feat: add agentcore feedback command by tejaskash · Pull Request #1321 · aws/agentcore-cli

tejaskash · 2026-05-20T17:21:30Z

Summary

Adds a new top-level agentcore feedback command:

One-shot:agentcore feedback "msg" [--screenshot path] [--json]
Wizard:agentcore feedback (no args) → multi-step Ink TUI

Both modes display the AWS Customer Agreement and require interactive y/N consent before submitting. Bare Enter defaults to No; non-TTY stdin is refused so the consent decision is always explicit.

The wizard walks Message → optional Screenshot → Consent → Submit, with Esc stepping back and message text preserved across declines/retries. Screenshots are validated locally (PNG/JPG/JPEG, ≤100MB) before any network call, then uploaded via presigned URL with SHA-256 checksum headers.

Submissions carry cli-version, os, node-version, and cli-mode metadata. The CLI does not attach AWS account IDs, credentials, project names, or telemetry IDs.

What's in the diff

src/cli/commands/feedback/ — Commander registration, action, TTY consent prompt
src/cli/operations/feedback/ — submission orchestrator, payload builder, presigned-URL + S3 upload + form-ingestion client
src/cli/tui/screens/feedback/ — Ink wizard with useFeedbackFlow hook
src/cli/cli.ts, src/cli/tui/copy.ts — wiring
README.md, docs/commands.md, docs/feedback.md — customer-facing docs

29 unit tests across operations, command, consent-prompt, and TUI hook.

Test plan

npm test -- feedback — 29/29 pass
eslint clean on all new files
tsc --noEmit clean on all new files
Manual prod submission verified end-to-end (message + screenshot, both submit and decline paths)
CLI consent prompt: y/yes/Y/YES accepts; n/no/bare-Enter declines; non-TTY refuses
TUI wizard: all phase transitions, Esc back-stepping, message preservation across consent decline
Plain mode and --json mode for success, declined, no-tty, validation, network error
Bug bash plan attached for testers (5 testers × 6 min)

github-actions · 2026-05-20T17:22:57Z

Package Tarball

aws-agentcore-0.14.1.tgz

How to install

gh release download pr-1321-tarball --repo aws/agentcore-cli --pattern "*.tgz" --dir /tmp/pr-tarball
npm install -g /tmp/pr-tarball/aws-agentcore-0.14.1.tgz

agentcore-devx-automation · 2026-05-20T17:25:24Z

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

agentcore-cli-automation

Thanks for the careful work on consent flow, validation, and docs — the wizard structure and cli/tui mode tagging are nicely done. A few things I'd like to see addressed before merging:

Missing telemetry instrumentation. Per src/cli/telemetry/README.md, every CLI command should emit a command_run metric, but feedback doesn't. Other recent commands (e.g. deploy, remove, create) all use withCommandRunTelemetry / runCliCommand.
Screenshot S3 key is fabricated client-side independent of the presigned URL. The form references a key the client invented rather than the actual upload location.
TUI validates message and screenshot only after consent. Bad input forces the user to redo consent.

Inline comments below.

github-actions · 2026-05-20T17:25:51Z

Coverage Report

Status	Category	Percentage	Covered / Total
🔵	Lines	44.3%	9531 / 21510
🔵	Statements	43.55%	10114 / 23223
🔵	Functions	41.01%	1655 / 4035
🔵	Branches	40.84%	6162 / 15086

Generated in workflow #3182 for commit 91cb5da by the Vitest Coverage Report Action

agentcore-devx-automation · 2026-05-20T17:41:43Z

Claude Security Review: the review run failed before completing. See the run for details.

agentcore-devx-automation · 2026-05-20T17:57:54Z

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

Hweinstock

implementation looks good, some questions on the testing, but we can revisit as a follow-up to bug bash today.

Adds a new top-level `agentcore feedback` command that lets users send feedback (and an optional screenshot) directly from the CLI: agentcore feedback "your message" [--screenshot path] [--json] agentcore feedback # multi-step wizard Both modes display the AWS Customer Agreement and require interactive y/N consent before submitting. Bare Enter defaults to No, and non-TTY stdin is refused so the consent decision is always explicit. The wizard walks through Message → optional Screenshot → Consent → Submit, with Esc stepping back and message text preserved across declines/retries. Screenshots are validated locally (PNG/JPG/JPEG, ≤100MB) before any network call, then uploaded via presigned URL with SHA-256 checksum headers. Submissions are tagged with cli-version, os, node-version, and cli-mode metadata to help triage. The CLI does not attach AWS account IDs, credentials, project names, or telemetry IDs. Includes README + docs/commands.md updates and a new docs/feedback.md covering syntax, consent text, and what not to include.

