): convert zod 4 schemas with the native toJSONSchema by @AbirAbbas in #935
• fix(sdk): send X-OpenRouter-Categories on OpenRouter requests by @santoshkumarradha in #930
• fix(sdk/typescript): honour AGENTFIELD_SERVER for the default control-plane URL by @AbirAbbas in #934
• fix(install): disable VCS stamping when building Go agent nodes by @AbirAbbas in #938
• Ship the desktop app with bundled agent nodes, missing-key surfacing, and the offload-by-default skill by @santoshkumarradha in #937
• fix(assets): correct garbled text in README features-strip banner by @AbirAbbas in #940
• Bundle sec-af and cloudsecurity-af with the desktop app by @AbirAbbas in #939
• fix(control-plane): stop counting agent restarts as failed executions by @santoshkumarradha in #941
• fix(ci): stop functional-test runs from reporting to production telemetry by @AbirAbbas in #942
• fix(execute): resolve webhook not triggering on async status callback (#936) by @7vignesh in #943
• fix(execute): propagate agent 4xx in async lane instead of blanket 502 by @7vignesh in #945
• fix(go-sdk): preserve structured AI API errors by @mikemikimike in #948
• fix(execute): make terminal status callbacks idempotent, reject cross-terminal rewrites by @AbirAbbas in #951

New Contributors

• @0xHerm made their first contribution in #897
• @mikemikimike made their first contribution in #908
• @storyicon made their first contribution in #861

Full Changelog: v0.1.126-rc.1...v0.1.133-rc.5

Fixed

• Fix(execute): make terminal status callbacks idempotent, reject cross-terminal rewrites (#951)

A status callback that re-delivers an execution's final status (e.g. an SDK retrying after a lost 200) was accepted and re-ran every side effect: the record was rewritten (result, duration_ms, completed_at), the webhook was re-notified, and another completed/failed lifecycle event was published to the execution bus — double-counting the execution for every bus consumer (SSE clients, tracing, product telemetry).

Worse, the terminal guard only rejected terminal→non-terminal writes despite its comment, so a late duplicate callback could flip "succeeded" to "failed" (or the reverse) after the outcome had already been observed, and emit an execution_failed event for an execution that had already completed.

• terminal→same-terminal: acknowledged with 200 as a strict no-op — no record rewrite, no webhook, no usage ingestion, no lifecycle event.
• terminal→different-terminal: rejected with 409 via the new errTerminalStatusConflict sentinel so bounded SDK retries fail fast instead of reading as a server fault.
• terminal→non-terminal: unchanged (still rejected; pinned by TestUpdateExecutionStatusHandler_TerminalRegression).

Verified live against an isolated control plane wired through the real telemetry relay: on main, a completed execution receiving two duplicate succeeded callbacks and one failed callback emitted three execution_completed events plus one execution_failed and left the record flipped to failed; with this fix the same sequence emits exactly one execution_completed, answers the duplicates with 200 no-ops and the flip with 409, and the record stays intact.

Co-authored-by: Claude Fable 5 [email protected] (ea897a0)

⚠️ This is a staging/pre-release version for testing. Not recommended for production use.

AgentField 0.1.133-rc.4

Installation

# Staging binary (use --staging flag)
curl -fsSL https://agentfield.ai/install.sh | bash -s -- --staging

# Python SDK (prerelease - requires --pre flag)
pip install --pre agentfield

# TypeScript SDK
npm install @agentfield/sdk@next

Version Pinning

VERSION=v0.1.133-rc.4 curl -fsSL https://agentfield.ai/install.sh | bash

Manual Download

Download the binary for your platform below, make it executable, and move it to your PATH.

Supported Platforms

• macOS Intel: agentfield-darwin-amd64
• macOS Apple Silicon: agentfield-darwin-arm64
• Linux x86_64: agentfield-linux-amd64
• Linux ARM64: agentfield-linux-arm64

What's Changed

• chore(deps): bump js-yaml from 4.3.0 to 4.3.1 in /desktop in the npm_and_yarn group across 1 directory by @dependabot[bot] in #892
• Fix macOS Go scaffold happy path by @AbirAbbas in #894
• fix(control-plane): make the execution cleanup initial delay testable by @0xHerm in #897
• fix(sdk/python): replace bare asyncio.run() with loop-aware helpers (#620) by @7vignesh in #899
• feat(harness): aforge provider + per-run isolation of the schema output file by @AbirAbbas in #891
• fix(sdk/python): resolve ResultCache cross-loop deadlock (#623) by @7vignesh in #799
• test: add coverage for GinLogger middleware (#557) by @santoshkumarradha in #895
• chore(deps): bump js-yaml to 4.3.1 by @santoshkumarradha in #898
• fix(sdk/python): run destructor cleanup synchronously when no loop is running (#620 follow-up) by @AbirAbbas in #902
• fix(harness): bound CLI output capture (16MB/stream, env-tunable) by @AbirAbbas in #903
• chore(deps): bump nanoid to 3.3.17 by @santoshkumarradha in #900
• fix(go-sdk): allow custom HTTP clients for AI requests by @mikemikimike in #908
• feat(harness): add Grok Build CLI provider by @storyicon in #861
• fix(sdk/python): add lock timeouts + offload blocking requests fallback (#620) by @7vignesh in #904
• feat(sdk/go): triggers package + OnEvent/OnSchedule sugar (#513) by @7vignesh in #906
• fix(sdk/python): harden cross-loop teardown by @mikemikimike in #907
• Seamless agent use: cloud-aware skill, native Codex install, self-healing skill sync by @AbirAbbas in #910
• feat(sdk/go): dispatch envelope unwrap + Context injection (#514) by @7vignesh in #914
• feat(sdk/go): trigger test helpers + fixture library (#515) by @7vignesh in #915
• feat(sdk/go): add Agent.Span for traced in-process sub-executions by @AbirAbbas in #916
• fix: repair broken quickstart flows — Go docs/examples, docker scaffolds, TS template by @AbirAbbas in #918
• fix(ui): only prompt for the admin token when the server enforces one by @AbirAbbas in #920
• feat(examples): triggers-demo-go end-to-end demo + Go skill docs (#516) by @7vignesh in #917
• fix(sdk): preserve multimodal args with user prompt by @mikemikimike in #923
• fix: resolve high-severity npm dependency alerts by @santoshkumarradha in #922
• feat(install): provision the aforge harness binary alongside af (curl, desktop, docker) by @AbirAbbas in #924
• feat(harness): make aforge the default provider across Python, Go, and TypeScript by @AbirAbbas in #905
• fix(sdk/python): exclude litellm 1.97.0 (crashes app.ai() on Python 3.10) by @AbirAbbas in #925
• fix(install): make --no-aforge actually skip the aforge download by @AbirAbbas in #926
• docs: feature Person to Brief AF by @santoshkumarradha in #912
• fix(harness): staging installs put aforge beside af; doctor exits 0 on a survey; doctor knows grok by @AbirAbbas in #927
• fix(harness): attribute default-provider runs to aforge, set HarnessResult.model, pass max_turns to aforge by @AbirAbbas in #928
• fix(sdk/go): attribute default-provider harness runs to aforge, prefer result.Model by @AbirAbbas in #929
• fix(sdk/typescript): pass claude-code options to claude-agent-sdk in camelCase by @AbirAbbas in #931
• fix(sdk/go,sdk/typescript): gemini argv — drop the nonexistent -C flag, map permission modes like Python by @AbirAbbas in #932
• fix(sdk/typescript): opencode surfaces exit-0 stderr failures and signals as errors by @AbirAbbas in #933
• fix(sdk/typescript): convert zod 4 schemas with the native toJSONSchema by @AbirAbbas in #935
• fix(sdk): send X-OpenRouter-Categories on OpenRouter requests by @santoshkumarradha in #930
• fix(sdk/typescript): honour AGENTFIELD_SERVER for the default control-plane URL by @AbirAbbas in #934
• fix(install): disable VCS stamping when building Go agent nodes by @AbirAbbas in #938
• Ship the desktop app with bundled agent nodes, missing-key surfacing, and the offload-by-default skill by @santoshkumarradha in #937
• fix(assets): correct garbled text in README features-strip banner by @AbirAbbas in #940
• Bundle sec-af and cloudsecurity-af with the desktop app by @AbirAbbas in #939
• fix(control-plane): stop counting agent restarts as failed executions by @santoshkumarradha in #941
• fix(ci): stop functional-test runs from reporting to production telemetry by @AbirAbbas in #942
• fix(execute): resolve webhook not triggering on async status callback (#936) by @7vignesh in #943
• fix(execute): propagate agent 4xx in async lane instead of blanket 502 by @7vignesh in #945
• fix(go-sdk): preserve structured AI API errors by @mikemikimike in #948

New Contributors

• @0xHerm made their first contribution in #897
• @mikemikimike made their first contribution in #908
• @storyicon made their first contribution in #861

Full Changelog: v0.1.126-rc.1...v0.1.133-rc.4

Fixed

• Fix(go-sdk): preserve structured AI API errors (#948)

• fix(go-sdk): preserve structured AI API errors

• test(go-sdk): cover non-2xx error edges (1af112c)

⚠️ This is a staging/pre-release version for testing. Not recommended for production use.

AgentField 0.1.133-rc.3

Installation

# Staging binary (use --staging flag)
curl -fsSL https://agentfield.ai/install.sh | bash -s -- --staging

# Python SDK (prerelease - requires --pre flag)
pip install --pre agentfield

# TypeScript SDK
npm install @agentfield/sdk@next

Version Pinning

VERSION=v0.1.133-rc.3 curl -fsSL https://agentfield.ai/install.sh | bash

Manual Download

Download the binary for your platform below, make it executable, and move it to your PATH.

Supported Platforms

• macOS Intel: agentfield-darwin-amd64
• macOS Apple Silicon: agentfield-darwin-arm64
• Linux x86_64: agentfield-linux-amd64
• Linux ARM64: agentfield-linux-arm64

What's Changed

• chore(deps): bump js-yaml from 4.3.0 to 4.3.1 in /desktop in the npm_and_yarn group across 1 directory by @dependabot[bot] in #892
• Fix macOS Go scaffold happy path by @AbirAbbas in #894
• fix(control-plane): make the execution cleanup initial delay testable by @0xHerm in #897
• fix(sdk/python): replace bare asyncio.run() with loop-aware helpers (#620) by @7vignesh in #899
• feat(harness): aforge provider + per-run isolation of the schema output file by @AbirAbbas in #891
• fix(sdk/python): resolve ResultCache cross-loop deadlock (#623) by @7vignesh in #799
• test: add coverage for GinLogger middleware (#557) by @santoshkumarradha in #895
• chore(deps): bump js-yaml to 4.3.1 by @santoshkumarradha in #898
• fix(sdk/python): run destructor cleanup synchronously when no loop is running (#620 follow-up) by @AbirAbbas in #902
• fix(harness): bound CLI output capture (16MB/stream, env-tunable) by @AbirAbbas in #903
• chore(deps): bump nanoid to 3.3.17 by @santoshkumarradha in #900
• fix(go-sdk): allow custom HTTP clients for AI requests by @mikemikimike in #908
• feat(harness): add Grok Build CLI provider by @storyicon in #861
• fix(sdk/python): add lock timeouts + offload blocking requests fallback (#620) by @7vignesh in #904
• feat(sdk/go): triggers package + OnEvent/OnSchedule sugar (#513) by @7vignesh in #906
• fix(sdk/python): harden cross-loop teardown by @mikemikimike in #907
• Seamless agent use: cloud-aware skill, native Codex install, self-healing skill sync by @AbirAbbas in #910
• feat(sdk/go): dispatch envelope unwrap + Context injection (#514) by @7vignesh in #914
• feat(sdk/go): trigger test helpers + fixture library (#515) by @7vignesh in #915
• feat(sdk/go): add Agent.Span for traced in-process sub-executions by @AbirAbbas in #916
• fix: repair broken quickstart flows — Go docs/examples, docker scaffolds, TS template by @AbirAbbas in #918
• fix(ui): only prompt for the admin token when the server enforces one by @AbirAbbas in #920
• feat(examples): triggers-demo-go end-to-end demo + Go skill docs (#516) by @7vignesh in #917
• fix(sdk): preserve multimodal args with user prompt by @mikemikimike in #923
• fix: resolve high-severity npm dependency alerts by @santoshkumarradha in #922
• feat(install): provision the aforge harness binary alongside af (curl, desktop, docker) by @AbirAbbas in #924
• feat(harness): make aforge the default provider across Python, Go, and TypeScript by @AbirAbbas in #905
• fix(sdk/python): exclude litellm 1.97.0 (crashes app.ai() on Python 3.10) by @AbirAbbas in #925
• fix(install): make --no-aforge actually skip the aforge download by @AbirAbbas in #926
• docs: feature Person to Brief AF by @santoshkumarradha in #912
• fix(harness): staging installs put aforge beside af; doctor exits 0 on a survey; doctor knows grok by @AbirAbbas in #927
• fix(harness): attribute default-provider runs to aforge, set HarnessResult.model, pass max_turns to aforge by @AbirAbbas in #928
• fix(sdk/go): attribute default-provider harness runs to aforge, prefer result.Model by @AbirAbbas in #929
• fix(sdk/typescript): pass claude-code options to claude-agent-sdk in camelCase by @AbirAbbas in #931
• fix(sdk/go,sdk/typescript): gemini argv — drop the nonexistent -C flag, map permission modes like Python by @AbirAbbas in #932
• fix(sdk/typescript): opencode surfaces exit-0 stderr failures and signals as errors by @AbirAbbas in #933
• fix(sdk/typescript): convert zod 4 schemas with the native toJSONSchema by @AbirAbbas in #935
• fix(sdk): send X-OpenRouter-Categories on OpenRouter requests by @santoshkumarradha in #930
• fix(sdk/typescript): honour AGENTFIELD_SERVER for the default control-plane URL by @AbirAbbas in #934
• fix(install): disable VCS stamping when building Go agent nodes by @AbirAbbas in #938
• Ship the desktop app with bundled agent nodes, missing-key surfacing, and the offload-by-default skill by @santoshkumarradha in #937
• fix(assets): correct garbled text in README features-strip banner by @AbirAbbas in #940
• Bundle sec-af and cloudsecurity-af with the desktop app by @AbirAbbas in #939
• fix(control-plane): stop counting agent restarts as failed executions by @santoshkumarradha in #941
• fix(ci): stop functional-test runs from reporting to production telemetry by @AbirAbbas in #942
• fix(execute): resolve webhook not triggering on async status callback (#936) by @7vignesh in #943
• fix(execute): propagate agent 4xx in async lane instead of blanket 502 by @7vignesh in #945

New Contributors

• @0xHerm made their first contribution in #897
• @mikemikimike made their first contribution in #908
• @storyicon made their first contribution in #861

Full Changelog: v0.1.126-rc.1...v0.1.133-rc.3

Fixed

• Fix(execute): propagate agent 4xx in async lane instead of blanket 502 (#862) (#945)

The async-completion branch in handleSync hardcoded HTTP 502 for all failed executions, making client-input rejections (e.g. 422) indistinguishable from upstream outages. The sync lane correctly propagated the agent's 4xx via writeExecutionError, but the async lane — which reads the execution from DB after the agent calls back — had no way to recover the original HTTP status.

Root cause: the callError.statusCode is lost once the execution is persisted; the async lane only sees the stored record.

