feat(dev): supervise every runtime in project dev by tejaskash · Pull Request #2041 · aws/agentcore-cli

tejaskash · 2026-08-19T19:12:12Z

Review after #1980

What this does

agentcore project dev without --agent runs all of the project's runtimes at once. Each agent's output is attributed ([orders] ... in the terminal, an agent field in NDJSON); one agent crashing reports its failure and leaves the others running; Ctrl-C tears the whole set down. Selecting one runtime (--agent, or a single-runtime project) keeps the direct path where a crash fails the command — scripts and CI rely on that exit code.

How

DevSupervisor (core/dev/supervisor.ts) owns per-agent lifecycle: an abort scope per agent chained to the command signal, a per-agent state machine (idle/starting/running/failed) with start dedup, TCP readiness probing, and one merged attributed event stream. It's also what the Inspector UI slice (next PR) drives via its start-on-demand API. Ports resolve sequentially — concurrent resolution raced two HTTP agents onto the same port.

Verification

Supervisor: state machine, dedup, retry-after-failure, merge, crash isolation, abort chaining.
Handler with fake runners + fake collector, including runners that bind real TCP listeners so the readiness probe runs for real.
End to end: built the bundle, scaffolded a project, added a second runtime to agentcore.json, ran project dev — both agents started on distinct ports with attributed logs, both answered real Bedrock invocations, and the trace directory holds spans for both service names (per-runtime OTEL_SERVICE_NAME verified end to end). SIGTERM exits 130.

tsc/lint/format clean.

codecov-commenter · 2026-08-19T19:53:59Z

Codecov Report

❌ Patch coverage is 98.79518% with 3 lines in your changes missing coverage. Please review.
✅ Project coverage is 97.37%. Comparing base (e51a676) to head (ec9377c).
⚠️ Report is 1 commits behind head on refactor.

Files with missing lines	Patch %	Lines
src/handlers/project/dev/index.ts	95.94%	3 Missing ⚠️

Additional details and impacted files

@@ Coverage Diff @@## refactor #2041 +/- ##
==========================================
Coverage 97.36% 97.37% ==========================================
Files 424 425 +1 Lines 25571 25761 +190 ==========================================
+ Hits 24897 25084 +187 - Misses 674 677 +3

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.

Hweinstock · 2026-08-24T16:45:12Z

agentcore project dev without --agent runs all of the project's runtimes at once.

Is this the behavior on the publicly released CLI? I thought we always had to select one?

The #1980 rebase carried a single-runtime host check (runtime.build) into the multi-agent dev handler, where the variable is the runtimes array. Bind all interfaces when any selected runtime runs in a container.

…batch events() spliced the queue, yielded each event, then installed its wake callback and blocked. A push landing during the yields ran while wake was undefined, so its wake was a no-op, and the newly installed waiter never noticed the queued event until the next push or shutdown. Re-check the queue inside the wait so a queued event resolves immediately.

…s review Remove the runSingleRuntime special case so every runtime, one or many, runs through DevSupervisor. Start all agents concurrently and stream their output live rather than awaiting each readiness first. End the event loop when no agent is starting or running, and exit non-zero when none started. Make start() async so an unknown name rejects rather than throwing synchronously. Time readiness off the agent's last output (idle timeout) so a slow build that keeps logging is not killed. Move waitForPort into src/io/port.ts to keep socket IO in the IO layer. Rename the environment accessor to getDevEnvVarsForRuntime and SupervisedEvent.agent to agentName.

agentcore-devx-automation · 2026-08-25T19:42:24Z

Claude Security Review: no high-confidence findings. (run)

agentcore-devx-automation · 2026-08-25T19:46:00Z

Claude Security Review: no high-confidence findings. (run)

Split renderEvent into renderAgentEvent (agent name required) and renderStatus for command-level trace-collector lines, so agent output is always tagged with [name] and the agent-optional branches are gone. Addresses review feedback on #2041.

agentcore-devx-automation · 2026-08-25T20:16:40Z

Claude Security Review: no high-confidence findings. (run)

tejaskash · 2026-08-19T19:12:12Z

Review after #1980

What this does

agentcore project dev without --agent runs all of the project's runtimes at once. Each agent's output is attributed ([orders] ... in the terminal, an agent field in NDJSON); one agent crashing reports its failure and leaves the others running; Ctrl-C tears the whole set down. Selecting one runtime (--agent, or a single-runtime project) keeps the direct path where a crash fails the command — scripts and CI rely on that exit code.