Three changes from automated review: 1. Telemetry: register `feedback` in command-run schema with `mode` + `has_screenshot` attrs, instrument the CLI handler with client.withCommandRun (CANCELLED on declined consent), and the TUI submit path with withCommandRunTelemetry. 2. Screenshot S3 key: parse the actual object key from the presigned URL path instead of fabricating one client-side. The fabricated key could drift from Aperture's bucket layout (UTC/local date, prefix changes) and silently produce form references to non-existent objects. 3. TUI input validation: validate message in setMessage and screenshot in setScreenshot, surface the error inline on the input phase via state.inputError. Previously bad input was only caught after the user walked through consent, forcing them to re-do consent on retry. Exposes validateFeedbackMessage and validateScreenshotPath from the operations module so the hook can reuse the same validators as the submission orchestrator.

agentcore-devx-automation · 2026-05-20T19:50:58Z

Claude Security Review: the review run failed before completing. See the run for details.

Three issues from bug bash: - 2.5: Esc/Ctrl+C in the wizard didn't terminate the process. The onExit callback unmounted the screen but Ink's stdin raw-mode listeners kept Node alive. Fix: await render()'s waitUntilExit() and call process.exit(0) once the wizard unmounts. Matches invoke/command.tsx's pattern. - 3.2: Pointing --screenshot at a directory ("/tmp") produced "Screenshot must be one of: .png, .jpg, .jpeg" because we extension- checked first. Fix: stat the path first; surface a directory-specific error before the extension check. - 3.6: Quoted paths starting with "~/" (e.g. ~/Desktop/foo.png) hit ENOENT because Node's fs APIs don't expand tildes — only the shell does, and quoting suppresses that. Fix: expand a leading "~" or "~/..." to os.homedir() before stat/read. Added unit tests for both new validator branches.

agentcore-devx-automation · 2026-05-20T20:06:02Z

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

…onal Per tester feedback: the explicit Y/N prompt felt like an extra step for something that's already optional. Drop the screenshot-prompt phase entirely — the wizard now goes Message → Screenshot → Consent. The Screenshot phase is the path picker itself, with allowEmpty=true on PathInput. Submit with a path to attach; Enter on empty input or Esc to skip. Help text reflects this: "Enter select / skip · Esc skip". Verified end-to-end via TUI harness: live submission id 96e2874e-3a61-4f65-8140-c144393d841e through the new flow.

agentcore-devx-automation · 2026-05-20T20:25:31Z

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

Hweinstock

LGTM! very minor nits and a question.

Per Hweinstock's review: - handleFeedback / useFeedbackFlow: replace the local `err instanceof Error ? err : new Error(String(err))` pattern with `toError(err)` from src/lib/errors/types.ts. - FeedbackValidationError: extend `ValidationError` from src/lib/errors/types.ts so the shared error classification helpers treat it as a user error, which means telemetry's classifyError(err)/isUserError(err) will set error_source: 'user' on failure metrics for invalid screenshots / empty messages.

agentcore-devx-automation · 2026-05-20T22:54:05Z

Claude Security Review: the review run failed before completing. See the run for details.

Hweinstock

Awesome!

…works (#1327) The /security-review slash command runs `git diff origin/HEAD...` as its first action to enumerate the PR's changes. actions/checkout doesn't set up the remote's symbolic HEAD ref, so that command fails with "ambiguous argument 'origin/HEAD...': unknown revision". Claude then loops trying variants until --max-turns 30 trips and the action exits 1. Set origin/HEAD to the PR's base ref right after checkout so the skill's git invocations resolve correctly. Run #26186045056 on PR #1321 was the trigger - 30 turns spent on shell-error recovery, no findings posted.

tejaskash · 2026-05-20T17:21:30Z

Summary

Adds a new top-level agentcore feedback command:

One-shot:agentcore feedback "msg" [--screenshot path] [--json]
Wizard:agentcore feedback (no args) → multi-step Ink TUI

Both modes display the AWS Customer Agreement and require interactive y/N consent before submitting. Bare Enter defaults to No; non-TTY stdin is refused so the consent decision is always explicit.