Fix (two-pronged):

Control plane:

• Add error_status_code field to executionStatusUpdateRequest so the SDK can forward the HTTP status in its failure callback.
• When error_status_code is 4xx, encode it in StatusReason as "agent_client_error:<code>" for persistence.
• Add httpStatusForFailedExecution() helper that resolves the HTTP status from StatusReason (encoded client errors, known categories) and ErrorMessage ("agent error (NNN):" pattern fallback).
• Replace hardcoded 502 in the async-completion branch with the helper.

Python SDK:

• In _execute_async_with_callback failure path, propagate error_status_code from exception.status_code / exception.code when the value is a valid HTTP status (400-599).

Backward-compatible: SDKs that don't send error_status_code continue to get the existing behavior (502 default). SDKs that do send it get correct 4xx propagation immediately. (ad79ab5)

Fixed

• Fix(execute): resolve webhook not triggering on async status callback (#936) (#943)

The WebhookRegistered field on types.Execution carries db:"-" and is never persisted to the database. This caused two failures:

1. Status callback path: when an async execution completes via the agent's PUT /executions/:id/status callback, the handler checked updated.WebhookRegistered (always false from DB) and therefore never triggered webhook delivery.

2. GET /executions/:id: renderStatus read the same unpersisted field, so it always reported webhook_registered=false to polling clients.

Fix: add HasExecutionWebhook to the ExecutionStore interface and use it in both the status-callback trigger path and renderStatusWithApproval to resolve the webhook state from the execution_webhooks table.

Includes regression tests for both paths. (db3c285)

• Fix(ci): stop functional-test runs from reporting to production telemetry (#942)

Every functional-test run starts the control plane on a fresh /data volume, which mints a new anonymous install ID. With telemetry enabled by default, each CI run therefore appears in product metrics as a brand-new first-time user whose activity is the suite's fixed set of successful and failed executions — the same counts every run, drowning out real first-run signal.

The test config is mounted by both the local and postgres compose stacks (and the log-demo one), so disabling telemetry here covers every functional CI path with one setting.

Co-authored-by: Claude Fable 5 [email protected] (0bd4a4a)

Fixed

• Fix(control-plane): stop counting agent restarts as failed executions (#941)

• fix(control-plane): stop counting agent restarts as failed executions

A long-running agent node restarts constantly during development — the user saves a file, hits Ctrl-C, or the process dies on a syntax error. The control plane kept believing the node was healthy for as long as it took the health checker to notice (~30s) or the heartbeat to go stale (60s). Every call that landed inside that window created an execution record, failed to dial the dead process, and was recorded as a failed execution. The user was charged a failure for a five-second restart, and it is the largest single source of failed executions a new user produces.

Three changes, in the order a call meets them:

1. Fail fast on a node we already know is down. prepareExecutionForTarget checked pending-approval but never health, so a dispatch into a dead node persisted a row and then failed it. It now returns 503 node_unavailable BEFORE the execution record exists — a request we never dispatched is a rejected request, not a failed execution. This mirrors the check reasoners.go has always had on the legacy proxy route. Only definitively down states are rejected: "unknown" (no heartbeat yet) still goes through, or the first call of every session would fail.

2. Absorb the restart when we do NOT know the node is down. A dial failure is the one transport error that is unambiguously safe to retry — no bytes reached the agent, so nothing can run twice — so the dispatch now waits for the node to come back and replays against its current address, which may be a different port. Recovery is read from the node record (a fresh instance_id, or a heartbeat that advanced), not guessed by blind retrying. Bounded by agent_restart_grace, default 15s, well inside the 90s agent call timeout; serverless targets are excluded because they have no resident process to come back.

3. Close the detection gap. When the wait expires we demote the node to inactive, conditional on the heartbeat we last observed, so the next caller takes path 1 and fails fast instead of repeating the wait.

Also exempts held executions from the orphan reaper. The re-registration that ENDS a restart is exactly what triggers MarkAgentExecutionsOrphaned, so the reaper was failing the very execution the retry was about to complete, and the sync caller received that failure even though the work went on to succeed. Executions marked awaiting_agent_restart are now skipped, on both the executions table and the workflow_executions row the DAG UI reads.

Finally, "agent 'x' not found" now classifies as target_not_found rather than internal_error. That is the quickstart curl run before python main.py: a normal mistake that was being reported as a broken control plane. The category already existed in canonicalFailureCategory and nothing ever assigned it.

Verified against a real Python SDK agent: a call fired while the agent was down, with the process returning 4s later, now returns HTTP 200 succeeded and emits execution_completed. Before this change the same sequence returned 502 with an agent_restart_orphaned message. The documented quickstart runs clean end to end.

Co-Authored-By: Claude Opus 5 [email protected]

• fix(control-plane): harden the restart-absorb path against review findings

Four behavioral fixes to the agent-restart grace mechanism, each found by adversarial review of the original commit:

1. Serverless nodes are exempt from the fail-fast health gate. They have no heartbeat loop and the health monitor never polls them, so the presence sweep marks every serverless node inactive shortly after registration — the gate would have rejected every serverless invocation with 503 node_unavailable, permanently. The gate now runs after serverless normalization and skips serverless targets entirely.

2. Replay requests bypass the gate. A replay hit is served from the recorded run without contacting the agent, so the node being down must not reject it. A replay miss dials and fails exactly as before the gate existed.

3. A cancel that lands during the restart wait now wins: the loop re-reads the execution record before replaying (mirroring callAgent's pre-dispatch check) and aborts without handing the agent work the caller disowned. A pause similarly waits for its resume.

4. The wait aborts as soon as the node record reports the node definitively down (demoted by the health checker or by another dispatch whose grace expired first). Queued dispatches aimed at one dead node no longer each burn their full grace serially on the async worker pool.

Also: the awaiting_agent_restart hold is now released on every exit path (including caller-context cancellation, via a detached write), the comment overstating what completeExecution can repair after an orphan-reaper race is corrected (workflow_executions has no exit from failed), and the test fake's UpdateAgentHealthAtomic now enforces the same conditional-heartbeat semantics as LocalStorage instead of discarding the argument.

Co-Authored-By: Claude Fable 5 [email protected]

Co-authored-by: Claude Opus 5 [email protected] Co-authored-by: Abir Abbas [email protected] (e486208)

• Fix(assets): correct garbled text in README features-strip banner (#940)

The AI-generated banner image read '90+ Production Features built into the controhe' — a garbled 'control plane'. Re-typeset the line as 'built into the control plane' in Inter (the repo's UI typeface), matching the original colors, cap height, and starfield background.

Co-authored-by: Claude Fable 5 [email protected] (0f91495)

Other

• Bundle sec-af and cloudsecurity-af with the desktop app (#939)

Promote the two security nodes from marketplace catalog rows to BUNDLED_NODES, so a fresh install provisions all four agent nodes (swe-planner, pr-af, sec-af, cloudsecurity-af) on first launch. Both repos follow the same superseded_by //go redirect as SWE-AF and pr-af, and their maintained nodes are Go, so the stale python language chips go away with the move. CATALOG stays as the (now empty) seam for future marketplace-only rows.

Co-authored-by: Claude Fable 5 [email protected] (30f821e)

Other

• Ship the desktop app with bundled agent nodes, missing-key surfacing, and the offload-by-default skill (#937)

• docs(desktop): document bundled agent nodes and first-launch provisioning

Explain what ships with the app, why delivery is fetch-on-first-launch rather than a payload in the DMG, why the bundled nodes are named at their bare repo URL (the superseded_by redirect is what migrates an existing Python node), and the three rules that make provisioning safe to run on every launch: uninstall sticks, failure retries, nothing is auto-started into a guaranteed missing-keys failure.

Co-Authored-By: Claude Opus 5 [email protected]

• feat(desktop): ship swe-planner and pr-af as bundled nodes

The two flagship nodes stop being marketplace cards and become nodes that ship with the app: shared/bundled.ts lists them, and main/bundledAgents.ts provisions them on first launch through the existing control-plane install API (fetch on first launch, so no packaging changes).

• shared/bundled.ts: BUNDLED_NODES / isBundled / bundledEntry, sourced at the bare repo URL so each root manifest's superseded_by redirect still migrates a user who has the older Python node.
• shared/catalog.ts: CATALOG keeps only sec-af and cloudsecurity-af; catalogEntry() now resolves over CATALOG + BUNDLED_NODES, which keeps a bundled node installable and --force updatable from the Agents view without widening the "renderer only passes a vetted name" boundary.
• shared/types.ts: BundledPhase / BundledStatus, snapshot.bundled, and settings.provisionedBundled.
• main/settings.ts: provisionedBundled defaults to [] and is coerced like autostartAgents, so uninstalling a bundled node sticks across launches.
• main/bundledAgents.ts: pure planBundledInstalls() plus a sequential, never-throwing ensureBundledAgents() driven by injected deps; failures stay visible as rows and are not recorded, so the next launch retries.

Co-Authored-By: Claude Opus 5 [email protected]

• feat(desktop): provision bundled agent nodes on first launch

Wire the main process for the two nodes that ship with the app (swe-planner, pr-af): they are fetched through the existing install API after autostart has a control plane, instead of sitting in the marketplace grid.

• getSnapshot() carries bundled, threaded through SnapshotOptions the same way the skillSync extra is, so the renderer's existing 5s poll delivers the provisioning rows with no new IPC channel.
• Provisioning is chained onto the userPathReady/runAutostart promise rather than started beside it — installing needs the control plane autostart adopted or brought up.
• installInFlight becomes a real mutex with a waiter queue: IPC handlers still refuse a concurrent install outright (a click deserves an answer), while provisioning parks and takes its turn.
• markProvisioned records the node in settings so an uninstall sticks across launches; onInstalled adds it to autostartAgents rather than starting it now, because both nodes need API keys a first-launch user has not entered yet.

Co-Authored-By: Claude Opus 5 [email protected]

• feat(desktop): show bundled agent provisioning in the Agents view

First launch fetches swe-planner and pr-af through the control-plane install API, so the Agents library must show them arriving instead of an empty panel or the marketplace.

• AgentsPanel takes bundled: BundledStatus[] and renders those nodes above the installed rows: name, provisioning badge, description, and the streamed message as the row progress line (.error-text when the install failed). The rows are inert — nothing to start or configure until the install lands — and fade in/out per DESIGN.md §5.2.
• The "No agents installed" empty state stays away while a bundled node is still provisioning; that library is unfinished, not empty.
• App passes snapshot.bundled through, keeps libraryEmpty false while any bundled row exists, and routes a cold launch to the Agents library when nodes are arriving (add-mode only when nothing is installed and nothing is coming).
• .badge.provisioning uses the accent tint from the shared tokens: mid-install is calm progress, not a fault.

Co-Authored-By: Claude Opus 5 [email protected]

• fix(desktop): drop the duplicate bundled log prefix

bundledAgents.ts already prefixes each line; index.ts prefixing again produced "bundled: bundled: …".

Co-Authored-By: Claude Opus 5 [email protected]

• docs(skill): make a missing agent key a blocking handoff in agentfield-use

An installed-but-unstarted node is the default first-run state: the desktop app ships swe-planner and pr-af pre-provisioned but deliberately not started, because their API keys are not entered yet. A coding agent driven by this skill then saw only HTTP 400 agent 'X' not found (or the MCP target ... not found text), never learned the real cause, and retried or silently substituted a different approach.

Bump agentfield-use to 0.7.0 and make the recovery path explicit:

• Start before dispatch. If a node is in af list but absent from discovery, or its health_status is not active, run af run <name> first. That start attempt is the diagnostic: it reads the encrypted store that actually gates startup and names the exact missing variable.
• Missing key is a blocking handoff. Name the variable(s) and the verbatim af secrets set ... --node <name> command, point at Desktop -> Agents -> <node> -> Keys, then wait. Never retry, never substitute another agent, never do the work silently instead, never ask for the value in chat. Consistent with the agentfield-personal handoff convention.
• Warn off the store-blind commands. af doctor reads os.Getenv and af config <pkg> --list reads the package .env, so both report a correctly-stored key as unset and neither renders require_one_of groups.
• Extend the failure table with the actually-observed 400 body (error_category: internal_error, not 404) and the MCP not-found text, both routed to the same recovery.

Adds a contract test pinning the new text, bumps the catalog version pins, and syncs the embedded skill_data mirror.

Co-Authored-By: Claude Opus 5 [email protected]

• feat(desktop): warn on the top bar when installed agents need keys

The bundled nodes are installed but never started, because both want API keys a first-launch user has not entered. Until now the only signal was the per-row "Needs keys" chip — invisible to someone who never opens the app and then asks a coding agent to review a PR.

KeysBanner rides the existing banner slot and reads getEnvReports, the only store-aware source of truth (af doctor and .env are store-blind). The unreachable-control-plane sentinel and the metadata-blind satisfied: true fallback both stay silent, so the banner cannot cry wolf. It is not dismissible: unlike the update offer, this reports that the product cannot work, and it clears itself the moment keys resolve.

Refresh is event-driven, not a second timer — reload on mount, on agent roster changes, and on view changes.

Co-Authored-By: Claude Opus 5 [email protected]

• fix(desktop): keep the star prompt from stacking under the keys banner

StarBanner already yields to an undismissed app update; it knew nothing about the keys banner, so a first launch could show both — asking for a GitHub star directly under a strip saying the installed agents cannot run.

KeysBanner reports its visibility upward and App carries the signal across, rather than StarBanner re-deriving it: getEnvReports fans out to the control plane per package, and one caller is enough.

Co-Authored-By: Claude Opus 5 [email protected]

• feat(desktop): notify once when bundled agents land without their keys

First-launch provisioning installs swe-planner and pr-af and deliberately does not start them, because both need an API key the user has not entered. On a login-item launch the app is hidden in the tray, so the Agents row's "Needs keys" chip is telling an empty room — the user finds out when a coding agent fails to call them.

main/keyNotice.ts closes that loop with one native notification, in the planner/runner shape of aforge-companion.ts and bundledAgents.ts: planKeyNotice() is pure and unit-tested (which agents, notify or not, given the env reports and what was already announced), and the Electron Notification is a thin injected effect in index.ts. Clicking it opens the app on the Agents view.

The verdict comes only from getEnvReports().satisfied — the control plane's per-agent secrets endpoint, i.e. the encrypted store af run actually reads — so a correctly stored key never triggers a notice, and the deliberate satisfied: true fallback for control planes too old to report requirement metadata is never notified on.

At-most-once is a new settings.keyNoticeShown list, the provisionedBundled precedent: announced names are persisted and filtered out, so the notice belongs to the provisioning event and no launch can raise it twice. Notification.isSupported() is checked, nothing throws, and nothing here can delay or break startup.

Co-Authored-By: Claude Opus 5 [email protected]

• docs(desktop): document the three missing-key surfaces

Provisioning installs the bundled nodes without starting them, so the README's "Nothing is started" needed the other half: how the user finds out a key is missing. Records why there are three surfaces (hidden-at-login, in-window, and coding-agent) and which single source of truth they all read — plus the warning that af doctor and af config --list are store-blind.

Co-Authored-By: Claude Opus 5 [email protected]

• feat(skills): rewrite agentfield-use around offloading (0.8.0)

Installed AgentField agents are subharnesses: workers a coding harness offloads to instead of doing the work inline. They run on smaller open models, in parallel, off the harness's own context, and every run is recorded on the control plane.