How

DevSupervisor (core/dev/supervisor.ts) owns per-agent lifecycle: an abort scope per agent chained to the command signal, a per-agent state machine (idle/starting/running/failed) with start dedup, TCP readiness probing, and one merged attributed event stream. It's also what the Inspector UI slice (next PR) drives via its start-on-demand API. Ports resolve sequentially — concurrent resolution raced two HTTP agents onto the same port.

Verification

Supervisor: state machine, dedup, retry-after-failure, merge, crash isolation, abort chaining.
Handler with fake runners + fake collector, including runners that bind real TCP listeners so the readiness probe runs for real.
End to end: built the bundle, scaffolded a project, added a second runtime to agentcore.json, ran project dev — both agents started on distinct ports with attributed logs, both answered real Bedrock invocations, and the trace directory holds spans for both service names (per-runtime OTEL_SERVICE_NAME verified end to end). SIGTERM exits 130.

tsc/lint/format clean.

codecov-commenter · 2026-08-19T19:53:59Z

Codecov Report

❌ Patch coverage is 98.79518% with 3 lines in your changes missing coverage. Please review.
✅ Project coverage is 97.37%. Comparing base (e51a676) to head (ec9377c).
⚠️ Report is 1 commits behind head on refactor.

Files with missing lines	Patch %	Lines
src/handlers/project/dev/index.ts	95.94%	3 Missing ⚠️

Additional details and impacted files

@@ Coverage Diff @@## refactor #2041 +/- ##
==========================================
Coverage 97.36% 97.37% ==========================================
Files 424 425 +1 Lines 25571 25761 +190 ==========================================
+ Hits 24897 25084 +187 - Misses 674 677 +3

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.

Hweinstock · 2026-08-24T16:45:12Z

agentcore project dev without --agent runs all of the project's runtimes at once.

Is this the behavior on the publicly released CLI? I thought we always had to select one?

The #1980 rebase carried a single-runtime host check (runtime.build) into the multi-agent dev handler, where the variable is the runtimes array. Bind all interfaces when any selected runtime runs in a container.

…batch events() spliced the queue, yielded each event, then installed its wake callback and blocked. A push landing during the yields ran while wake was undefined, so its wake was a no-op, and the newly installed waiter never noticed the queued event until the next push or shutdown. Re-check the queue inside the wait so a queued event resolves immediately.

…s review Remove the runSingleRuntime special case so every runtime, one or many, runs through DevSupervisor. Start all agents concurrently and stream their output live rather than awaiting each readiness first. End the event loop when no agent is starting or running, and exit non-zero when none started. Make start() async so an unknown name rejects rather than throwing synchronously. Time readiness off the agent's last output (idle timeout) so a slow build that keeps logging is not killed. Move waitForPort into src/io/port.ts to keep socket IO in the IO layer. Rename the environment accessor to getDevEnvVarsForRuntime and SupervisedEvent.agent to agentName.

agentcore-devx-automation · 2026-08-25T19:42:24Z

Claude Security Review: no high-confidence findings. (run)

agentcore-devx-automation · 2026-08-25T19:46:00Z

Claude Security Review: no high-confidence findings. (run)

Split renderEvent into renderAgentEvent (agent name required) and renderStatus for command-level trace-collector lines, so agent output is always tagged with [name] and the agent-optional branches are gone. Addresses review feedback on #2041.

agentcore-devx-automation · 2026-08-25T20:16:40Z

Claude Security Review: no high-confidence findings. (run)

tejaskash · 2026-08-19T19:12:12Z

Review after #1980

What this does

agentcore project dev without --agent runs all of the project's runtimes at once. Each agent's output is attributed ([orders] ... in the terminal, an agent field in NDJSON); one agent crashing reports its failure and leaves the others running; Ctrl-C tears the whole set down. Selecting one runtime (--agent, or a single-runtime project) keeps the direct path where a crash fails the command — scripts and CI rely on that exit code.

How

DevSupervisor (core/dev/supervisor.ts) owns per-agent lifecycle: an abort scope per agent chained to the command signal, a per-agent state machine (idle/starting/running/failed) with start dedup, TCP readiness probing, and one merged attributed event stream. It's also what the Inspector UI slice (next PR) drives via its start-on-demand API. Ports resolve sequentially — concurrent resolution raced two HTTP agents onto the same port.