The wizard walks Message → optional Screenshot → Consent → Submit, with Esc stepping back and message text preserved across declines/retries. Screenshots are validated locally (PNG/JPG/JPEG, ≤100MB) before any network call, then uploaded via presigned URL with SHA-256 checksum headers.

Submissions carry cli-version, os, node-version, and cli-mode metadata. The CLI does not attach AWS account IDs, credentials, project names, or telemetry IDs.

What's in the diff

src/cli/commands/feedback/ — Commander registration, action, TTY consent prompt
src/cli/operations/feedback/ — submission orchestrator, payload builder, presigned-URL + S3 upload + form-ingestion client
src/cli/tui/screens/feedback/ — Ink wizard with useFeedbackFlow hook
src/cli/cli.ts, src/cli/tui/copy.ts — wiring
README.md, docs/commands.md, docs/feedback.md — customer-facing docs

29 unit tests across operations, command, consent-prompt, and TUI hook.

Test plan

npm test -- feedback — 29/29 pass
eslint clean on all new files
tsc --noEmit clean on all new files
Manual prod submission verified end-to-end (message + screenshot, both submit and decline paths)
CLI consent prompt: y/yes/Y/YES accepts; n/no/bare-Enter declines; non-TTY refuses
TUI wizard: all phase transitions, Esc back-stepping, message preservation across consent decline
Plain mode and --json mode for success, declined, no-tty, validation, network error
Bug bash plan attached for testers (5 testers × 6 min)

github-actions · 2026-05-20T17:22:57Z

Package Tarball

aws-agentcore-0.14.1.tgz

How to install

gh release download pr-1321-tarball --repo aws/agentcore-cli --pattern "*.tgz" --dir /tmp/pr-tarball
npm install -g /tmp/pr-tarball/aws-agentcore-0.14.1.tgz

agentcore-devx-automation · 2026-05-20T17:25:24Z

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

agentcore-cli-automation

Thanks for the careful work on consent flow, validation, and docs — the wizard structure and cli/tui mode tagging are nicely done. A few things I'd like to see addressed before merging:

Missing telemetry instrumentation. Per src/cli/telemetry/README.md, every CLI command should emit a command_run metric, but feedback doesn't. Other recent commands (e.g. deploy, remove, create) all use withCommandRunTelemetry / runCliCommand.
Screenshot S3 key is fabricated client-side independent of the presigned URL. The form references a key the client invented rather than the actual upload location.
TUI validates message and screenshot only after consent. Bad input forces the user to redo consent.

Inline comments below.

github-actions · 2026-05-20T17:25:51Z

Coverage Report

Status	Category	Percentage	Covered / Total
🔵	Lines	44.3%	9531 / 21510
🔵	Statements	43.55%	10114 / 23223
🔵	Functions	41.01%	1655 / 4035
🔵	Branches	40.84%	6162 / 15086

Generated in workflow #3182 for commit 91cb5da by the Vitest Coverage Report Action

agentcore-devx-automation · 2026-05-20T17:41:43Z

Claude Security Review: the review run failed before completing. See the run for details.

agentcore-devx-automation · 2026-05-20T17:57:54Z

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

Hweinstock

implementation looks good, some questions on the testing, but we can revisit as a follow-up to bug bash today.

Adds a new top-level `agentcore feedback` command that lets users send feedback (and an optional screenshot) directly from the CLI: agentcore feedback "your message" [--screenshot path] [--json] agentcore feedback # multi-step wizard Both modes display the AWS Customer Agreement and require interactive y/N consent before submitting. Bare Enter defaults to No, and non-TTY stdin is refused so the consent decision is always explicit. The wizard walks through Message → optional Screenshot → Consent → Submit, with Esc stepping back and message text preserved across declines/retries. Screenshots are validated locally (PNG/JPG/JPEG, ≤100MB) before any network call, then uploaded via presigned URL with SHA-256 checksum headers. Submissions are tagged with cli-version, os, node-version, and cli-mode metadata to help triage. The CLI does not attach AWS account IDs, credentials, project names, or telemetry IDs. Includes README + docs/commands.md updates and a new docs/feedback.md covering syntax, consent text, and what not to include.