The doctrine now leads the skill:

• Offload by default. Coverage decides, not task size, and the fleet is discovered at runtime rather than enumerated here — users install new subharnesses at any time. One af ls -e answers "does anything cover this?", cached per session.
• Announce every offload with the run's live UI link (<server>/ui/runs/<run_id>), which is for the user to watch in parallel, never a substitute for the harness's own monitoring.
• The user can always override; a failed or stalled run is reported and asked about, never silently redone inline and presented as the subharness's work.
• Vocabulary rule: agent/reasoner/node in commands and fields, "subharness" in what the harness says to its user.

The golden path is now CLI-first and async: af call --schema → af call --in … --async (client-side schema validation before dispatch) → a deliberate retrieval mode (wait / tail / group poll / webhook for services only, rendered as a decision table) → report. Adds the previously undocumented webhook body, scoped explicitly away from coding harnesses, and honest cost reporting: duration_ms is per-run truth, /api/ui/v1/usage/stats is a window aggregate and there is no per-run cost endpoint to quote.

Every operational fact from 0.7.0 is preserved: server resolution, MCP, discovery gotchas, start-before-dispatch and the missing-key blocking handoff, entry-points-only, contract-before-call, no-coverage offer, concurrency/load/canary, workspace_handle/furrow, wedge protocol, sessions, audit trail, failure table (plus a row for af wait exit 2 = timeout, not failure) and hard rules.

Frontmatter, catalog and mirror pins bumped to 0.8.0; adds contract tests for the offload doctrine and the async golden path.

• fix(desktop): keep bundled provisioning local-only and make uninstall stick

First-launch provisioning ran against whatever control plane was active, so a user with a cloud control plane configured would get swe-planner and pr-af installed onto the remote server on the first launch after upgrading, and the app-global provisionedBundled latch would then suppress local provisioning for good. planBundledInstalls now takes cloudActive and skips while a cloud connection is active; it also treats an unreadable registry as "not ready" instead of "nothing installed", and the runner asks the control plane whether it has the install API before seeding rows, so an older control plane is skipped once per launch instead of producing two red rows every time.

"Uninstall sticks" only held for nodes the app itself installed: a node the user already had was skipped by the plan but never recorded, so removing it later brought it back on the next launch. Installed-but-unrecorded bundled nodes are now adopted into provisionedBundled before any install runs.

Co-Authored-By: Claude Fable 5 [email protected]

• fix(desktop): land a first launch on the arriving bundled rows, not the marketplace

The cold-launch default route is applied on the first snapshot, which the renderer fetches the moment it mounts — long before main has finished autostart and seeded any provisioning rows. defaultView therefore always saw bundled=0 / agents=0 on a fresh install, chose add-mode, and the user sat in a marketplace (now without swe-planner and pr-af) while the two nodes installed out of sight. Re-evaluate once: if the auto-applied view is still add-mode, no deep link or user navigation has happened, and bundled rows appear, switch to the Agents library. shouldRerouteToBundled is the pure decision with a truth-table test.

Co-Authored-By: Claude Fable 5 [email protected]

• fix(desktop): refresh env reports when the registry gains a row; hide redundant bundled rows

AgentsPanel loaded env reports on mount only. Provisioning is the first flow where registry rows appear while the panel stays mounted, so a freshly-landed bundled node rendered without its "Needs keys" chip or Keys button (both gated on a report) and without the pre-start key gate. The load effect is now keyed on a stable roster key (the set of agent names), so it re-runs exactly when a row is added or removed.

A bundled node's 'installed' provisioning row was kept until the whole run finished, so it was rendered twice — provisioning row plus registry row — for the duration of the next node's install. visibleBundledRows drops any provisioning row whose name already has a registry row.

Co-Authored-By: Claude Fable 5 [email protected]

• fix(desktop): keep bundled nodes reachable from the marketplace

Removing swe-planner and pr-af from the curated catalog meant a user who uninstalled one had no curated way to get it back (the provisioned latch correctly never re-installs it). The catalog IPC now lists the bundled nodes ahead of the marketplace rows; installed ones render the usual "Installed" state, and a node the app is currently provisioning renders a disabled "Installing…" so the marketplace never asks the user to install what the app is already installing.

Co-Authored-By: Claude Fable 5 [email protected]

• fix(control-plane,desktop): report keys supplied by the control plane's environment as set

GET /api/ui/v1/agents/:id/secrets computed is_set from the encrypted store alone, while the runner's EnvResolver resolves the control plane's own process environment first. A key exported in the shell (or a Windows user env var) therefore started the agent fine but was reported missing — which the desktop turned into a non-dismissible "needs API keys" banner, a "Needs keys" chip, a blocked Start button and a one-shot OS notification, all false. The handler now checks the process environment too (presence only, never values), reports it as env: true, and is_set reflects runtime truth; scope still names only where a STORED value lives. The desktop maps env-only keys to the existing 'env' status ("From environment") and keeps storedScopes store-backed.

Co-Authored-By: Claude Fable 5 [email protected]

• fix(desktop): tell a spawned control plane its own URL for the agents it starts

The control plane hands every agent it starts the URL from its own AGENTFIELD_SERVER environment, falling back to http://localhost:8080 regardless of the port it actually listens on. The app only pinned AGENTFIELD_PORT, so a control plane on any other port — the auto-picked one when 8080 is busy, or a configured one — told swe-planner, pr-af and the swe-pro engine to register with localhost:8080: with whatever else was there (observed live: the nodes joined an unrelated control plane on 8080), or nothing. Spawn the server with AGENTFIELD_SERVER=http://localhost:<port> too.

Co-Authored-By: Claude Fable 5 [email protected]

• fix(desktop): seed the local control-plane port before the first poll

Until autostart resolved the port, every early snapshot went to localhost:8080 — not the configured port, nor the one the last launch ended on. With something else on 8080 the renderer's first snapshot described a foreign server (a healthy control plane with an empty registry), and the cold-launch route was decided on it. applyConnectionProfile now seeds the local port from settings.controlPlanePort ?? lastControlPlanePort, so the first poll of a launch already targets this app's control plane.

Co-Authored-By: Claude Fable 5 [email protected]

• fix(desktop): decide the cold-launch route only once the registry is readable

The registry is read through the control plane, so the first poll after a cold autostart sees "no registry" while the server is coming up. Routing on that snapshot sent a user with a stocked library into the marketplace on every cold launch. canDecideDefaultRoute waits for a readable registry (or provisioning rows, which only exist once the control plane answered); until then the initial Home view with its control-plane status is what shows.

Co-Authored-By: Claude Fable 5 [email protected]

• test(control-plane): cover the agent-secrets handler's default process-env lookup

Co-Authored-By: Claude Fable 5 [email protected]

• test(control-plane): make the runner-resolution secrets test hermetic

EnvResolver prefers a non-empty process env value, so a developer machine with OPENAI_API_KEY exported resolved the host value instead of the stored sk-test and the assertion failed outside CI. Neutralize the variable with t.Setenv — empty counts as unset — so the test is deterministic everywhere.

Co-Authored-By: Claude Fable 5 [email protected]

Co-authored-by: Claude Opus 5 [email protected] Co-authored-by: santoshkumarradha [email protected] Co-authored-by: Abir Abbas [email protected] (2bbf5ff)

Fixed

• Fix(install): disable VCS stamping when building Go agent nodes (#938)

• fix(install): disable VCS stamping when building Go agent nodes

The installed package copy under ~/.agentfield/packages is not a git checkout, so there is nothing truthful for go build's -buildvcs to stamp. Worse, the go tool's VCS auto-detection walks up PAST the package root: a .git anywhere above the packages dir — a dotfiles-managed $HOME, a stray /tmp/.git — either hard-fails the install with "error obtaining VCS status: exit status 128" or stamps an unrelated repository's revision into the node binary.

Found live: a stock 'af install https://github.com/Agent-Field/sec-af' (and cloudsecurity-af) failed at the go-build step on a machine with an empty /tmp/.git directory. The regression test reproduces exactly that layout with the real toolchain and fails without the flag.

Co-Authored-By: Claude Fable 5 [email protected]

• test(templates): make the scaffold-build test immune to parent VCS state

TestRenderedGoScaffoldBuilds runs a plain 'go build' inside t.TempDir(), so the same walk-up VCS detection fixed in InstallGoDependencies could fail this test for reasons unrelated to the template (reproduced on a machine with a stray /tmp/.git). Build the rendered scaffold with -buildvcs=false, mirroring what the installer now does with that same scaffold.

Co-Authored-By: Claude Fable 5 [email protected]

Co-authored-by: Claude Fable 5 [email protected] (eb522f7)

Fixed

• Fix(sdk/typescript): honour AGENTFIELD_SERVER for the default control-plane URL (#934)

The TS SDK hardcoded http://localhost:8080 in three places and never read the environment, so nodes launched by af run / docker compose with AGENTFIELD_SERVER set registered against localhost. Resolve like Python: explicit agentFieldUrl > AGENTFIELD_SERVER > AGENTFIELD_SERVER_URL > localhost:8080, from one helper used by all three sites.

Co-authored-by: Claude Fable 5 [email protected] Co-authored-by: Santosh kumar [email protected] (513dd64)

Fixed

• Fix(sdk): send X-OpenRouter-Categories on OpenRouter requests (#930)

All three SDKs now attach X-OpenRouter-Categories=cli-agent,programming-app on OpenRouter traffic, matching the control-plane defaults, with the same AGENTFIELD_OPENROUTER_CATEGORIES / OR_CATEGORIES env override chain used for site URL and app name.

Co-authored-by: Cursor Agent [email protected] Co-authored-by: Santosh kumar [email protected] (5718a25)

Fixed

• Fix(sdk/typescript): convert zod 4 schemas with the native toJSONSchema (#935)

• fix(sdk/typescript): convert zod 4 schemas with the native toJSONSchema

harness()/ai() schema conversion only knew zod-to-json-schema, which does not understand zod 4 instances (a fresh npm i zod is 4.x), so zod 4 schemas were serialized as zod internals. Detect zod 4 schemas (_zod) and convert with zod's own toJSONSchema (zod, then zod/v4); zod 3 keeps using zod-to-json-schema unchanged; plain JSON Schema still passes through. No dependency or lockfile changes.

Co-Authored-By: Claude Fable 5 [email protected]

• fix(sdk/typescript): resolve zod 4's toJSONSchema from the application's zod copy first

The schema instance is built by the app's zod; converting it with a different copy (e.g. the SDK's nested zod 3.25 via zod/v4) keeps the structure but drops registry-backed metadata — .describe() descriptions and the .int() refinement (integer → number). Try the zod resolvable from the working directory before the SDK-relative copies. Verified end to end with zod 4.4.3: the provider now receives description and integer.

Co-Authored-By: Claude Fable 5 [email protected]

Co-authored-by: Claude Fable 5 [email protected] (a9d634d)

• Fix(sdk/typescript): opencode surfaces exit-0 stderr failures and signals as errors (#933)

TS only treated (exit != 0 && no output) as an error. Python and Go also flag: a negative exit code (killed by signal), and exit 0 with empty output whose stderr matches opencode's known failure patterns (^Error:, Model not found, AuthenticationError, Unauthorized, APIError) — surfacing the matching stderr window instead of an empty non-error result. Same pattern list as the other two SDKs; ANSI stripped before matching; failureType set.

Co-authored-by: Claude Fable 5 [email protected] (83e2921)

• Fix(sdk/go,sdk/typescript): gemini argv — drop the nonexistent -C flag, map permission modes like Python (#932)

The Gemini CLI has no -C flag; both providers passed -C <cwd> (the cwd is already applied as the subprocess working directory) and mapped permission_mode=auto to --sandbox, which restricts execution rather than granting it. Match the Python provider (agentfield#687): no -C, auto → --yolo, plan → --approval-mode plan, then -m <model> and -p <prompt>. Tests that pinned the old argv are rewritten table-driven from the CLI's real flags.

Co-authored-by: Claude Fable 5 [email protected] (ff6e28c)

• Fix(sdk/typescript): pass claude-code options to claude-agent-sdk in camelCase (#931)

The provider built the query() options with snake_case keys (max_turns, allowed_tools, system_prompt, max_budget_usd, permission_mode). @anthropic-ai/claude-agent-sdk's Options type is camelCase (maxTurns, allowedTools, systemPrompt, maxBudgetUsd, permissionMode), so unknown keys were dropped and turn caps, USD budget, tool allowlist, system prompt and permission mode were silently ignored on TypeScript. The existing test asserted the snake_case keys; it now asserts the contract at the SDK boundary.

Co-authored-by: Claude Fable 5 [email protected] (23a6965)

Fixed

• Fix(sdk/go): attribute default-provider harness runs to aforge, prefer result.Model (#929)

Since v0.1.130 the harness provider is optional (env AGENTFIELD_HARNESS_PROVIDER, else aforge). recordHarnessUsage still derived the harness column only from opts.Provider / HarnessConfig.Provider, so a default-provider run was recorded as harness="" and model="harness". Resolve the effective provider through harness.ResolveProviderName (same precedence the runner uses) and let the provider-reported result.Model win over the configured model, mirroring the Python fix in #928.

Co-authored-by: Claude Fable 5 [email protected] (937641a)

Documentation

• Docs: feature Person to Brief AF (#912) (bc290a4)

Fixed

• Fix(harness): attribute default-provider runs to aforge, set HarnessResult.model, pass max_turns to aforge (#928)

• fix(sdk/python): attribute default-provider harness runs to aforge

_record_harness_usage resolved the provider from self.harness_config, which is None unless the caller built a HarnessConfig explicitly. On the default path that left the cost entry with harness=None and model="harness" even though aforge is what actually ran — the entries were unattributable.

Resolve through agentfield.harness._defaults.resolve_harness_provider instead, which is the same precedence the runner uses to pick a provider: explicit value > AGENTFIELD_HARNESS_PROVIDER > "aforge". Explicit and configured providers keep winning; only the previously-empty case changes.

Co-Authored-By: Claude Fable 5 [email protected]

• fix(harness): report the model aforge actually ran on every result

The aforge providers only reported a model when the caller pinned one, so the default path — now the common one — produced results with no model at all. Go had no model field on Metrics/Result to report through in the first place.

Report the effective model: the resolved caller model when given, else AFORGE_MODEL as it stands in the final env overlay (caller-supplied env wins), else aforge's own built-in default, pinned as a named constant per SDK. Go gains Metrics.Model and Result.Model, mirroring the Python SDK, with the same first-non-empty-wins aggregation across retries.

Co-Authored-By: Claude Fable 5 [email protected]

• fix(harness): pass max_turns through to aforge as --turns

The aforge providers dropped the caller's turn cap, so a run asking for two turns got aforge's 200-turn runaway backstop instead. With aforge as the default provider that is now the ordinary path, not a corner case.

Map a positive turn cap onto --turns N, right after --timeout, on the exec argv only — aforge do has no such flag and would fail to parse it. Cost caps stay unmapped on purpose: aforge's --budget counts tokens, not dollars, so there is no honest conversion from max_budget_usd.

Co-Authored-By: Claude Fable 5 [email protected]

Co-authored-by: Claude Fable 5 [email protected] (ca88757)

• Fix(harness): staging installs put aforge beside af; doctor exits 0 on a survey; doctor knows grok (#927)

• fix(install): install aforge beside af on the staging channel

The staging channel installs af into ~/.agentfield-staging/bin and adds only that directory to PATH, but install_aforge() delegates to af aforge ensure, which installs into $AGENTFIELD_HOME/bin — ~/.agentfield/bin by default. So a staging install downloaded aforge into a directory it never put on PATH, and the SDK's "run af aforge ensure" remedy could not fix it either, because the remedy resolves the same default home.