Verification

Supervisor: state machine, dedup, retry-after-failure, merge, crash isolation, abort chaining.
Handler with fake runners + fake collector, including runners that bind real TCP listeners so the readiness probe runs for real.
End to end: built the bundle, scaffolded a project, added a second runtime to agentcore.json, ran project dev — both agents started on distinct ports with attributed logs, both answered real Bedrock invocations, and the trace directory holds spans for both service names (per-runtime OTEL_SERVICE_NAME verified end to end). SIGTERM exits 130.

tsc/lint/format clean.

codecov-commenter · 2026-08-19T19:53:59Z

Codecov Report

❌ Patch coverage is 98.79518% with 3 lines in your changes missing coverage. Please review.
✅ Project coverage is 97.37%. Comparing base (e51a676) to head (ec9377c).
⚠️ Report is 1 commits behind head on refactor.

Files with missing lines	Patch %	Lines
src/handlers/project/dev/index.ts	95.94%	3 Missing ⚠️

Additional details and impacted files

@@ Coverage Diff @@## refactor #2041 +/- ##
==========================================
Coverage 97.36% 97.37% ==========================================
Files 424 425 +1 Lines 25571 25761 +190 ==========================================
+ Hits 24897 25084 +187 - Misses 674 677 +3

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.

Hweinstock · 2026-08-24T16:45:12Z

agentcore project dev without --agent runs all of the project's runtimes at once.

Is this the behavior on the publicly released CLI? I thought we always had to select one?

The #1980 rebase carried a single-runtime host check (runtime.build) into the multi-agent dev handler, where the variable is the runtimes array. Bind all interfaces when any selected runtime runs in a container.

…batch events() spliced the queue, yielded each event, then installed its wake callback and blocked. A push landing during the yields ran while wake was undefined, so its wake was a no-op, and the newly installed waiter never noticed the queued event until the next push or shutdown. Re-check the queue inside the wait so a queued event resolves immediately.

…s review Remove the runSingleRuntime special case so every runtime, one or many, runs through DevSupervisor. Start all agents concurrently and stream their output live rather than awaiting each readiness first. End the event loop when no agent is starting or running, and exit non-zero when none started. Make start() async so an unknown name rejects rather than throwing synchronously. Time readiness off the agent's last output (idle timeout) so a slow build that keeps logging is not killed. Move waitForPort into src/io/port.ts to keep socket IO in the IO layer. Rename the environment accessor to getDevEnvVarsForRuntime and SupervisedEvent.agent to agentName.

agentcore-devx-automation · 2026-08-25T19:42:24Z

Claude Security Review: no high-confidence findings. (run)

agentcore-devx-automation · 2026-08-25T19:46:00Z

Claude Security Review: no high-confidence findings. (run)

Split renderEvent into renderAgentEvent (agent name required) and renderStatus for command-level trace-collector lines, so agent output is always tagged with [name] and the agent-optional branches are gone. Addresses review feedback on #2041.

agentcore-devx-automation · 2026-08-25T20:16:40Z

Claude Security Review: no high-confidence findings. (run)

tejaskash · 2026-08-19T19:12:12Z

Review after #1980

What this does

agentcore project dev without --agent runs all of the project's runtimes at once. Each agent's output is attributed ([orders] ... in the terminal, an agent field in NDJSON); one agent crashing reports its failure and leaves the others running; Ctrl-C tears the whole set down. Selecting one runtime (--agent, or a single-runtime project) keeps the direct path where a crash fails the command — scripts and CI rely on that exit code.

How

DevSupervisor (core/dev/supervisor.ts) owns per-agent lifecycle: an abort scope per agent chained to the command signal, a per-agent state machine (idle/starting/running/failed) with start dedup, TCP readiness probing, and one merged attributed event stream. It's also what the Inspector UI slice (next PR) drives via its start-on-demand API. Ports resolve sequentially — concurrent resolution raced two HTTP agents onto the same port.

Verification

Supervisor: state machine, dedup, retry-after-failure, merge, crash isolation, abort chaining.
Handler with fake runners + fake collector, including runners that bind real TCP listeners so the readiness probe runs for real.
End to end: built the bundle, scaffolded a project, added a second runtime to agentcore.json, ran project dev — both agents started on distinct ports with attributed logs, both answered real Bedrock invocations, and the trace directory holds spans for both service names (per-runtime OTEL_SERVICE_NAME verified end to end). SIGTERM exits 130.

tsc/lint/format clean.

codecov-commenter · 2026-08-19T19:53:59Z

Codecov Report

❌ Patch coverage is 98.79518% with 3 lines in your changes missing coverage. Please review.
✅ Project coverage is 97.37%. Comparing base (e51a676) to head (ec9377c).
⚠️ Report is 1 commits behind head on refactor.