Three changes from automated review: 1. Telemetry: register `feedback` in command-run schema with `mode` + `has_screenshot` attrs, instrument the CLI handler with client.withCommandRun (CANCELLED on declined consent), and the TUI submit path with withCommandRunTelemetry. 2. Screenshot S3 key: parse the actual object key from the presigned URL path instead of fabricating one client-side. The fabricated key could drift from Aperture's bucket layout (UTC/local date, prefix changes) and silently produce form references to non-existent objects. 3. TUI input validation: validate message in setMessage and screenshot in setScreenshot, surface the error inline on the input phase via state.inputError. Previously bad input was only caught after the user walked through consent, forcing them to re-do consent on retry. Exposes validateFeedbackMessage and validateScreenshotPath from the operations module so the hook can reuse the same validators as the submission orchestrator.

agentcore-devx-automation · 2026-05-20T19:50:58Z

Claude Security Review: the review run failed before completing. See the run for details.

Three issues from bug bash: - 2.5: Esc/Ctrl+C in the wizard didn't terminate the process. The onExit callback unmounted the screen but Ink's stdin raw-mode listeners kept Node alive. Fix: await render()'s waitUntilExit() and call process.exit(0) once the wizard unmounts. Matches invoke/command.tsx's pattern. - 3.2: Pointing --screenshot at a directory ("/tmp") produced "Screenshot must be one of: .png, .jpg, .jpeg" because we extension- checked first. Fix: stat the path first; surface a directory-specific error before the extension check. - 3.6: Quoted paths starting with "~/" (e.g. ~/Desktop/foo.png) hit ENOENT because Node's fs APIs don't expand tildes — only the shell does, and quoting suppresses that. Fix: expand a leading "~" or "~/..." to os.homedir() before stat/read. Added unit tests for both new validator branches.

agentcore-devx-automation · 2026-05-20T20:06:02Z

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

…onal Per tester feedback: the explicit Y/N prompt felt like an extra step for something that's already optional. Drop the screenshot-prompt phase entirely — the wizard now goes Message → Screenshot → Consent. The Screenshot phase is the path picker itself, with allowEmpty=true on PathInput. Submit with a path to attach; Enter on empty input or Esc to skip. Help text reflects this: "Enter select / skip · Esc skip". Verified end-to-end via TUI harness: live submission id 96e2874e-3a61-4f65-8140-c144393d841e through the new flow.

agentcore-devx-automation · 2026-05-20T20:25:31Z

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

Hweinstock

LGTM! very minor nits and a question.

Per Hweinstock's review: - handleFeedback / useFeedbackFlow: replace the local `err instanceof Error ? err : new Error(String(err))` pattern with `toError(err)` from src/lib/errors/types.ts. - FeedbackValidationError: extend `ValidationError` from src/lib/errors/types.ts so the shared error classification helpers treat it as a user error, which means telemetry's classifyError(err)/isUserError(err) will set error_source: 'user' on failure metrics for invalid screenshots / empty messages.

agentcore-devx-automation · 2026-05-20T22:54:05Z

Claude Security Review: the review run failed before completing. See the run for details.

Hweinstock

Awesome!

…works (#1327) The /security-review slash command runs `git diff origin/HEAD...` as its first action to enumerate the PR's changes. actions/checkout doesn't set up the remote's symbolic HEAD ref, so that command fails with "ambiguous argument 'origin/HEAD...': unknown revision". Claude then loops trying variants until --max-turns 30 trips and the action exits 1. Set origin/HEAD to the PR's base ref right after checkout so the skill's git invocations resolve correctly. Run #26186045056 on PR #1321 was the trigger - 30 turns spent on shell-error recovery, no findings posted.

tejaskash · 2026-05-20T17:21:30Z

Summary

Adds a new top-level agentcore feedback command:

One-shot:agentcore feedback "msg" [--screenshot path] [--json]
Wizard:agentcore feedback (no args) → multi-step Ink TUI

Both modes display the AWS Customer Agreement and require interactive y/N consent before submitting. Bare Enter defaults to No; non-TTY stdin is refused so the consent decision is always explicit.