Pin AGENTFIELD_HOME to the parent of INSTALL_DIR for the ensure call so aforge always lands beside the af that installed it. Production is unchanged: ~/.agentfield/bin strips to ~/.agentfield, which is what ensure already defaulted to. install.ps1 has the same split whenever AGENTFIELD_INSTALL_DIR is set, so Install-Aforge pins and restores the variable the same way.

--no-aforge / AFORGE_MODE=none is untouched: ensure is still never invoked.

Co-Authored-By: Claude Fable 5 [email protected]

• fix(cli): af harness doctor exits 0 when it is a survey

Without --provider the command surveys the machine, but it still returned "requested harness provider is unavailable: <first missing>" for the first provider it could not use. A healthy fresh install that simply has not installed every coding harness therefore exited 1, and the message was printed three times: cobra's own copy, the AgentHintJSON "invalid_command" blob from main.go, and the zerolog error line.

A bare doctor is informational — a machine legitimately has providers it has not installed — so the unusable gate now applies only when the caller named providers explicitly. When it does fire it returns cliExitError, which makes main.go print the plain message instead of the invalid_command blob (the command was invoked correctly), and SilenceErrors/SilenceUsage on the subcommand drops cobra's duplicate. --provider unknown is untouched: that really is a mis-invocation, so the agent hint stays.

Co-Authored-By: Claude Fable 5 [email protected]

• feat(cli): teach af harness doctor about the grok provider

The Python SDK ships a grok entry in PROVIDER_SPECS (sdk/python/agentfield/harness/_availability.py), so an agent can be asked to run it, but the Go doctor's provider table did not list it: af harness doctor silently omitted grok and --provider grok was rejected as an unknown provider. Add the row and list it in the flag's help text, plus a test that pins the two lists to the same provider-name set so they cannot drift again.

Co-Authored-By: Claude Fable 5 [email protected]

Co-authored-by: Claude Fable 5 [email protected] (7f88434)

• Fix(install): make --no-aforge / AFORGE_MODE=none actually skip the aforge download (#926)

The installer only skipped its own explicit af aforge ensure step, but af skill install --all (which runs first) triggers the same best-effort aforge provisioning hook (skillkit/install.go), so a user who opted out still got the 35 MB binary downloaded into ~/.agentfield/bin. Regressed with #924, which added both the hook and the flag.

Export AGENTFIELD_SKIP_AFORGE=1 for the whole run when AFORGE_MODE=none — that env gate is the single switch the Go side honours — so every af (and af-tray) invocation the script makes respects the opt-out.

Verified against v0.1.130-rc.5 into an isolated HOME: patched installer with --no-aforge installs af + skills + furrow and no aforge; the unpatched script installs aforge despite the flag.

Co-authored-by: Claude Fable 5 [email protected] (4930075)

• Fix(sdk/python): exclude litellm 1.97.0, which crashes on Python 3.10 (#925)

litellm 1.97.0 (PyPI 2026-08-16) fails while constructing ModelResponse on Python 3.10 with pydantic 2.13:

PydanticUserError: `Message` is not fully defined; you should define
all referenced types, then call `Message.model_rebuild()`

(BerriAI/litellm#36384). The SDK declares a bare litellm dependency, so a fresh pip install agentfield on 3.10 — a version we support and test — resolves to 1.97.0 and every app.ai() call dies inside litellm.completion() before reaching any provider. Verified standalone: 1.97.0 breaks on 3.10, works on 3.11/3.12; 1.96.2 works on 3.10.

Exclude just that release so pip resolves 1.96.2 (or a fixed 1.97.x+) and the docs quickstart works again on a clean install.

Co-authored-by: Claude Fable 5 [email protected] (b149cfd)

Added

• Feat(harness): make aforge the default provider across Python, Go, and TypeScript (#905)

• feat(harness): add aforge do parity across SDKs

• feat(harness): preserve TypeScript failure metadata

• fix(harness): parse live aforge JSON envelopes

• feat(harness): support aforge exec across SDKs

• feat(harness): default aforge to exec

• feat(harness): make aforge the default provider in the Python SDK

app.harness("...") with nothing configured now runs AForge, AgentField's native coding harness, instead of raising "No harness provider specified".

Provider precedence is explicit value > AGENTFIELD_HARNESS_PROVIDER > "aforge", implemented once in harness/_defaults.py and applied both where HarnessConfig materialises its default and where the runner resolves options (so a runner built without any config follows the same chain).

HarnessConfig.model stops defaulting to "sonnet" — that was Claude-specific and wrong for every other provider. The default is now empty, meaning "use the provider's own default", and the claude-code provider carries "sonnet" internally so explicit claude-code users see no change.

Also drops the private-repo build instruction from the aforge install hint; the binary ships alongside af and af aforge ensure (re)installs it.

Co-Authored-By: Claude Fable 5 [email protected]

• feat(harness): make aforge the default provider in the Go SDK

Runner.Run no longer errors on an empty Options.Provider. BuildProvider and Run both route the name through ResolveProviderName, which applies the same precedence as the Python SDK: explicit value > AGENTFIELD_HARNESS_PROVIDER > DefaultProvider ("aforge"). The resolved name is written back onto the options so error messages and provider construction see the real provider.

Model stays empty by default and means "the provider's own default"; the Harness doc example no longer suggests a Claude-specific "sonnet".

Co-Authored-By: Claude Fable 5 [email protected]

• docs(harness): lead with AForge as the default harness

Rewrites the harness-providers lead so the zero-setup path is the headline: AForge is the default, af aforge ensure installs it alongside af, and picking Claude Code / Codex / Gemini CLI / OpenCode is an override of one option rather than a prerequisite. Documents the provider precedence chain (explicit > AGENTFIELD_HARNESS_PROVIDER > aforge) and that an unset model means the provider's own default.

Drops the "go build -o aforge ./cmd/aforge" instruction — that repo is not public and is no longer how anyone gets the binary. Adds the grok row (Python SDK only) so the install table matches the supported provider set.

Co-Authored-By: Claude Fable 5 [email protected]

• docs(harness): correct the v2 design doc's "provider is required" rule

The design doc still asserted that HarnessConfig.provider has no implicit default and that a call without one raises. Both are false now: provider resolves through explicit > AGENTFIELD_HARNESS_PROVIDER > "aforge", and model defaults to the provider's own rather than "sonnet".

Co-Authored-By: Claude Fable 5 [email protected]

• feat(harness): make aforge the default provider in the TypeScript SDK

HarnessRunner.run no longer throws when nothing sets a provider. Both the runner and buildProvider route the name through resolveProviderName, which applies the same precedence as the Python and Go SDKs: explicit value > AGENTFIELD_HARNESS_PROVIDER > DEFAULT_HARNESS_PROVIDER ("aforge"). The resolved name is written back onto the options so providers and error messages see the real provider.

HarnessConfig.provider becomes optional to match, so {} is a complete config.

Co-Authored-By: Claude Fable 5 [email protected]

• fix(harness): stop pointing users at a private repo to get aforge

The Go and TypeScript adapters told anyone hitting a missing binary to "Build it from https://github.com/Agent-Field/aforge-v2" — a repo they cannot open. AForge ships with af, so the message now names af aforge ensure and the AFORGE_BIN escape hatch, matching the Python install hint.

Co-Authored-By: Claude Fable 5 [email protected]

Co-authored-by: Claude Fable 5 [email protected] (6b8c9be)

Added

• Feat(install): provision the aforge harness binary alongside af (curl, desktop, docker) (#924)

• feat(aforge): pinned-binary provisioner for the aforge coding harness

AgentField's harness providers spawn aforge exec --json, but nothing in the product ever put that binary on a machine — every install surface assumed the user had built it from a private repo. This adds the provisioning half.

Shaped on internal/furrow's provisioner (same lock, marker, atomic-rename and best-effort contract) with three deliberate differences:

• assets are distributed gzipped, so the stream is decompressed before it is hashed — checksums.txt carries the sha256 of the UNCOMPRESSED binary — and the decompressed size is capped so a bad endpoint cannot gzip-bomb us;
• all six platforms are mapped, not the three furrow happens to ship;
• Options.Force exists so af aforge ensure --force can bypass the marker.

AGENTFIELD_AFORGE_BASE_URL overrides the whole base (mirrors, staging hosts); AGENTFIELD_SKIP_AFORGE=1 makes the whole thing a no-op.

Co-Authored-By: Claude Fable 5 [email protected]

• feat(cli): af aforge ensure, and teach the harness doctor about aforge

af aforge ensure mirrors af furrow ensure — Ensure, not EnsureBestEffort, because someone who asks for the binary by name is owed the failure — plus a --force that bypasses the version marker.

The doctor needed two new ideas to describe aforge honestly:

• VersionArgs, because providers do not agree on how to be asked. aforge will answer version; every other provider answers --version.
• VersionOptional, because the pinned aforge build answers BOTH with its whole usage banner on exit 1. A present, executable binary is enough to call it usable; it reports version "unknown" with an informational "version_unavailable" issue until the next aforge release adds the subcommand. The probe requires exit 0 before it believes any output — without that it would file the usage banner as the installed version.

The probe also falls back to $AGENTFIELD_HOME/bin when PATH misses: the shell that just ran af aforge ensure has not re-read PATH, and reporting a binary we installed thirty seconds ago as missing is the wrong answer.

Co-Authored-By: Claude Fable 5 [email protected]

• feat(install): provision aforge from the skill install and both shell installers

Three entry points, one implementation. af skill install --all, the curl installer and the PowerShell installer all end up calling the same Go provisioner, so there is exactly one place that knows the pinned version, the download host and the checksum rules.

The skillkit hook goes in InstallAll rather than install(): InstallAll is the one call every fresh machine makes, and hooking each skill would re-download a 35MB binary once per catalog entry. --dry-run stays side-effect free.

Both shell installers keep aforge strictly optional — a failed provision warns and moves on, because by that point the control plane is already installed and working. Opt out with --no-aforge / AFORGE_MODE=none (sh) or -NoAforge / $env:AFORGE_MODE='none' (ps1; a piped iwr | iex cannot pass a switch, hence the env var).

Co-Authored-By: Claude Fable 5 [email protected]

• feat(desktop): ensure aforge on launch, once, without blocking startup

A desktop-only install never runs the curl installer, so it would have had af and no harness. The app now shells out to af aforge ensure right after it resolves the CLI — no bundled payload, so there is one download path and the upgrade rules stay in Go.

Deliberately not extraResources: bundling the binary would fork the install path and freeze the pinned version at package time, and the app would still need the runtime check for machines that already had an older copy.

Shaped like tray-companion.ts: planAforge() is pure so the skip rules are unit tested, and the effect takes injected deps so no test ever spawns anything. Fire-and-forget with a once-per-launch latch set before the await, so a dead network delays nothing and two callers still produce one download.

Co-Authored-By: Claude Fable 5 [email protected]

• feat(docker): ship aforge in the images that actually run agent nodes

The cloud control plane (single-container topology), the python-agent image and the go-agent image all host agent nodes, and a node that reaches for a harness in a container without one fails at spawn. Dockerfile.control-plane is left alone: it is distroless, has no shell, and by construction runs agents elsewhere.

Soft fetch, hard verify. The asset host goes live with the website deploy, so a 404 has to degrade to "this image ships without aforge" rather than break every image build in the meantime — but a download that does land is always checked against the published sha256, taken after gunzip because checksums.txt hashes the uncompressed binary. A mismatch fails the build.

The COPY takes the fetch stage's output DIRECTORY, not a fixed file path: when the fetch was skipped the directory is empty and the COPY is a no-op, instead of planting a zero-byte aforge on PATH that af harness doctor would cheerfully report as installed.

aforge is a statically linked ELF, so the debian-built asset runs unchanged on the musl/alpine go-agent base.

Co-Authored-By: Claude Fable 5 [email protected]

• docs(aforge): point every install hint at af aforge ensure

The Python SDK told users to go build -o aforge ./cmd/aforge from a private repo when a harness call failed — advice nobody could follow. It now names the command that actually exists on their machine.

Also documents the two knobs the provisioner reads (AGENTFIELD_AFORGE_BASE_URL for mirrors, AGENTFIELD_SKIP_AFORGE for air-gapped hosts) and the shell installers' equivalents.

Co-Authored-By: Claude Fable 5 [email protected]

• test(aforge): cover the provisioner's failure paths

CI's patch-coverage gate came in at 79.00% against an 80% floor — all of the misses were the error branches of the new code, which is exactly the code you want covered: a provisioner is mostly failure handling.

Adds tests for the paths that were only reachable by breaking something: an unwritable bin directory (MkdirAll and flock failures), a bin/aforge path occupied by a directory so the atomic rename fails, a marker path likewise, a corrupt/truncated/non-gzip body, an unreachable host, a checksums file with malformed hex or no line for the asset, and every branch of home resolution. Also exercises the default GOOS/GOARCH path and asserts — via a fake RoundTripper rather than a real request — that the default base URL is the one actually dialled.

internal/aforge: 78.0% → 95.5% of statements. No production code changed.

Co-Authored-By: Claude Fable 5 [email protected]

• chore(aforge): pin the installer to aforge v0.1.0 and require a real version from the doctor

aforge-v2 cut its first semver release, so every surface that provisioned build-9b3ff482de3f now provisions v0.1.0: the pinned constant in control-plane/internal/aforge, the AFORGE_VERSION build arg in the cloud control-plane / python-agent / go-agent images, and the base-URL example in .env.example.

The release also ships a real version subcommand (and --version), which retires the reason aforge was allowed to be "usable" without one. Drop the aforge-only VersionOptional allowance from the harness doctor: a binary that cannot name itself is now version_probe_failed and unusable, exactly like every other provider. VersionArgs still tries "version" before "--version", and a non-zero exit is still rejected so a usage banner never lands in the version field.

The Docker fetch stays soft-fail on a 404 (hard-verify on a hit) — that is a separate documented follow-up, not this change.

Co-Authored-By: Claude Fable 5 [email protected]

Co-authored-by: Claude Fable 5 [email protected] (480342e)

Fixed

• Fix high-severity npm dependency alerts (#922)

Co-authored-by: Abir Abbas [email protected] (efbaca3)

Added

• Feat(examples): triggers-demo-go end-to-end demo + Go skill docs (#516) (#917)

• chore(release): v0.1.118-rc.1 [skip ci]

• chore(release): v0.1.118-rc.3 [skip ci]

• feat(examples): triggers-demo-go end-to-end demo + Go skill docs (#516)

Go counterpart to examples/triggers-demo/ (Python) and examples/triggers-demo-ts/ (TypeScript), driven by the same unmodified scripts/fire-events.sh and producing equivalent memory writes.

New examples/triggers-demo-go/:

• main.go with three deterministic reasoners: handle_payment (stripe, with a Transform flattening data.object), handle_pr (github), handle_tick (cron). Memory keys and record shapes match the Python demo exactly, so the UI surfaces read the same data regardless of which demo is running.
• Dockerfile: multi-stage build against the in-tree SDK, CGO disabled, alpine runtime with ca-certificates for outbound HTTPS.
• docker-compose.yml: control plane plus Go agent, sharing the demo secrets so signature verification roundtrips with no external configuration. DID enabled so the run-detail trigger enrichment has a VC chain to walk.
• scripts/fire-events.sh: byte-identical copy of the Python demo's script.
• README.md: quick start, UI tour, memory-key table, and a note on the script's Slack/HMAC/Bearer sections (they target a handle_inbound catch-all that this demo, like the TypeScript one, does not define).
• main_test.go: 24 tests exercising the reasoners through triggers.SimulateEvent / SimulateSchedule / LoadFixture, doubling as a worked example of testing trigger reasoners with no control plane.