Files with missing lines	Patch %	Lines
src/handlers/project/dev/index.ts	95.94%	3 Missing ⚠️

Additional details and impacted files

@@ Coverage Diff @@## refactor #2041 +/- ##
==========================================
Coverage 97.36% 97.37% ==========================================
Files 424 425 +1 Lines 25571 25761 +190 ==========================================
+ Hits 24897 25084 +187 - Misses 674 677 +3

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.

Hweinstock · 2026-08-24T16:45:12Z

agentcore project dev without --agent runs all of the project's runtimes at once.

Is this the behavior on the publicly released CLI? I thought we always had to select one?

The #1980 rebase carried a single-runtime host check (runtime.build) into the multi-agent dev handler, where the variable is the runtimes array. Bind all interfaces when any selected runtime runs in a container.

…batch events() spliced the queue, yielded each event, then installed its wake callback and blocked. A push landing during the yields ran while wake was undefined, so its wake was a no-op, and the newly installed waiter never noticed the queued event until the next push or shutdown. Re-check the queue inside the wait so a queued event resolves immediately.

…s review Remove the runSingleRuntime special case so every runtime, one or many, runs through DevSupervisor. Start all agents concurrently and stream their output live rather than awaiting each readiness first. End the event loop when no agent is starting or running, and exit non-zero when none started. Make start() async so an unknown name rejects rather than throwing synchronously. Time readiness off the agent's last output (idle timeout) so a slow build that keeps logging is not killed. Move waitForPort into src/io/port.ts to keep socket IO in the IO layer. Rename the environment accessor to getDevEnvVarsForRuntime and SupervisedEvent.agent to agentName.

agentcore-devx-automation · 2026-08-25T19:42:24Z

Claude Security Review: no high-confidence findings. (run)

agentcore-devx-automation · 2026-08-25T19:46:00Z

Claude Security Review: no high-confidence findings. (run)

Split renderEvent into renderAgentEvent (agent name required) and renderStatus for command-level trace-collector lines, so agent output is always tagged with [name] and the agent-optional branches are gone. Addresses review feedback on #2041.

agentcore-devx-automation · 2026-08-25T20:16:40Z

Claude Security Review: no high-confidence findings. (run)

tejaskash · 2026-08-19T19:12:12Z

Review after #1980

What this does

agentcore project dev without --agent runs all of the project's runtimes at once. Each agent's output is attributed ([orders] ... in the terminal, an agent field in NDJSON); one agent crashing reports its failure and leaves the others running; Ctrl-C tears the whole set down. Selecting one runtime (--agent, or a single-runtime project) keeps the direct path where a crash fails the command — scripts and CI rely on that exit code.

How

DevSupervisor (core/dev/supervisor.ts) owns per-agent lifecycle: an abort scope per agent chained to the command signal, a per-agent state machine (idle/starting/running/failed) with start dedup, TCP readiness probing, and one merged attributed event stream. It's also what the Inspector UI slice (next PR) drives via its start-on-demand API. Ports resolve sequentially — concurrent resolution raced two HTTP agents onto the same port.

Verification

Supervisor: state machine, dedup, retry-after-failure, merge, crash isolation, abort chaining.
Handler with fake runners + fake collector, including runners that bind real TCP listeners so the readiness probe runs for real.
End to end: built the bundle, scaffolded a project, added a second runtime to agentcore.json, ran project dev — both agents started on distinct ports with attributed logs, both answered real Bedrock invocations, and the trace directory holds spans for both service names (per-runtime OTEL_SERVICE_NAME verified end to end). SIGTERM exits 130.

tsc/lint/format clean.

codecov-commenter · 2026-08-19T19:53:59Z

Codecov Report

❌ Patch coverage is 98.79518% with 3 lines in your changes missing coverage. Please review.
✅ Project coverage is 97.37%. Comparing base (e51a676) to head (ec9377c).
⚠️ Report is 1 commits behind head on refactor.

Files with missing lines	Patch %	Lines
src/handlers/project/dev/index.ts	95.94%	3 Missing ⚠️

Additional details and impacted files

@@ Coverage Diff @@## refactor #2041 +/- ##
==========================================
Coverage 97.36% 97.37% ==========================================
Files 424 425 +1 Lines 25571 25761 +190 ==========================================
+ Hits 24897 25084 +187 - Misses 674 677 +3

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.