The wizard walks Message → optional Screenshot → Consent → Submit, with Esc stepping back and message text preserved across declines/retries. Screenshots are validated locally (PNG/JPG/JPEG, ≤100MB) before any network call, then uploaded via presigned URL with SHA-256 checksum headers.

Submissions carry cli-version, os, node-version, and cli-mode metadata. The CLI does not attach AWS account IDs, credentials, project names, or telemetry IDs.

What's in the diff

src/cli/commands/feedback/ — Commander registration, action, TTY consent prompt
src/cli/operations/feedback/ — submission orchestrator, payload builder, presigned-URL + S3 upload + form-ingestion client
src/cli/tui/screens/feedback/ — Ink wizard with useFeedbackFlow hook
src/cli/cli.ts, src/cli/tui/copy.ts — wiring
README.md, docs/commands.md, docs/feedback.md — customer-facing docs

29 unit tests across operations, command, consent-prompt, and TUI hook.

Test plan

npm test -- feedback — 29/29 pass
eslint clean on all new files
tsc --noEmit clean on all new files
Manual prod submission verified end-to-end (message + screenshot, both submit and decline paths)
CLI consent prompt: y/yes/Y/YES accepts; n/no/bare-Enter declines; non-TTY refuses
TUI wizard: all phase transitions, Esc back-stepping, message preservation across consent decline
Plain mode and --json mode for success, declined, no-tty, validation, network error
Bug bash plan attached for testers (5 testers × 6 min)

github-actions · 2026-05-20T17:22:57Z

Package Tarball

aws-agentcore-0.14.1.tgz

How to install

gh release download pr-1321-tarball --repo aws/agentcore-cli --pattern "*.tgz" --dir /tmp/pr-tarball
npm install -g /tmp/pr-tarball/aws-agentcore-0.14.1.tgz

agentcore-devx-automation · 2026-05-20T17:25:24Z

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

agentcore-cli-automation

Thanks for the careful work on consent flow, validation, and docs — the wizard structure and cli/tui mode tagging are nicely done. A few things I'd like to see addressed before merging:

Missing telemetry instrumentation. Per src/cli/telemetry/README.md, every CLI command should emit a command_run metric, but feedback doesn't. Other recent commands (e.g. deploy, remove, create) all use withCommandRunTelemetry / runCliCommand.
Screenshot S3 key is fabricated client-side independent of the presigned URL. The form references a key the client invented rather than the actual upload location.
TUI validates message and screenshot only after consent. Bad input forces the user to redo consent.

Inline comments below.

github-actions · 2026-05-20T17:25:51Z

Coverage Report

Status	Category	Percentage	Covered / Total
🔵	Lines	44.3%	9531 / 21510
🔵	Statements	43.55%	10114 / 23223
🔵	Functions	41.01%	1655 / 4035
🔵	Branches	40.84%	6162 / 15086

Generated in workflow #3182 for commit 91cb5da by the Vitest Coverage Report Action

agentcore-devx-automation · 2026-05-20T17:41:43Z

Claude Security Review: the review run failed before completing. See the run for details.

agentcore-devx-automation · 2026-05-20T17:57:54Z

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

Hweinstock

implementation looks good, some questions on the testing, but we can revisit as a follow-up to bug bash today.

Adds a new top-level `agentcore feedback` command that lets users send feedback (and an optional screenshot) directly from the CLI: agentcore feedback "your message" [--screenshot path] [--json] agentcore feedback # multi-step wizard Both modes display the AWS Customer Agreement and require interactive y/N consent before submitting. Bare Enter defaults to No, and non-TTY stdin is refused so the consent decision is always explicit. The wizard walks through Message → optional Screenshot → Consent → Submit, with Esc stepping back and message text preserved across declines/retries. Screenshots are validated locally (PNG/JPG/JPEG, ≤100MB) before any network call, then uploaded via presigned URL with SHA-256 checksum headers. Submissions are tagged with cli-version, os, node-version, and cli-mode metadata to help triage. The CLI does not attach AWS account IDs, credentials, project names, or telemetry IDs. Includes README + docs/commands.md updates and a new docs/feedback.md covering syntax, consent text, and what not to include.