The reasoners are declared as eventReasoner / scheduleReasoner values that main() iterates, rather than inline inside registration calls, so the handlers and trigger wiring are testable without reaching into SDK internals.

Collapses the duplicated transform matcher flagged in #915: SimulateEvent now uses the shared ApplyTransform and NewContext from dispatch.go instead of a private copy. This fixes a real defect the demo surfaced — the helpers attached the context under a private key while handlers read it via FromContext, so a handler that worked in production saw a nil context under test. Added a regression test asserting FromContext sees what SimulateEvent attaches.

skills/agentfield-multi-reasoner-builder/references/triggers.md gains a Go section covering both declaration forms, Context fields, the FromContext accessor, the test helpers, and a three-way comparison table. Every Go snippet was compiled against the real SDK. Updated the shared wire-format and envelope sections from "both SDKs" to "all three".

Verified: build, vet, gofmt, and go test -race clean on both modules; docker compose config valid; the Dockerfile's exact steps reproduced with -mod=readonly to confirm go.sum completeness; and the real binary registers all three triggers with the expected wire payload (sources, event types, secret env vars, cron config, accepts_webhook auto-set, code_origin stamped) against a stub control plane. Docker image build itself not run, as no daemon was available in this environment.

Part of #508. Closes #516.

• fix(examples): drop committed Windows binary from triggers-demo-go

Removes examples/triggers-demo-go/triggers-demo-go.exe, an 11 MB build artifact that should never have been committed. It came from compiling the demo to verify it and was swept in by adding the directory rather than naming files.

Adds a .gitignore for the binary so a local build cannot reintroduce it. The example remains buildable from source via the Dockerfile or go build.

Addresses @santoshkumarradha's review on #917.

Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com> (a1dd50b)

Fixed

• Fix(sdk): merge positional multimodal args with user prompt (#923) (243d906)

Fixed

• Fix(ui): only prompt for the admin token when the server enforces one (#920)

• fix(ui): only prompt for the admin token when the server enforces one

The Access management page rendered the admin-token input unconditionally, so deployments that never configured AGENTFIELD_AUTHORIZATION_ADMIN_TOKEN were still asked for a token the server would never read (AdminTokenAuth is a no-op when the configured token is empty). Operators pasted whatever they had — usually the API key — into a field that does nothing.

The governance probe already received the discriminating signal and threw it away: it collapsed the status of GET /api/v1/admin/policies into status !== 404, which also misread 401 (API-key problem) as "admin routes available".

Replace the boolean probe with a tri-state one that deliberately omits X-Admin-Token — since AdminTokenAuth wraps the whole admin group, the tokenless status is authoritative for every admin route:

• 404 → authorization feature disabled (existing banner, no prompt)
• 403 → an admin token is enforced: the only state that shows the prompt; policy/tag queries and mutations are gated until a token is stored
• 200 → no admin token enforced: prompt hidden, admin APIs just work; a stored stale token gets a hint that it can be cleared
• 401 → surfaced as an API-key error instead of "routes available"

Also drop the "unchanged repo default is often admin-secret" hint copy: it is only true for a source checkout run from control-plane/ (or a deploy pointing AGENTFIELD_CONFIG_FILE at the bundled YAML) and misled operators of env-configured deployments into pasting wrong values.

Co-Authored-By: Claude Fable 5 [email protected]

• fix(ui): drop the admin-token field from the login screen

At sign-in the client cannot yet know whether the server enforces an admin token (an unauthenticated probe always 401s), so the "Admin Token (optional)" field was another unconditional prompt inviting wrong pastes. Access management now asks for the token in the one state where it is actually required, so the login screen only asks for the API key.

Co-Authored-By: Claude Fable 5 [email protected]

Co-authored-by: Claude Fable 5 [email protected] (f3e5ec9)

Fixed

• Fix: repair broken quickstart flows — Go docs/examples, docker scaffolds, TS template (#918)

• fix(examples): regenerate stale go.sum in Go example modules

examples/go_agent_nodes and examples/go_harness_demo stopped building when sdk/go added jsonschema/v5 (#750) without their go.sum being refreshed: 'missing go.sum entry for module providing package .../jsonschema/v5'.

Co-Authored-By: Claude Fable 5 [email protected]

• docs(sdk/go): fix broken quickstart and approvals snippets

• Quick Start omitted the required Config.Version, so a copy-paste run exited with 'config.Version is required'; the CLAUDE.md variant also discarded the error from New and nil-panicked on RegisterSkill.
• Approvals snippet called client.New with one return value (it returns (*Client, error)) and passed a nil variadic Option, which panics.

Both snippets now compile and start against the current SDK.

Co-Authored-By: Claude Fable 5 [email protected]

• fix(cli): make 'af init --docker' work for Go and TypeScript scaffolds

The generated docker-compose.yml builds the agent from 'dockerfile: Dockerfile' for every language, but only Python mapped a Dockerfile template — Go and TypeScript scaffolds failed at 'docker compose up'.

• Add go.Dockerfile.tmpl and typescript.Dockerfile.tmpl and map them in GetDockerTemplateFiles; the test now asserts every language maps exactly one Dockerfile instead of codifying the gap.
• Wire AGENT_CALLBACK_URL (set by the compose template) through to the SDK's public URL in the Go and TypeScript scaffolds so the control plane can reach containerized agents; Python already reads it.

Co-Authored-By: Claude Fable 5 [email protected]

• fix(cli): TypeScript scaffold never started from a path containing spaces

The main-module guard compared import.meta.url (percent-encoded) against a raw 'file://' + process.argv[1] string, so under any path with a space (e.g. 'My Projects/') the guard was false and 'npm run dev' exited silently without starting the agent. Compare against pathToFileURL(process.argv[1]).href instead.

Co-Authored-By: Claude Fable 5 [email protected]

• fix(cli): unify scaffold env var names on AGENTFIELD_SERVER / AGENT_NODE_ID

The TypeScript scaffold read AGENTFIELD_URL/AGENT_ID while Python, Go, the docs, and the docker-compose template all use AGENTFIELD_SERVER/ AGENT_NODE_ID — a TS agent given the documented variable silently registered with the default localhost:8080 instead. The old names are kept as fallbacks for existing setups.

Python and Go .env.example advertised AGENTFIELD_CONTROL_PLANE_URL, which nothing reads; they now list the variables the scaffolds honor.

Co-Authored-By: Claude Fable 5 [email protected]

• fix(cli): scaffold polish — accurate next steps, current dep floors, ruff-clean output

• Python next steps (README + af init output) told users to uncomment an ai_config that the scaffold already ships enabled, and the README promised port auto-discovery while main.py pins port 8001.
• Fresh Python scaffolds failed 'ruff check' out of the box (F401 on pydantic imports only used by the commented AI sample); the imports now live inside that commented block.
• TypeScript AI sample never passed the input text into the prompt.
• Dep floors were stale: @agentfield/sdk pinned ^0.1.0 (128 releases old) and requirements.txt had no floor at all. Both now sit at the current release, and bump_version.py updates them on stable releases the same way it already maintains go.mod.tmpl.

Co-Authored-By: Claude Fable 5 [email protected]

• fix(cli): close scaffold gaps found in verification

• .dockerignore.tmpl (shared by all languages) did not exclude node_modules, so a host install was COPY'd over the container's in the TypeScript image, shadowing it with host-arch binaries; it now covers Node and Python artifacts for every language.
• The Go scaffold hardcoded NodeID, silently ignoring the AGENT_NODE_ID override that the generated docker-compose.yml and .env.example set.
• Uncommenting the Python AI sample tripped ruff E402 (mid-file import); the commented import now carries noqa for when it is enabled.
• .env.example now lists AGENT_NODE_ID (and AI_MODEL for Python) to match what the scaffolds actually read.

Co-Authored-By: Claude Fable 5 [email protected]

• chore(scripts): anchor agentfield matcher in bump_version + cover template updaters

The requirements updater matched any line starting with 'agentfield', which would silently rewrite e.g. 'agentfield-cli>=1.0'; it now matches only the agentfield distribution (bare or with a version specifier). Adds regression tests for update_ts_template and the scaffold requirements floor, which previously had no coverage.

Co-Authored-By: Claude Fable 5 [email protected]

Co-authored-by: Claude Fable 5 [email protected] (a5f1573)

Added

• Feat(sdk/go): add Agent.Span for traced in-process sub-executions (#916)

• feat(sdk/go): add Agent.Span for traced in-process sub-executions

Span(ctx, name, input, fn) runs fn as a child execution of the context carried by ctx and emits workflow events so it appears as a node in the run's DAG — CallLocal's lineage and event emission without requiring the target to be a registered reasoner. The child context is injected into fn's ctx, so nested Span/CallLocal/Note calls chain correctly and concurrent Spans become siblings.

Terminal events are sent synchronously (a finished span can never be left dangling in "running"); start events go through a bounded async queue and are shed first under load, which is safe because the events endpoint upserts nodes from the terminal event alone. fn panics emit a "failed" terminal event and re-raise. Input/result payloads are size- capped agent-side (16KB, 4KB preview) since the events path stores them verbatim.

Motivation: the deep-research Go agent runs its internal pipeline stages as plain function calls, so a 12-minute prepare_research_package renders as a single DAG node — while the Python SDK's decorator tracking gave every internal stage its own node. Span restores that trace parity for in-process stages.

Co-Authored-By: Claude Fable 5 [email protected]

• fix(sdk/go): make Agent.Span nil-receiver safe

A nil *Agent runs fn untraced instead of panicking in buildChildContext. Unit tests for wrapped pipeline functions commonly pass a nil agent for stages that never touch the LLM; those must keep working when the stage gains a Span wrapper.

Co-Authored-By: Claude Fable 5 [email protected]

• fix(sdk/go): snapshot span input before handing it to the async event queue

truncateTraceInput returned the caller's map by identity whenever it fit under the size cap, and enqueueSpanStart handed that same map to the background sender goroutine, which json.Marshals it at an unbounded delay. Any caller that mutates the input map after Span entry — including inside fn itself — raced that marshal; without -race that is a process-fatal 'concurrent map read and map write'. Reproduced with go test -race.

The traced input is now deep-copied from its marshaled bytes at Span entry, so events carry a consistent as-of-call snapshot and never alias caller memory. Also aligns enqueueSpanStart's control-plane-URL guard with emitWorkflowEvent's TrimSpace semantics.

Co-Authored-By: Claude Fable 5 [email protected]

• fix(control-plane): survive insert races and shed starts on the events endpoint

The workflow events endpoint did Get→Create with no transaction. Two events for the same execution_id arriving concurrently — exactly what sdk/go Span produces, an async 'running' event racing the synchronous terminal event for any sub-millisecond span — both saw no row and raced to INSERT; the loser got a unique-violation 500 and was dropped. When the loser was the terminal event, the node stayed 'running' forever: the run never left /executions/active, and the stale-execution reaper eventually flipped it to timeout, making a finished run render as timed out. Reproduced live: 550-span flood → 11 nodes stranded, 30 dropped events.

The handler now retries in a bounded loop: a failed create re-reads and merges through the update path (the row necessarily exists once the racing winner lands), and a row that vanishes between read and update re-creates instead of silently no-oping. Terminal-state immutability is unchanged. Post-fix the same flood persists 551/551 nodes terminal with zero errors.

Also backdate StartedAt by duration_ms when a node is created from its terminal event alone (start event shed under queue pressure), so such nodes get a real timeline bar instead of a zero-width one at arrival time.

Co-Authored-By: Claude Fable 5 [email protected]

Co-authored-by: Claude Fable 5 [email protected] (526466c)

• Feat(sdk/go): dispatch envelope unwrap + Context injection (#514) (#914)

• chore(release): v0.1.118-rc.1 [skip ci]

• chore(release): v0.1.118-rc.3 [skip ci]

• feat(sdk/go): dispatch envelope unwrap + Context injection (#514)

Wire the dispatch-side envelope unwrap and *triggers.Context injection into the Go SDK reasoner dispatch paths, so a webhook delivery from the control plane lands on the handler with the transformed payload and populated trigger metadata.

New sdk/go/triggers/dispatch.go:

• IsEnvelope(body) detects the dispatcher shape {event, _meta} (requires _meta.trigger_id, so ordinary payloads are never misread)
• Unwrap(body) peels the event and builds a *Context from _meta, with multi-layout received_at parsing and a now() fallback
• NewContext / FromContext propagate the trigger metadata through context.Context, so HandlerFunc keeps its existing signature (no breaking change for current reasoners)
• ApplyTransform picks the best-matching binding (source match, exact or dotted-prefix event type, specific beats catch-all) and runs its Transform, recovering from a panicking transform to raw input

Wired into every dispatch path via a single applyTriggerDispatch helper so they cannot drift: handleReasoner (sync HTTP), executeReasonerAsync (async goroutine), handleExecute (/execute route), Execute (local), and HandleServerlessEvent (serverless). CallLocal is deliberately left alone, since agent-to-agent calls are direct invocations by definition.

Direct calls are unchanged: input passes through untouched, Transform is skipped, and FromContext returns nil.

Tests: 17 tests in triggers/dispatch_test.go (envelope detection table, unwrap, timestamp fallbacks, context round-trip, transform matching, specificity, prefix match, panic recovery) plus 6 end-to-end tests in agent/agent_dispatch_triggers_test.go proving both shapes through the real Execute path. go build/vet clean, go test -race ./triggers/ green.

Part of #508. Closes #514.

• docs(sdk/go): drop EXPERIMENTAL caveats now that dispatch is wired (#514)

The Context, Transform, and EventOpts.Transform docs carried EXPERIMENTAL notes saying dispatch-time execution and context injection would ship with #514. This PR is that work, so the caveats are now stale and describe the opposite of reality.

Replaced with accurate contracts: FromContext(ctx) retrieval with the nil check distinguishing trigger dispatches from direct calls, Transform running before the handler on dispatches only, and the panic-degrades-to-passthrough behaviour. Kept the VCID caveat, which is still pending the DID/VC chain work tracked separately.

Follows up on @AbirAbbas's review of #906, which flagged these docs as promising behaviour that was not yet wired.

Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com> (f6e6f11)

• Feat(sdk/go): trigger test helpers + fixture library (#515) (#915)

• chore(release): v0.1.118-rc.1 [skip ci]

• chore(release): v0.1.118-rc.3 [skip ci]

• feat(sdk/go): trigger test helpers + fixture library (#515)

Add Go test helpers and a captured-fixture library so reasoners can be unit-tested without spinning up a control plane, mirroring the Python agentfield/testing.py surface.

New sdk/go/triggers/testing.go:

• SimulateEvent(t, handler, opts) builds the *Context the runtime would have produced, applies the matching binding's Transform, and invokes the handler with the context attached
• SimulateSchedule(t, handler, opts) wraps SimulateEvent for cron handlers (source "cron", event type "tick")
• SimulatedContextFrom(ctx) retrieves the synthetic context in tests
• Identifiers default to fresh random values so repeated simulations are independently dedup-safe; every field is overridable
• Transform matching mirrors the other SDKs exactly (source match, exact or dotted-prefix event type, specific beats catch-all) and recovers from a panicking transform to raw input

New sdk/go/triggers/load_fixture.go:

• LoadFixture(t, name) reads a captured payload, accepting "stripe" or "stripe.json"
• RawFixture(t, name) returns undecoded bytes for byte-level assertions
• FixtureNames() lists all six sources for table-driven tests
• Fixtures are embedded with go:embed rather than read from a relative testdata path, so LoadFixture works from any caller's working directory including the module cache

New sdk/go/triggers/testdata/*.json: six fixtures (stripe, github, slack, cron, generic_hmac, generic_bearer) copied byte-for-byte from the Python SDK.