Hweinstock · 2026-08-24T16:45:12Z

agentcore project dev without --agent runs all of the project's runtimes at once.

Is this the behavior on the publicly released CLI? I thought we always had to select one?

The #1980 rebase carried a single-runtime host check (runtime.build) into the multi-agent dev handler, where the variable is the runtimes array. Bind all interfaces when any selected runtime runs in a container.

…batch events() spliced the queue, yielded each event, then installed its wake callback and blocked. A push landing during the yields ran while wake was undefined, so its wake was a no-op, and the newly installed waiter never noticed the queued event until the next push or shutdown. Re-check the queue inside the wait so a queued event resolves immediately.

…s review Remove the runSingleRuntime special case so every runtime, one or many, runs through DevSupervisor. Start all agents concurrently and stream their output live rather than awaiting each readiness first. End the event loop when no agent is starting or running, and exit non-zero when none started. Make start() async so an unknown name rejects rather than throwing synchronously. Time readiness off the agent's last output (idle timeout) so a slow build that keeps logging is not killed. Move waitForPort into src/io/port.ts to keep socket IO in the IO layer. Rename the environment accessor to getDevEnvVarsForRuntime and SupervisedEvent.agent to agentName.

agentcore-devx-automation · 2026-08-25T19:42:24Z

Claude Security Review: no high-confidence findings. (run)

agentcore-devx-automation · 2026-08-25T19:46:00Z

Claude Security Review: no high-confidence findings. (run)

Split renderEvent into renderAgentEvent (agent name required) and renderStatus for command-level trace-collector lines, so agent output is always tagged with [name] and the agent-optional branches are gone. Addresses review feedback on #2041.

agentcore-devx-automation · 2026-08-25T20:16:40Z

Claude Security Review: no high-confidence findings. (run)

tejaskash · 2026-08-19T19:12:12Z

Review after #1980

What this does

agentcore project dev without --agent runs all of the project's runtimes at once. Each agent's output is attributed ([orders] ... in the terminal, an agent field in NDJSON); one agent crashing reports its failure and leaves the others running; Ctrl-C tears the whole set down. Selecting one runtime (--agent, or a single-runtime project) keeps the direct path where a crash fails the command — scripts and CI rely on that exit code.

How

DevSupervisor (core/dev/supervisor.ts) owns per-agent lifecycle: an abort scope per agent chained to the command signal, a per-agent state machine (idle/starting/running/failed) with start dedup, TCP readiness probing, and one merged attributed event stream. It's also what the Inspector UI slice (next PR) drives via its start-on-demand API. Ports resolve sequentially — concurrent resolution raced two HTTP agents onto the same port.

Verification

Supervisor: state machine, dedup, retry-after-failure, merge, crash isolation, abort chaining.
Handler with fake runners + fake collector, including runners that bind real TCP listeners so the readiness probe runs for real.
End to end: built the bundle, scaffolded a project, added a second runtime to agentcore.json, ran project dev — both agents started on distinct ports with attributed logs, both answered real Bedrock invocations, and the trace directory holds spans for both service names (per-runtime OTEL_SERVICE_NAME verified end to end). SIGTERM exits 130.

tsc/lint/format clean.

codecov-commenter · 2026-08-19T19:53:59Z

Codecov Report

❌ Patch coverage is 98.79518% with 3 lines in your changes missing coverage. Please review.
✅ Project coverage is 97.37%. Comparing base (e51a676) to head (ec9377c).
⚠️ Report is 1 commits behind head on refactor.

Files with missing lines	Patch %	Lines
src/handlers/project/dev/index.ts	95.94%	3 Missing ⚠️

Additional details and impacted files

@@ Coverage Diff @@## refactor #2041 +/- ##
==========================================
Coverage 97.36% 97.37% ==========================================
Files 424 425 +1 Lines 25571 25761 +190 ==========================================
+ Hits 24897 25084 +187 - Misses 674 677 +3

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.

Hweinstock · 2026-08-24T16:45:12Z

agentcore project dev without --agent runs all of the project's runtimes at once.

Is this the behavior on the publicly released CLI? I thought we always had to select one?

The #1980 rebase carried a single-runtime host check (runtime.build) into the multi-agent dev handler, where the variable is the runtimes array. Bind all interfaces when any selected runtime runs in a container.