Tests: 61 test cases covering all six fixtures through SimulateEvent, known field values per provider, transform application and source mismatch, identifier overrides and uniqueness, parent-context propagation, handler error propagation, non-object and panicking transforms, and the schedule helper. Includes a parity guard asserting the fixtures are byte-identical to the Python SDK copies, plus a check that FixtureNames() stays in sync with the embedded files. go test -race ./triggers/ green, build/vet clean.

Part of #508. Closes #515.

Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com> (4a7529c)

Added

• Feat(sdk/go): triggers package + OnEvent/OnSchedule sugar (#513) (#906)

• chore(release): v0.1.118-rc.1 [skip ci]

• chore(release): v0.1.118-rc.3 [skip ci]

• feat(sdk/go): triggers package + OnEvent/OnSchedule sugar (#513)

Add the public triggers package and agent-level sugar methods for declaring inbound webhook and cron-schedule bindings on Go SDK reasoners.

New sdk/go/triggers/ package:

• Context type (TriggerID, Source, EventType, EventID, IdempotencyKey, ReceivedAt, VCID) — nil for direct calls, populated by dispatch
• EventOpts / ScheduleOpts configuration structs
• Event() / Schedule() factory functions producing Binding values
• Transform function type for pre-handler event transformation
• BindingKind enum (EventBinding / ScheduleBinding)

New agent sugar (sdk/go/agent/agent_triggers.go):

• Agent.OnEvent(opts, name, handler) — registers a reasoner with an event trigger binding in one call
• Agent.OnSchedule(expression, name, handler, ...OnScheduleOption) — registers a reasoner with a cron trigger binding
• WithTimezone(tz) option for OnSchedule
• withTriggersBinding() bridges triggers.Binding into the existing ReasonerOption machinery
• bindingToWire() converts triggers.Binding to types.TriggerBinding

Modified sdk/go/agent/agent.go:

• Added triggerBindings field to Reasoner (stores Transform for dispatch)
• Import triggers package
• triggerToBinding() now also accepts triggers.Binding (backward compat)

Tests: 9 tests in triggers/triggers_test.go (types, factories, config, transform, defaults) + 6 tests in agent/agent_triggers_test.go (OnEvent, OnSchedule, WithTimezone, parity with WithTriggers, backward compat). All pass. go build/vet clean.

Part of #508. Closes #513.

• fix(sdk/go): address review feedback on triggers package (#513)

1. WithTriggers(triggers.Event(...)) now stores triggerBindings (preserving Transform) alongside the wire binding — parity with OnEvent path. Added regression test.

2. Schedule() always merges expression/timezone into Config, even when custom Config is provided. Prevents silent loss of the cron expression.

3. Renamed ScheduleOpts.Expression to ScheduleOpts.Cron for consistency with the Python and TypeScript SDKs.

4. Removed unused captureCallerOrigin function.

5. Marked Context as EXPERIMENTAL until #514 ships the dispatch injection.

Addresses @santoshkumarradha's review on #906.

• docs(sdk/go): mark triggers Transform as experimental pending #514

The Transform doc claimed the SDK runs Transform(rawEvent) before invoking the reasoner. Nothing on the Go side consumes TriggerBinding.TransformFn yet — the binding stores it and dispatch never executes it, so a reasoner invoked by an inbound event silently receives the raw event.

Restate the doc to match reality and add the same EXPERIMENTAL caveat that Context already carries, on both the Transform type and EventOpts.Transform.

Co-Authored-By: Claude Fable 5 [email protected]

• fix(sdk/go): keep Cron authoritative when merging schedule config

Schedule() seeded the config map with expression/timezone and then merged custom Config over it, so a custom "expression" key silently replaced opts.Cron — the same silent cron loss the previous fix targeted, reached from the other direction.

Merge custom keys first and apply Cron unconditionally afterwards. Timezone precedence is now explicit: opts.Timezone wins, else a custom "timezone" key survives, else the UTC default stands. Malformed and non-object Config is still ignored rather than propagated, which is now documented on ScheduleOpts.Config instead of being an implicit quirk.

Co-Authored-By: Claude Fable 5 [email protected]

• style(sdk/go): gofmt triggers files touched by this PR

Two spots this branch introduced fail gofmt: the trailing blank line left in agent_triggers.go by the captureCallerOrigin removal, and struct-key alignment in triggers_test.go after the Expression -> Cron rename.

Scoped to files this PR already touches; the other pre-existing gofmt offenders under sdk/go are left alone.

Co-Authored-By: Claude Fable 5 [email protected]

Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com> Co-authored-by: Abir Abbas [email protected] Co-authored-by: Claude Fable 5 [email protected] (081b53e)

Fixed

• Fix(sdk/python): add lock timeouts + offload blocking requests fallback (#620) (#904)

• chore(release): v0.1.118-rc.1 [skip ci]

• chore(release): v0.1.118-rc.3 [skip ci]

• fix(sdk/python): add lock timeouts + offload blocking requests fallback (#620)

Slice 4 of #620: prevents indefinite hangs from contended locks and offloads the remaining blocking HTTP call in an async function.

Lock timeouts:

• New agentfield/lock_utils.py: timed_lock() context manager that acquires with a configurable timeout (default 30s, env var AGENTFIELD_LOCK_TIMEOUT_SECONDS) and raises LockTimeoutError with diagnostic info instead of hanging.
• Applied to all lock sites in result_cache.py (11), cost_tracker.py (8), node_logs.py (7) — the 26 highest-contention acquisitions.

Blocking request offload:

• memory_events.py history() fallback: the blocking requests.get() in the async function's ImportError path is now offloaded to loop.run_in_executor() so it doesn't freeze the event loop.
• Removed the ASYNC210 per-file-ignore for memory_events.py (resolved).

Running-loop guard:

• client.execute_sync() now emits a RuntimeWarning when called from within a running event loop, directing users to await execute() instead.

Tests: 7 tests in test_lock_timeout.py covering timeout behaviour, reentrant locks, cross-thread contention, error attributes, and the execute_sync warning. 63 tests pass across the affected test surface.

Part of #620.

• fix(sdk/python): parse AGENTFIELD_LOCK_TIMEOUT_SECONDS defensively

DEFAULT_LOCK_TIMEOUT was resolved with a bare float() at import time, so a malformed value took down import agentfield altogether. The empty-string case is the common one: AGENTFIELD_LOCK_TIMEOUT_SECONDS= in a compose env: block makes float("") raise from init.py -> result_cache.py -> lock_utils.py. A negative value imported fine but broke every lock op, since lock.acquire(timeout=-5) raises ValueError.

Parsing now falls back to 30s for missing, empty, non-numeric, non-positive and non-finite values, warning through the module logger for the cases that look like a misconfiguration. Tests drive a fresh interpreter per value so the import-time path is the one under test.

Co-Authored-By: Claude Fable 5 [email protected]

• fix(sdk/python): keep LockTimeoutError diagnostics visible on 3.11+

On Python 3.11+ asyncio.TimeoutError is TimeoutError, so deriving from it made LockTimeoutError catchable by every except asyncio.TimeoutError up the stack. Agent.call wraps client.execute in asyncio.wait_for and that path goes through the result cache, so a real lock deadlock surfaced as "Execute call timed out" and the holder/wait diagnostics were lost.

Deriving from RuntimeError instead keeps the message intact on all matrix versions. Nothing in the repo catches LockTimeoutError, so no call sites change.

Co-Authored-By: Claude Fable 5 [email protected]

• test(sdk/python): harden execute_sync warning fixture, fix module docstring

The new-based client only avoided real I/O by accident: execution died on a missing caller_agent_id attribute inside except Exception: pass. Give caller_agent_id a value and stub _submit_execution_sync with a sentinel, so the test asserts the RuntimeWarning and proves nothing was submitted — rather than depending on a crash that a class-level default would silence, turning the test into a live POST to localhost:8080 plus a polling loop.

The module docstring also advertised a memory_events.history() test that was never written; say where that change is actually guarded instead.

Co-Authored-By: Claude Fable 5 [email protected]

Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com> Co-authored-by: Abir Abbas [email protected] Co-authored-by: Claude Fable 5 [email protected] (25821fc)

• Fix(sdk/python): harden cross-loop teardown (#907)

• fix(sdk/python): harden cross-loop teardown

• fix(sdk/python): absorb both CPython cross-loop error wordings in task teardown

cancel_and_await_if_same_loop only suppressed the asyncio/tasks.py wording ("attached to a different loop"). CPython raises the same error class with a second wording from asyncio/mixins.py ("is bound to a different event loop") when a cancellation cleanup awaits a Lock/Event/Condition bound to a foreign loop, so that case re-raised out of teardown instead of being absorbed.

Match both wordings. The regression test binds an asyncio.Event to a background thread's loop and awaits it from a cancelled task's cleanup on another loop, reproducing the mixins wording genuinely rather than raising a synthetic RuntimeError; the existing test asserting unrelated RuntimeErrors still propagate is unchanged.

Co-Authored-By: Claude Fable 5 [email protected]

Co-authored-by: Abir Abbas [email protected] Co-authored-by: Claude Fable 5 [email protected] (a2ea9e6)

Other

• Seamless agent use: cloud-aware skill, native Codex install, self-healing skill sync (#910)

• feat(skills): agentfield-use v0.6.0 — cloud-vs-local resolution, contract fetch, entry points only

The skill now resolves which control plane work goes to before anything else: an explicit server or AGENTFIELD_SERVER wins, then the AgentField Desktop cloud config (settings.json cloud.enabled/serverUrl/apiKey, read from the per-OS app-data path), else the local default. A configured but unreachable cloud is a stop-and-report, never a silent fallback — the local and cloud fleets are disjoint. The target is passed explicitly per call (--server/-k), never exported.

Dispatch gains two preconditions: fetch the reasoner's exact contract first (agent-summary / get_reasoner_schema; a vacuous schema means the description IS the contract), and only call entrypoint-tagged or described reasoners — undescribed run_* reasoners are internal pipeline stages. The workspace-handle flow is capability-gated (not every build ships get_workspace_handle), and process-wide serialization (swe-pro) is called out in the concurrency guidance.

Also adds agentfield-personal to the embed sync script's list — edits to that skill previously never reached the embedded mirror.

Co-Authored-By: Claude Fable 5 [email protected]

• fix(skillkit): install Codex skills natively, self-heal broken targets, isolate tests from the real home

Codex reads the cross-tool skills standard (~/.codex/skills/<name>/ SKILL.md), not the AGENTS.override.md rules file the marker-block target wrote — so skills were never visible to Codex at all. The codex target now symlinks into ~/.codex/skills like claude-code, and install/update/ uninstall strip any legacy agentfield marker block (deleting the rules file when only whitespace remains).

Installs self-heal: the already-at-this-version skip now also requires the recorded artifact to be valid — symlinks must resolve inside the skill's canonical store, marker blocks must exist and point at a real SKILL.md, and a recorded method that differs from the target's current one is invalid by definition (which is what migrates legacy Codex entries). Machines in the wild carry state pointing at deleted temp directories; a matching version over a broken artifact previously meant the target was never repaired.

Those deleted temp paths came from skillkit's own tests: they isolated AGENTFIELD_HOME but resolved targets against the real home, so go test rewrote real ~/.codex/AGENTS.override.md files. A package TestMain now pins HOME/USERPROFILE/AGENTFIELD_HOME to temp dirs, and a regression test runs a full InstallAll and fingerprints the real home before/after.

af skill install/update now exit non-zero when any skill or target fails (report printed first), and InstallAll continues past a failing skill instead of abandoning the rest of the catalog.

Co-Authored-By: Claude Fable 5 [email protected]

• fix(agentic): bound agent-summary payloads and return descriptions in reasoner search

GET /api/v1/agentic/agent/:id/summary inlined every execution payload from the last 24h — unbounded, often megabytes — making the best contract-discovery surface unusable for coding agents. The query now excludes payloads, sorts newest-first, and is capped at 500 rows; metrics_24h stays exact over the window while recent_executions is sliced to 20. Reasoner descriptions in the summary now apply the same legacy-metadata fallback as discovery.

Reasoner search results carry the (truncated) description — previously description was scored for ranking and then discarded, forcing a second round trip before dispatch — and the search index now prefers the registered description over the legacy metadata map, which also makes current-SDK descriptions searchable.

Co-Authored-By: Claude Fable 5 [email protected]

• feat(desktop): observable skill sync with re-sync triggers and honest dashboard state

Skill sync previously spawned af skill install with stdio ignored, an empty error handler, and no exit-code check, ran only once at launch, and the dashboard claimed skills were installed based on the settings boolean alone. A new SkillSync module captures output (64KB cap, 120s timeout), records the last sync result, appends to <logs>/skill-sync.log, and serializes concurrent triggers. Syncs run at launch, when installSkills flips on, and after a CLI update. The dashboard row now renders from the recorded result: installed only after an exit-0 sync, a visible failure message otherwise.

Co-Authored-By: Claude Fable 5 [email protected]

• fix(skills): §0 probes the Windows desktop config from WSL

On WSL machines the desktop app usually lives on the Windows side, but a Linux-side settings.json (written by a Linux desktop install, often with no cloud key at all) can exist too. §0's "first file that exists" rule stopped there, never saw the Windows cloud config, and resolved to local — reproducing the exact silent-local-fallback incident this skill exists to prevent.

§0 now checks every path that applies to the machine: a config file without an enabled cloud no longer ends the search, and under WSL (detected via /proc/version) the Windows side is probed at /mnt/c/Users/*/AppData/Roaming/agentfield-desktop/settings.json after the Linux path. First file declaring an enabled cloud wins.

Found by rehearsing the desktop-to-cloud happy path on a WSL+Windows machine where the split-brain is live.

Co-Authored-By: Claude Fable 5 [email protected]

Co-authored-by: Claude Fable 5 [email protected] (0256b37)

Added

• Feat(harness): add Grok Build CLI provider (#861)

• feat(harness): add Grok Build CLI provider

Introduce a first-class harness provider for the local Grok Build CLI so AgentField can drive Grok headless sessions the same way it drives Codex, OpenCode, Gemini, and Claude Code.

The provider wraps the Grok CLI with a PTY (via script) because plain pipes currently fail with "Device not configured", feeds prompts through --prompt-file to avoid argv limits, and normalizes missing token usage to zero so metrics aggregation does not crash on incomplete usage objects.

• fix(harness): use util-linux script -c form on Linux for grok PTY wrapper

script -q /dev/null grok ... only runs the command on BSD/macOS. util-linux script takes at most one positional (the typescript file) and ignores the rest, falling back to spawning $SHELL interactively — so on Linux runners the grok CLI never started and the process sat on a bash prompt until the harness idle watchdog killed it. Verified on util-linux 2.37.2: both script -q /dev/null /bin/echo hi and the ---separated script -q /dev/null -- /bin/echo hi hang on an interactive prompt and never echo.

Route the wrapper through a _pty_command helper that picks the right form per flavor: script -q -e -c "<cmd>" /dev/null on Linux, the existing trailing-argv form on BSD/macOS, and the command untouched on Windows or when script(1) is absent. -e is required on util-linux, otherwise script exits 0 regardless of the child's status and masks every non-zero grok exit from the returncode handling below it.

The -c form goes through a shell, so argv is joined with shlex.quote: grok arguments carry caller-controlled text (--system-prompt-override, project paths) that must reach the CLI verbatim and never be re-interpreted as shell syntax. A round-trip test asserts shlex.split of the generated string equals the original argv for input containing quotes, spaces and shell metacharacters.

Co-Authored-By: Claude Fable 5 [email protected]

Co-authored-by: Abir Abbas [email protected] Co-authored-by: Claude Fable 5 [email protected] (f3bc5c5)

Fixed

• Fix(go-sdk): allow custom HTTP clients for AI requests (#908) (5c369b5)

Chores

• Chore(deps): bump nanoid to 3.3.17 (#900)

Co-authored-by: Abir Abbas [email protected] (5902d71)

Fixed

• Fix(harness): bound CLI output capture at 16MB per stream (#903)

run_cli buffered a child's entire stdout/stderr unbounded — chunk list, joined str, and (in providers) the parsed JSONL event list all live at once, so a runaway stream is held in memory several times over, and N concurrent harness calls multiply that. This was one contributing layer in pr-af's OOM crash (Agent-Field/pr-af#65): 16 concurrent opencode calls, each buffering everything, on one container.

Capture is now bounded per stream via AGENTFIELD_HARNESS_MAX_OUTPUT_BYTES (default 16MB, <=0 disables). Real provider streams are completion-boundary events (hundreds of KB), so normal runs stay byte-identical. On overflow the head (session/model info, first error) and the tail (final result + cumulative usage events) are kept around a truncation marker line; the marker and any partial seam line parse as invalid JSON, which parse_jsonl already skips, so extract_final_text and token extraction keep working on a truncated stream.

Co-authored-by: Claude Fable 5 [email protected] (33979de)

Fixed

• Fix(sdk/python): run destructor cleanup synchronously when no loop is running (#620 follow-up) (#902)

• fix(sdk/python): retain and observe fire_and_forget tasks on a running loop

fire_and_forget()'s running-loop branch did a bare loop.create_task(coro). asyncio only keeps a weak reference to a task, so the task could be garbage-collected mid-flight and silently never complete, and because nobody ever retrieved the result, a failing task printed the noisy "Task exception was never retrieved" traceback on collection — exactly what the thread branch of #899 fixed for the no-loop case.

Hold the task in a module-level set and attach a done callback that drops the reference and logs any failure at debug level, matching the thread branch's message.

Co-Authored-By: Claude Fable 5 [email protected]

• fix(sdk/python): run destructor cleanup synchronously when no loop is running

#899 replaced the bare asyncio.run() in Agent.del with fire_and_forget(). That fixed the running-loop case (previously it raised RuntimeError and the exception was swallowed), but regressed the common destructor case: with no running loop, fire_and_forget() hands the coroutine to a daemon thread, and at interpreter exit that thread is killed before it does any work. AsyncExecutionManager.stop(), the background-task gather and the notification dispatcher shutdown were all silently dropped.

Dispatch on loop presence instead: asyncio.run() when there is no running loop so the cleanup actually completes before del returns, and fire_and_forget() only when a loop is already running.

Co-Authored-By: Claude Fable 5 [email protected]

Co-authored-by: Claude Fable 5 [email protected] (2523fae)

Chores

• Chore(deps): bump js-yaml to 4.3.1 (#898) (5e0da59)

Fixed

• Fix(sdk/python): resolve ResultCache cross-loop deadlock (#623) (#799)

• fix(sdk/python): resolve ResultCache cross-loop deadlock (#623)

ResultCache mixed a threading.RLock (for its data) with loop-bound asyncio primitives (asyncio.Event for shutdown, asyncio.Task for the cleanup loop). When start() and stop() ran on different event loops — which happens when the AgentFieldClient's sync and async execution paths are mixed (#620) — stop() would raise 'got Future attached to a different loop' and could wedge the process waiting on a task it can never await.

Fix: make the cache lifecycle loop-aware.

• Record the event loop the cleanup task/shutdown event are bound to.
• start() is now idempotent on the same loop and rebinds cleanly when called on a new loop, discarding the stale task via call_soon_threadsafe(task.cancel) on its owning loop — never a cross-loop await.
• stop() only awaits the cleanup task when on its owning loop; from a different loop it cancels without awaiting. The cache is always cleared regardless (that path only needs the thread lock).
• Shrink the cleanup loop's critical section so stats logging no longer runs while holding the lock, reducing contention with sync callers.

Adds tests/test_result_cache_deadlock.py covering cross-loop stop, idempotent/rebinding start, concurrent sync access during cleanup, and the disabled-cache no-op path. result_cache.py coverage: 88% -> 95%.

• fix(test): remove unused pytest import (ruff F401)

• fix(sdk/python): extend loop-aware teardown to manager + connection pool (#623)

Addresses review feedback on #799: the same foreign-loop 'task.cancel(); await task' hazard that affected ResultCache also lived in AsyncExecutionManager.stop() and http_connection_manager ConnectionManager.close(). Since client.aclose() flows through manager.stop() -> connection_manager.close(), the end-to-end sync/async mixing case (#620/#623) still raised 'got Future attached to a different loop' one level up, before result_cache.stop() was reached.

Changes:

• New agentfield/async_lifecycle.py with shared loop-aware teardown helpers: current_running_loop(), cancel_task_cross_loop(), cancel_and_await_if_same_loop(). Single source of truth for the safe pattern.
• ResultCache refactored to use the shared helpers (behaviour unchanged).
• AsyncExecutionManager records its owning loop at start(); stop() only awaits background tasks / sets the shutdown Event on that loop, and cancels cross-loop without awaiting otherwise.
• http_connection_manager ConnectionManager records its owning loop at start(); close() takes the async lock + awaits the session only on the owning loop, and on a foreign loop schedules session/connector close on the owning loop via call_soon_threadsafe without awaiting.

Tests: new tests/test_async_lifecycle_deadlock.py covers cross-loop teardown of both components (3 of its 4 checks fail pre-fix on the merge-base). Full result_cache + manager + connection suite green (110 passed); result_cache.py coverage 98%.

• fix(sdk/python): address review feedback — stop/start race + cross-loop lock (#623)

1. ResultCache.stop(): compare-before-clear so a concurrent start() that installs fresh references between the await and the cleanup doesn't get its new task orphaned.

2. AsyncExecutionManager.stop(): skip the _execution_lock section on cross-loop stop — the lock is bound to the owning loop, taking it from a foreign loop raises the same 'got Future attached to a different loop' error one line below the fix.

• fix(test): prevent flaky pending-task warnings in cross-loop tests

Add time.sleep(0.2) before stopping loops to let scheduled cross-loop cancels process, and guard loop.close() against still-running threads. Addresses Copilot review feedback on test reliability.

• fix(sdk/python): schedule real teardown on owning loop for cross-loop stop (#623)

Addresses @santoshkumarradha's review: cross-loop stop/close now schedules the actual teardown work on the owning loop rather than skipping it or mutating state unserialized.

AsyncExecutionManager.stop(): on cross-loop, schedules execution cancellation (under _execution_lock) on the owning loop via call_soon_threadsafe so active executions are actually terminated rather than left dangling.

ConnectionManager._close_cross_loop(): moves _session/_connector mutation inside a coroutine scheduled on the owning loop under the async lock, so an in-flight get_session() on the owning loop never sees half-torn-down state. Only _closed is set immediately (to gate new requests); the rest is serialized on the correct loop. (29498e4)

Testing

• Test: add coverage for GinLogger middleware (#557) (#895)

• test: add coverage for GinLogger middleware (#557)

• test(middleware): cover aborted-status and error-field logging in GinLogger (#557)

Ports the non-200 case from #896 so that draft can close in favor of this one, and additionally asserts the error field is populated via c.Error(), a path neither draft covered. Also restores the missing trailing newline.

Co-Authored-By: Claude Fable 5 [email protected]

Co-authored-by: Abir Abbas [email protected] Co-authored-by: Claude Fable 5 [email protected] (c3fd728)

Added

• Feat(harness): aforge provider + per-run isolation of the schema output file (#891)

• fix(harness): isolate schema output file in a per-run temp dir for cwd-only runs

With only cwd set, every harness call wrote the fixed cwd/.agentfield_output.json, so N concurrent .harness() calls sharing a cwd (e.g. pr-af fanning out 8 review dimensions over one checkout) overwrote each other's output and cleanup deleted the file from under still-running siblings — surfacing as spurious "schema parse failed" zero-finding results. Reuse the project_dir strategy unconditionally: mkdtemp(".agentfield-out-*") under project_dir or cwd, cleaned up per run. Mock providers in the runner tests now learn the output path from the prompt suffix, the same way real coding agents do.

Co-Authored-By: Claude Fable 5 [email protected]

• feat(harness): add aforge coding-agent provider

Registers aforge (github.com/Agent-Field/aforge-v2) as a harness provider alongside claude-code/codex/gemini/opencode. The provider invokes aforge exec --json -w <root> with the prompt over stdin, maps the JSON envelope (text/stop/usage/turns) onto RawResult/Metrics with provider-reported cost, translates model slugs to AFORGE_MODEL (leading openrouter/ prefix stripped) and #variant to AFORGE_EXEC_REASONING, and treats budget/turn-cap landings that still produced a final message as usable results while mapping all other non-zero exits to CRASH. The no-progress watchdog is disabled for this provider because aforge only writes stdout once, at completion. HarnessConfig grows aforge_bin, which is threaded through the runner's options whitelist, the factory, the availability spec (probed via aforge version, needs OPENROUTER_API_KEY), and the doctor.

Co-Authored-By: Claude Fable 5 [email protected]

Co-authored-by: Claude Fable 5 [email protected] (f2dd31f)

Fixed

• Fix(sdk/python): replace bare asyncio.run() with loop-aware helpers (#620) (#899)

• chore(release): v0.1.118-rc.1 [skip ci]

• chore(release): v0.1.118-rc.3 [skip ci]

• fix(sdk/python): replace bare asyncio.run() with loop-aware helpers (#620)

Bare asyncio.run() raises RuntimeError when called from within an already-running event loop (e.g. a sync reasoner dispatched by FastAPI, a serverless handler wrapping an async framework, or a destructor on the loop thread). This is slice 3 of #620.

New module agentfield/run_async.py provides two helpers:

• run_coroutine(coro): blocks until result. If a loop is running, runs in a new thread with its own loop so the caller can safely block.
• fire_and_forget(coro): non-blocking. If a loop is running, creates a task; otherwise spawns a daemon thread.

Applied to all 6 asyncio.run() sites:

• agent.py handle_serverless: run_coroutine() for async reasoners
• agent_serverless.py: same pattern
• agent_cli.py: run_coroutine() (safe for CLI, consistent API)
• agent.py note(): fire_and_forget() replaces manual loop detection + threading (was 15 lines, now 1 line)
• agent.py destructor: fire_and_forget() for cleanup

Tests: 9 tests in test_run_async.py covering both loop-running and no-loop cases, exception propagation, and fire-and-forget semantics. Updated test_agent_core.py to match new dispatch pattern.

Part of #620.

• fix(sdk/python): catch/log exceptions in fire_and_forget daemon thread

Wrap the no-loop fire_and_forget() worker thread with try/except so exceptions are logged at debug level rather than leaking as unhandled thread exception tracebacks. Addresses @santoshkumarradha's review.

Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com> (ce137d1)

Fixed

• Fix(control-plane): make the execution cleanup initial delay testable (#897)

cleanupLoop waited on a hardcoded 30 second timer before its first cleanup pass. That branch could not be reached by any test without sleeping for 30 seconds, so the first-pass path was never exercised.

Move the value to a named constant and hold it in an unexported field that the constructor always populates. Production timing is unchanged: NewExecutionCleanupService is the only construction site, so every instance still gets 30 seconds, and no config or exported API surface is added. Tests in this package can shorten the field directly.

A guard test asserts the default is still 30 seconds, so the value is now pinned by an assertion rather than left as an implicit assumption.

Adds loop coverage for the initial-cleanup branch, the ticker branch, and the RetryStaleWorkflowExecutions error path. Each loop test neutralises the other timer so it can only pass for the right reason.

execution_cleanup.go now reports 100% across all seven functions (cleanupLoop 84.6% -> 100%, performCleanup 97.9% -> 100%). (53bad8c)

Other

• Add request logging middleware for #557

Replace gin's verbose stdout [GIN] logger with a structured zerolog middleware (middleware.GinLogger) that emits one DEBUG-level line per request, keeping default info output free of duplicated, overly verbose request logs. (0a0eaa4)

Chores

• Chore(deps): bump js-yaml (#892)

Bumps the npm_and_yarn group with 1 update in the /desktop directory: js-yaml.

Updates js-yaml from 4.3.0 to 4.3.1

• Changelog
• Commits

updated-dependencies:

• dependency-name: js-yaml dependency-version: 4.3.1 dependency-type: direct:production dependency-group: npm_and_yarn ...

Signed-off-by: dependabot[bot] [email protected] Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> (da525c3)

Other

• Fix macOS Go scaffold happy path (#894) (902737b)

Added

• Feat: make the workspace handle reachable — provision furrow, fix the skill, expose the address (#890)

• feat(af): provision the pinned furrow client from its release

The workspace handle needs furrow on the CALLER's machine, and furrow had no distribution channel, so the only instruction anyone could give was "build it from Rust source" — which meant the feature was unreachable in practice.

Download the pinned release asset into ~/.agentfield/bin, verified against the release's SHA256SUMS and written atomically. Unsupported platforms (Windows has no asset; furrow uses std::os::unix unconditionally) are a clean no-op, and the installed version is recorded beside the binary so bumping the pin actually upgrades machines that already have it.

Co-Authored-By: Claude Fable 5 [email protected]

• feat(af): install furrow with the skill that uses it, and expose af furrow ensure

Provisioning belongs in the install path rather than in documentation that each caller re-implements. Ensure furrow when the agentfield-use skill installs — best-effort, so a failed download never fails the install — and add an explicit af furrow ensure for repair, which does surface the error since someone asking for it by name is owed the failure.

Co-Authored-By: Claude Fable 5 [email protected]

• fix(skills): sync the embedded agentfield-use mirror

The workspace-handle section added in #885 landed in skills/ only, leaving the embedded copy the control plane actually serves 41 lines behind. Three skillkit tests have been failing on main since that merge.

Co-Authored-By: Claude Fable 5 [email protected]

• feat(control-plane): expose furrow address in health

The desktop's workspace-sync probe (PR #885) reads furrow_public_addr from the health response body, but nothing emitted the field, so the probe could never report availability. Emit it from the shared health handler when the FURROW_PUBLIC_ADDR env var is set; omit the key entirely when it is not.

Co-Authored-By: Claude Fable 5 [email protected]

• docs(skill): resolve furrow from where AgentField installs it

furrow has no release channel today, so "use it only if furrow is on PATH" silently disabled the workspace handle for every caller. Point the lookup at ~/.agentfield/bin (where provisioning puts it) and at a node's own vendored copy, and keep the silent-skip when neither exists. Provisioning itself belongs in the install path, not in this document.

Co-Authored-By: Claude Fable 5 [email protected]

• fix(furrow): serialize concurrent installs and stop timing out slow downloads

Two processes running Ensure at once (a desktop skill sync racing a manual af skill install, possibly different af versions) could interleave the binary rename and marker write, leaving an old binary marked as current — permanently skipping the repair. An flock around the whole check-download- install sequence serializes them, and the loser re-checks under the lock so it skips instead of re-downloading.