…batch events() spliced the queue, yielded each event, then installed its wake callback and blocked. A push landing during the yields ran while wake was undefined, so its wake was a no-op, and the newly installed waiter never noticed the queued event until the next push or shutdown. Re-check the queue inside the wait so a queued event resolves immediately.

…s review Remove the runSingleRuntime special case so every runtime, one or many, runs through DevSupervisor. Start all agents concurrently and stream their output live rather than awaiting each readiness first. End the event loop when no agent is starting or running, and exit non-zero when none started. Make start() async so an unknown name rejects rather than throwing synchronously. Time readiness off the agent's last output (idle timeout) so a slow build that keeps logging is not killed. Move waitForPort into src/io/port.ts to keep socket IO in the IO layer. Rename the environment accessor to getDevEnvVarsForRuntime and SupervisedEvent.agent to agentName.

agentcore-devx-automation · 2026-08-25T19:42:24Z

Claude Security Review: no high-confidence findings. (run)

agentcore-devx-automation · 2026-08-25T19:46:00Z

Claude Security Review: no high-confidence findings. (run)

Split renderEvent into renderAgentEvent (agent name required) and renderStatus for command-level trace-collector lines, so agent output is always tagged with [name] and the agent-optional branches are gone. Addresses review feedback on #2041.

agentcore-devx-automation · 2026-08-25T20:16:40Z

Claude Security Review: no high-confidence findings. (run)

tejaskash · 2026-08-19T19:12:12Z

Review after #1980

What this does

agentcore project dev without --agent runs all of the project's runtimes at once. Each agent's output is attributed ([orders] ... in the terminal, an agent field in NDJSON); one agent crashing reports its failure and leaves the others running; Ctrl-C tears the whole set down. Selecting one runtime (--agent, or a single-runtime project) keeps the direct path where a crash fails the command — scripts and CI rely on that exit code.

How

DevSupervisor (core/dev/supervisor.ts) owns per-agent lifecycle: an abort scope per agent chained to the command signal, a per-agent state machine (idle/starting/running/failed) with start dedup, TCP readiness probing, and one merged attributed event stream. It's also what the Inspector UI slice (next PR) drives via its start-on-demand API. Ports resolve sequentially — concurrent resolution raced two HTTP agents onto the same port.

Verification

Supervisor: state machine, dedup, retry-after-failure, merge, crash isolation, abort chaining.
Handler with fake runners + fake collector, including runners that bind real TCP listeners so the readiness probe runs for real.
End to end: built the bundle, scaffolded a project, added a second runtime to agentcore.json, ran project dev — both agents started on distinct ports with attributed logs, both answered real Bedrock invocations, and the trace directory holds spans for both service names (per-runtime OTEL_SERVICE_NAME verified end to end). SIGTERM exits 130.

tsc/lint/format clean.

codecov-commenter · 2026-08-19T19:53:59Z

Codecov Report

❌ Patch coverage is 98.79518% with 3 lines in your changes missing coverage. Please review.
✅ Project coverage is 97.37%. Comparing base (e51a676) to head (ec9377c).
⚠️ Report is 1 commits behind head on refactor.

Files with missing lines	Patch %	Lines
src/handlers/project/dev/index.ts	95.94%	3 Missing ⚠️

Additional details and impacted files

@@ Coverage Diff @@## refactor #2041 +/- ##
==========================================
Coverage 97.36% 97.37% ==========================================
Files 424 425 +1 Lines 25571 25761 +190 ==========================================
+ Hits 24897 25084 +187 - Misses 674 677 +3

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.

Hweinstock · 2026-08-24T16:45:12Z

agentcore project dev without --agent runs all of the project's runtimes at once.

Is this the behavior on the publicly released CLI? I thought we always had to select one?

The #1980 rebase carried a single-runtime host check (runtime.build) into the multi-agent dev handler, where the variable is the runtimes array. Bind all interfaces when any selected runtime runs in a container.

…batch events() spliced the queue, yielded each event, then installed its wake callback and blocked. A push landing during the yields ran while wake was undefined, so its wake was a no-op, and the newly installed waiter never noticed the queued event until the next push or shutdown. Re-check the queue inside the wait so a queued event resolves immediately.