The 15s client timeout bounded the entire request including the ~7.5MB body, failing spuriously below ~500KB/s. Phase timeouts (dial 10s, TLS 10s, response header 30s) with a 3-minute ceiling replace it.

Co-Authored-By: Claude Fable 5 [email protected]

• fix(skill): make the furrow resolver POSIX sh, honor AGENTFIELD_HOME, bump to 0.5.0

The resolver snippet used {bin,go/bin} brace expansion, which dash leaves literal — any agent running it under sh would never find furrow-dial inside installed packages. Spell the two package dirs out. It also hardcoded ~/.agentfield while provisioning honors AGENTFIELD_HOME, so a custom home could install furrow somewhere the skill never looks.

The catalog says to bump Version on every content change; the furrow sections (here and #885) shipped on 0.4.0, leaving reconcilers no signal.

Co-Authored-By: Claude Fable 5 [email protected]

• fix(deploy): opt the cloud image out of furrow client provisioning

The furrow client is a laptop-side tool — cloud agents get furrowd vendored by their own packages, and nothing in the container clones workspaces. Any skill install run in the container would otherwise pull ~7.5MB from GitHub onto the volume for no consumer.

Co-Authored-By: Claude Fable 5 [email protected]

• feat(desktop): tell users when a cloud upgrade exists, and stop skipping skill sync in cloud mode

Re-run deploy has been a safe upgrade path since source_image pinning, but nothing said an upgrade existed — users had to know the button doubles as one. The panel now compares the deployed pin from Terraform state against the release tag Docker Hub resolves for :latest, shows 'Control plane vX -> vY available', and relabels the action 'Upgrade & redeploy' while one is pending. The connection test also gets a Workspace sync row, kept neutral when the server predates the health field it reads.

syncSkills was skipped whenever a cloud profile was active — a guard the cloud-mode PR added wholesale. Skills (and the furrow client their install provisions) belong to local coding agents regardless of where the control plane runs; a cloud-connected laptop is exactly the machine that needs the workspace client.

Co-Authored-By: Claude Fable 5 [email protected]

• chore(furrow): check or explicitly discard every error lint sees

golangci-lint is advisory in CI, but the new provisioning code should not ship with its own errcheck noise.

Co-Authored-By: Claude Fable 5 [email protected]

• test(furrow): cover the provisioning error paths the patch gate flagged

CI's 80% patch-coverage gate measured the furrow provisioning code at 74%: every error branch (unresolvable home, bin-dir collision, lock acquisition, missing or malformed checksums, failed binary download) and the runtime-platform defaulting path were untested. Exercise each of them, plus 'af furrow ensure' end to end through cobra in both its silent-success and surfaced-failure shapes. The lock test lives behind a unix build tag because only the flock implementation can fail; the flock() syscall error itself stays uncovered rather than contorting the code to inject it.

Co-Authored-By: Claude Fable 5 [email protected]

Co-authored-by: Claude Fable 5 [email protected] (1fc0eb7)

• Feat(desktop): open the workspace-sync port when we deploy a control plane (#885)

• feat(desktop): open the workspace-sync port when we deploy a control plane

SWE-AF can now mirror a build's workspace so the coding agent that started a run can read the files while the run is still going. Reaching it needs one exposed TCP port and an address the node can advertise, and neither exists on a deployment we provision today.

The module now declares a railway_tcp_proxy for port 8802 and sets FURROW_PUBLIC_ADDR from its computed domain and port, so a managed deploy comes with workspace sync already on and nothing for the user to configure or even know about. This stays declarative rather than another GraphQL side-channel like the volume: the pinned provider (0.6.2) has railway_tcp_proxy, so it composes with destroy, and Railway's public GraphQL exposes no TCP-proxy create mutation anyway. A port-targeted HTTP domain is not an option either — railway_service_domain in this provider version has no port attribute.

Workspace sync is an extra and is treated like one. Its outputs are read separately from the ones a deploy needs, so a control plane that is up and reachable is a success whether or not a furrow address came back with it — without that split, a Railway that declined the proxy would have reported the whole deployment as failed. testCloudConnection reports reachability as a status, never a health failure, so a user who brought their own control plane sees no change beyond one passive line.

The agentfield-use skill now tells a coding agent what to do with a workspace_handle when it finds one in a result, and says nothing when there isn't one. That includes the one carve-out from "never POST to an agent's own port": the handle's endpoint is a furrow transport authorized by a per-run token, not the agent's HTTP surface.

Co-Authored-By: Claude Fable 5 [email protected]

• fix(desktop): probe the workspace-sync port without disabling TLS validation

The reachability probe opened a TLS connection with certificate validation turned off, because furrowd's default certificate is self-signed. CodeQL flagged it as a high-severity finding and was right to: the flag is a real security control, and "it is only a probe" is not a reason to switch one off.

A plain TCP connect answers the same question — is anything listening on the advertised workspace-sync port — without weakening anything. Confidentiality here comes from furrow's payload encryption and the per-run token, neither of which this probe is involved in.

Co-Authored-By: Claude Fable 5 [email protected]

• docs(skill): pair with the handle's path verbatim

The handle now carries the run's own store directory rather than the root above it, so appending the run id to it points at nothing. Says so explicitly, since the wrong version fails with a message about a missing HEAD that gives no hint the path was the problem.

Co-Authored-By: Claude Fable 5 [email protected]

• fix(desktop): keep the workspace-sync address stable across deploys

The Railway provider returns the TCP proxy domain as an absolute FQDN when it creates the proxy — "altaria.proxy.rlwy.net." — and without the trailing dot when it later reads it back. Interpolating that value raw meant a fresh deploy wrote FURROW_PUBLIC_ADDR with the dot and the next deploy rewrote it without, and because a changed service variable restarts the service, a redeploy that should have been a no-op bounced the control plane.

Normalising both places the domain is read makes the published address identical on create and on refresh, so the follow-up plan is empty.

Found by applying the module against a real Railway account rather than reading the provider schema: create returned the dotted form, the next plan showed the variable changing underneath it. Both address forms do reach furrowd — a client on this machine cloned a workspace out of a Railway container over each — so this is about deploy idempotence, not reachability.

Co-Authored-By: Claude Opus 5 [email protected]

Co-authored-by: Claude Fable 5 [email protected] (1515101)

Other

• Make agent installs work on cloud deployments, and make re-run deploy a real upgrade path (#887)

• fix(desktop): make re-run deploy a safe, actual upgrade path for remote deployments

Remote deployments were frozen on whatever image existed the day they were created: the Terraform module pinned source_image to the floating :latest, Railway resolves a floating tag once and never re-pulls, and an unchanged string is a no-op apply — so 'Re-run deploy' could never upgrade anything. That is how a deployment from July kept failing SWE-AF installs with the pre-superseded_by 'requires Python >=3.12' error long after the redirect shipped in v0.1.121.

• source_image is now a variable, set at deploy time by resolveCloudImage(), which asks Docker Hub which release tag shares latest's digest and pins that concrete tag. A new release changes the string, the diff redeploys the service (the provider calls redeployAllInstances on update), and re-running deploy becomes the upgrade path.
• A failed lookup returns null and falls back to the pin already recorded in state — never rewriting a working deployment's image to :latest, which would itself have forced a pointless redeploy. Fresh deployments fall back to :latest. The lookup is bounded by AbortSignal.timeout(5000).
• The service resource now ignores changes to its volume attribute. The /data volume is created out-of-band, the provider refreshes it into state, and a re-apply planned the undeclared attribute back to null — which the provider's update handler turns into volumeDelete, silently destroying the control plane's databases, secrets, and installed agents on every re-run. With re-runs promoted to the routine upgrade action, that pre-existing hazard had to close.

Co-Authored-By: Claude Fable 5 [email protected]

• fix(deploy): ship uv in the cloud image so requires-python nodes install in cloud deployments

The cloud image's python3 is Debian bookworm's 3.11.2 and nothing could provision anything newer, so installing an agent node with requires-python

=3.12 on a cloud deployment failed with 'no compatible interpreter is available' — the installer's provisionViaUv path already handles this, but only when uv is on PATH.

Interpreters uv downloads go under /data (UV_PYTHON_INSTALL_DIR): venvs on the volume symlink back to the interpreter, and one left in $HOME would vanish with the container on the next deploy, breaking every venv built from it. The smoke test asserts both the binary and the install dir.

Validated by building the image and installing a fixture node declaring requires-python ">=3.12": before, it reproduces the exact production failure; after, uv provisions CPython 3.12 under /data/uv/python and the install succeeds.

Co-Authored-By: Claude Fable 5 [email protected]

• fix(ci): actually gate required-checks on the cloud image job

The summary job lists control-plane-cloud-image in needs but only ever inspected control-plane-image.result, so with if: always() a cloud image build or smoke-test failure could not fail the aggregate check.

Co-Authored-By: Claude Fable 5 [email protected]

Co-authored-by: Claude Fable 5 [email protected] (75fa55e)

Fixed

• Fix(sdk/python): forward v1 video optional fields regressed by #854 (#886)

PR #854 promoted content/ratio/callback_url/aigc_watermark to named parameters of MiniMaxProvider.generate_video and rejected all four for non-H3 models. Before that, the last three fell through **kwargs into the v1 request body and were forwarded verbatim — callback_url and aigc_watermark are documented v1 fields — so existing Hailuo callers now raise ValueError.

Forward ratio/callback_url/aigc_watermark verbatim on the v1 path again, matching v0.1.123 byte-for-byte. Keep the guard for content only: structured content is a v2 request shape that never worked against v1, so the client-side error stays.

Co-authored-by: Claude Fable 5 [email protected] (de30ec6)

• Fix(ci): build UI images on bookworm and pin desktop security overrides (#868)

• fix(security): close open npm Dependabot vulnerability alerts

Bump transitive overrides across web client, desktop, TypeScript SDK, and mastra-bench lockfiles:

• brace-expansion → 1.1.18 / 2.1.4 / 5.0.9 (CVE-2026-14257, CVE-2026-69152)
• postcss → 8.5.25 (GHSA-fxqj-rqcc-2cmp incomplete sourceMappingURL fix)
• fast-uri → 3.1.5 (backslash authority host confusion)
• undici → 6.28.0 (cookie injection, retry desync, CRLF blob type)

Closes Dependabot alerts #355, #357, #363, #365, #366, #370, #371, #372,

Co-authored-by: Santosh kumar [email protected]

• fix(ci): build web UI on bookworm to avoid lightningcss musl flake

npm ci on alpine intermittently omits lightningcss-linux-*-musl when the lockfile lacks libc metadata, breaking vite build in functional-test image builds. Use glibc Node images for the UI builder stages instead.

Co-authored-by: Santosh kumar [email protected]

Co-authored-by: Cursor Agent [email protected] Co-authored-by: Santosh kumar [email protected] Co-authored-by: Abir Abbas [email protected] (5c7e773)

Fixed

• Fix(security): bump fast-uri, postcss, and hono for Dependabot alerts (#883)

Upgrade overrides/lockfiles to clear open npm advisories:

• fast-uri 3.1.4 → 3.1.5 (GHSA-7p8r-x3mc-p8w7)
• postcss 8.5.18 → 8.5.25 (GHSA-fxqj-rqcc-2cmp)
• hono 4.12.32 → 4.12.34 (GHSA-8j4g-w8fx-2239)

Co-authored-by: Cursor Agent [email protected] Co-authored-by: Santosh kumar [email protected] (cd644a8)

Fixed

• Fix(sdk/go/ai): two follow-ups to the Infron gateway integration (#874) (#884)

• fix(sdk/go/ai): never fabricate zero-token usage from a top-level cost

normalizeNativeCost synthesized an empty Usage{} when a body carried a top-level cost without a usage block. On the streaming path every consumer accumulates usage last-non-nil-wins, so a cost-only chunk arriving after the real usage chunk replaced genuine token counts with zeros — recorded downstream as input=0/output=0 with cost_source "provider", an authoritative-looking row that has lost its tokens. Fold the cost only into a usage block the provider actually sent.

Co-Authored-By: Claude Fable 5 [email protected]

• fix(sdk/go/ai): don't inherit attribution values past their opt-out

Infron attribution fell back to the OpenRouter-scoped site URL and app name but never consulted AGENTFIELD_OPENROUTER_ATTRIBUTION, so values a deployment had explicitly suppressed — often internal hostnames or product names — were sent to a different vendor on the first Infron call. Inherit the values only while OpenRouter attribution is enabled; the Infron defaults apply otherwise.

Co-Authored-By: Claude Fable 5 [email protected]

Co-authored-by: Claude Fable 5 [email protected] (a2dee0b)

Added

• Feat(sdk/go/ai): support the Infron gateway (#874)

• feat(sdk/go/ai): support the Infron gateway

Infron is an OpenAI-compatible inference gateway that serves the standard <provider>/<model> ids, so a model moves across by prefix alone: infron/moonshotai/kimi-k2.6 routes the same model the bare id names.

Follows the provider shape already in this package rather than inventing a new one:

• infron_attribution.go mirrors the existing attribution helper. Infron accepts the same HTTP-Referer / X-Title pair, and the attribution env vars already configured for the existing gateway are honored as fallbacks, so a deployment that already declares itself as 'AgentField AI' keeps that identity after switching gateways.
• Config gains IsInfron(); DefaultConfig() reads INFRON_API_KEY and points at https://llm.onerouter.pro/v1.
• client.go attaches attribution on both the sync and streaming paths.
• marshalRequest opts Infron into native usage accounting and strips the routing-only 'infron/' model prefix before the request goes out (stripInfronPrefix, mirroring the prefix handling on the media path). The gateway serves the bare id, so leaving the prefix on returns 'No available providers for model infron/...'. Only a copy of the Request is rewritten; the caller's Request is untouched.

One real difference is handled rather than papered over: Infron returns the native cost at the top level of the body and of the final stream chunk, rather than nested under usage. Parsed naively that leaves Usage.Cost nil, which the cost tracker reads as 'price unknown' -- usage still recorded, but with no cost and an empty cost_source instead of 'provider'. Response/StreamChunk now carry the top-level field and normalizeNativeCost folds it into Usage.Cost, so every existing consumer keeps reading one place. An explicit usage.cost always wins.

A gateway key that was already honored before Infron existed keeps precedence, so adding an Infron key never reroutes an existing deployment.

llm.onerouter.pro is deliberately NOT added to vouchedRewriteDomains: max_tokens and max_completion_tokens behaved identically in probing and neither could be shown to be enforced, so the conservative legacy max_tokens path stays, per the reasoning already in that comment.

• fix(sdk/go/ai): keep OPENAI_API_KEY precedence over INFRON_API_KEY

DefaultConfig applied the Infron block unconditionally, so an environment with OPENAI_API_KEY set and INFRON_API_KEY added resolved to the Infron key and base URL. That contradicts the guarantee stated in DefaultConfig's own doc comment and in ENVIRONMENT_VARIABLES.md, and it matters because spawned agent processes inherit the parent environment -- one exported INFRON_API_KEY would move every Go agent's traffic and credential to a different gateway.

The existing precedence test cleared OPENAI_API_KEY on its first line, so it only exercised the OpenRouter branch and the gap passed CI green. Adds the regression test for the OpenAI case plus one pinning that Infron still applies when it is the only gateway key set, and names the OpenRouter attribution fallback vars in the docs so operators can audit what feeds the gateway.

Co-authored-by: Abir Abbas [email protected] (64d27aa)

• Feat: add a Parallel search option to the deep research example (#863)

Signed-off-by: georgeatparallel [email protected] Co-authored-by: Santosh kumar [email protected] (f721fca)