…s review Remove the runSingleRuntime special case so every runtime, one or many, runs through DevSupervisor. Start all agents concurrently and stream their output live rather than awaiting each readiness first. End the event loop when no agent is starting or running, and exit non-zero when none started. Make start() async so an unknown name rejects rather than throwing synchronously. Time readiness off the agent's last output (idle timeout) so a slow build that keeps logging is not killed. Move waitForPort into src/io/port.ts to keep socket IO in the IO layer. Rename the environment accessor to getDevEnvVarsForRuntime and SupervisedEvent.agent to agentName.

agentcore-devx-automation · 2026-08-25T19:42:24Z

Claude Security Review: no high-confidence findings. (run)

agentcore-devx-automation · 2026-08-25T19:46:00Z

Claude Security Review: no high-confidence findings. (run)

Split renderEvent into renderAgentEvent (agent name required) and renderStatus for command-level trace-collector lines, so agent output is always tagged with [name] and the agent-optional branches are gone. Addresses review feedback on #2041.

agentcore-devx-automation · 2026-08-25T20:16:40Z

Claude Security Review: no high-confidence findings. (run)

tejaskash · 2026-08-19T19:12:12Z

Review after #1980

What this does

agentcore project dev without --agent runs all of the project's runtimes at once. Each agent's output is attributed ([orders] ... in the terminal, an agent field in NDJSON); one agent crashing reports its failure and leaves the others running; Ctrl-C tears the whole set down. Selecting one runtime (--agent, or a single-runtime project) keeps the direct path where a crash fails the command — scripts and CI rely on that exit code.

How

DevSupervisor (core/dev/supervisor.ts) owns per-agent lifecycle: an abort scope per agent chained to the command signal, a per-agent state machine (idle/starting/running/failed) with start dedup, TCP readiness probing, and one merged attributed event stream. It's also what the Inspector UI slice (next PR) drives via its start-on-demand API. Ports resolve sequentially — concurrent resolution raced two HTTP agents onto the same port.

Verification

Supervisor: state machine, dedup, retry-after-failure, merge, crash isolation, abort chaining.
Handler with fake runners + fake collector, including runners that bind real TCP listeners so the readiness probe runs for real.
End to end: built the bundle, scaffolded a project, added a second runtime to agentcore.json, ran project dev — both agents started on distinct ports with attributed logs, both answered real Bedrock invocations, and the trace directory holds spans for both service names (per-runtime OTEL_SERVICE_NAME verified end to end). SIGTERM exits 130.

tsc/lint/format clean.

codecov-commenter · 2026-08-19T19:53:59Z

Codecov Report

❌ Patch coverage is 98.79518% with 3 lines in your changes missing coverage. Please review.
✅ Project coverage is 97.37%. Comparing base (e51a676) to head (ec9377c).
⚠️ Report is 1 commits behind head on refactor.

Files with missing lines	Patch %	Lines
src/handlers/project/dev/index.ts	95.94%	3 Missing ⚠️

Additional details and impacted files

@@ Coverage Diff @@## refactor #2041 +/- ##
==========================================
Coverage 97.36% 97.37% ==========================================
Files 424 425 +1 Lines 25571 25761 +190 ==========================================
+ Hits 24897 25084 +187 - Misses 674 677 +3

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.

Hweinstock · 2026-08-24T16:45:12Z

agentcore project dev without --agent runs all of the project's runtimes at once.

Is this the behavior on the publicly released CLI? I thought we always had to select one?

The #1980 rebase carried a single-runtime host check (runtime.build) into the multi-agent dev handler, where the variable is the runtimes array. Bind all interfaces when any selected runtime runs in a container.

…batch events() spliced the queue, yielded each event, then installed its wake callback and blocked. A push landing during the yields ran while wake was undefined, so its wake was a no-op, and the newly installed waiter never noticed the queued event until the next push or shutdown. Re-check the queue inside the wait so a queued event resolves immediately.

…s review Remove the runSingleRuntime special case so every runtime, one or many, runs through DevSupervisor. Start all agents concurrently and stream their output live rather than awaiting each readiness first. End the event loop when no agent is starting or running, and exit non-zero when none started. Make start() async so an unknown name rejects rather than throwing synchronously. Time readiness off the agent's last output (idle timeout) so a slow build that keeps logging is not killed. Move waitForPort into src/io/port.ts to keep socket IO in the IO layer. Rename the environment accessor to getDevEnvVarsForRuntime and SupervisedEvent.agent to agentName.

agentcore-devx-automation · 2026-08-25T19:42:24Z

Claude Security Review: no high-confidence findings. (run)

agentcore-devx-automation · 2026-08-25T19:46:00Z

Claude Security Review: no high-confidence findings. (run)

Split renderEvent into renderAgentEvent (agent name required) and renderStatus for command-level trace-collector lines, so agent output is always tagged with [name] and the agent-optional branches are gone. Addresses review feedback on #2041.

agentcore-devx-automation · 2026-08-25T20:16:40Z

Claude Security Review: no high-confidence findings. (run)

github-actionsBot added the size/l PR size: L label Aug 19, 2026

tejaskash had a problem deploying to e2e-testing August 19, 2026 19:12 — with GitHub Actions Failure

github-actionsBot added agentcore-harness-reviewing AgentCore Harness review in progress and removed agentcore-harness-reviewing AgentCore Harness review in progress labels Aug 19, 2026

tejaskash force-pushed the feat/dev-supervisor branch from 1864da4 to c539836Compare August 19, 2026 19:51

tejaskash force-pushed the feat/dev-supervisor branch from c539836 to 6080aa0Compare August 20, 2026 16:09

tejaskash force-pushed the feat/dev-supervisor branch from 6080aa0 to 82af18aCompare August 20, 2026 16:18

tejaskash force-pushed the feat/dev-supervisor branch from 82af18a to 2db7743Compare August 21, 2026 15:23

tejaskash force-pushed the feat/dev-supervisor branch from 2db7743 to 136a454Compare August 21, 2026 16:14

tejaskash force-pushed the feat/dev-supervisor branch from 136a454 to 35775f9Compare August 21, 2026 17:51

tejaskash force-pushed the feat/dev-supervisor branch from 35775f9 to a9f11feCompare August 21, 2026 18:55

Base automatically changed from feat/dev-otel-collector to refactorAugust 21, 2026 21:32

tejaskash force-pushed the feat/dev-supervisor branch 2 times, most recently from 7b4b60b to 0ca91f0Compare August 24, 2026 16:13

tejaskash mentioned this pull request Aug 24, 2026
feat(dev): add Agent Inspector HTTP layer (server, security, assets) #2082
Merged

tejaskash force-pushed the feat/dev-supervisor branch 2 times, most recently from fc5db62 to 9904ee1Compare August 24, 2026 21:51

notgitika self-requested a review August 24, 2026 22:14

tejaskash force-pushed the feat/dev-supervisor branch from 9904ee1 to 0a77c0cCompare August 25, 2026 15:06

tejaskash added 2 commits August 25, 2026 15:39

fix(dev): bind collector to 0.0.0.0 when any runtime is a container
a697e15
The #1980 rebase carried a single-runtime host check (runtime.build) into the multi-agent dev handler, where the variable is the runtimes array. Bind all interfaces when any selected runtime runs in a container.

tejaskash force-pushed the feat/dev-supervisor branch from 175d896 to abf8c23Compare August 25, 2026 19:39

github-actionsBot added size/l PR size: L and removed size/l PR size: L labels Aug 25, 2026

agentcore-devx-automationBot added the claude-security-reviewing Claude Code /security-review in progress label Aug 25, 2026

tejaskash dismissed jariy17’s stale review via cf625caAugust 25, 2026 19:41

github-actionsBot added size/l PR size: L and removed size/l PR size: L labels Aug 25, 2026

agentcore-devx-automationBot added claude-security-reviewing Claude Code /security-review in progress and removed claude-security-reviewing Claude Code /security-review in progress labels Aug 25, 2026

agentcore-devx-automationBot removed the claude-security-reviewing Claude Code /security-review in progress label Aug 25, 2026

tejaskash requested review from jariy17 and notgitika August 25, 2026 19:48

github-actionsBot added size/l PR size: L and removed size/l PR size: L labels Aug 25, 2026

agentcore-devx-automationBot added the claude-security-reviewing Claude Code /security-review in progress label Aug 25, 2026

agentcore-devx-automationBot removed the claude-security-reviewing Claude Code /security-review in progress label Aug 25, 2026

jariy17 approved these changes Aug 25, 2026
View reviewed changes

notgitika approved these changes Aug 25, 2026
View reviewed changes

tejaskash merged commit 1d96d1f into refactorAug 25, 2026
22 of 23 checks passed

tejaskash deleted the feat/dev-supervisor branch August 25, 2026 20:25

This was referenced Aug 26, 2026
fix(dev): cancel readiness retries on abort #2109
Merged
bug: windows tests hanging on refactor branch. #2110
Closed

tejaskash mentioned this pull request Aug 28, 2026
feat(dev): supervise every runtime in headless mode #2138
Open