Three changes from automated review: 1. Telemetry: register `feedback` in command-run schema with `mode` + `has_screenshot` attrs, instrument the CLI handler with client.withCommandRun (CANCELLED on declined consent), and the TUI submit path with withCommandRunTelemetry. 2. Screenshot S3 key: parse the actual object key from the presigned URL path instead of fabricating one client-side. The fabricated key could drift from Aperture's bucket layout (UTC/local date, prefix changes) and silently produce form references to non-existent objects. 3. TUI input validation: validate message in setMessage and screenshot in setScreenshot, surface the error inline on the input phase via state.inputError. Previously bad input was only caught after the user walked through consent, forcing them to re-do consent on retry. Exposes validateFeedbackMessage and validateScreenshotPath from the operations module so the hook can reuse the same validators as the submission orchestrator.

agentcore-devx-automation · 2026-05-20T19:50:58Z

Claude Security Review: the review run failed before completing. See the run for details.

Three issues from bug bash: - 2.5: Esc/Ctrl+C in the wizard didn't terminate the process. The onExit callback unmounted the screen but Ink's stdin raw-mode listeners kept Node alive. Fix: await render()'s waitUntilExit() and call process.exit(0) once the wizard unmounts. Matches invoke/command.tsx's pattern. - 3.2: Pointing --screenshot at a directory ("/tmp") produced "Screenshot must be one of: .png, .jpg, .jpeg" because we extension- checked first. Fix: stat the path first; surface a directory-specific error before the extension check. - 3.6: Quoted paths starting with "~/" (e.g. ~/Desktop/foo.png) hit ENOENT because Node's fs APIs don't expand tildes — only the shell does, and quoting suppresses that. Fix: expand a leading "~" or "~/..." to os.homedir() before stat/read. Added unit tests for both new validator branches.

agentcore-devx-automation · 2026-05-20T20:06:02Z

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

…onal Per tester feedback: the explicit Y/N prompt felt like an extra step for something that's already optional. Drop the screenshot-prompt phase entirely — the wizard now goes Message → Screenshot → Consent. The Screenshot phase is the path picker itself, with allowEmpty=true on PathInput. Submit with a path to attach; Enter on empty input or Esc to skip. Help text reflects this: "Enter select / skip · Esc skip". Verified end-to-end via TUI harness: live submission id 96e2874e-3a61-4f65-8140-c144393d841e through the new flow.

agentcore-devx-automation · 2026-05-20T20:25:31Z

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

Hweinstock

LGTM! very minor nits and a question.

Per Hweinstock's review: - handleFeedback / useFeedbackFlow: replace the local `err instanceof Error ? err : new Error(String(err))` pattern with `toError(err)` from src/lib/errors/types.ts. - FeedbackValidationError: extend `ValidationError` from src/lib/errors/types.ts so the shared error classification helpers treat it as a user error, which means telemetry's classifyError(err)/isUserError(err) will set error_source: 'user' on failure metrics for invalid screenshots / empty messages.

agentcore-devx-automation · 2026-05-20T22:54:05Z

Claude Security Review: the review run failed before completing. See the run for details.

Hweinstock

Awesome!

…works (#1327) The /security-review slash command runs `git diff origin/HEAD...` as its first action to enumerate the PR's changes. actions/checkout doesn't set up the remote's symbolic HEAD ref, so that command fails with "ambiguous argument 'origin/HEAD...': unknown revision". Claude then loops trying variants until --max-turns 30 trips and the action exits 1. Set origin/HEAD to the PR's base ref right after checkout so the skill's git invocations resolve correctly. Run #26186045056 on PR #1321 was the trigger - 30 turns spent on shell-error recovery, no findings posted.

tejaskash · 2026-05-20T17:21:30Z

Summary

Adds a new top-level agentcore feedback command:

One-shot:agentcore feedback "msg" [--screenshot path] [--json]
Wizard:agentcore feedback (no args) → multi-step Ink TUI

Both modes display the AWS Customer Agreement and require interactive y/N consent before submitting. Bare Enter defaults to No; non-TTY stdin is refused so the consent decision is always explicit.

The wizard walks Message → optional Screenshot → Consent → Submit, with Esc stepping back and message text preserved across declines/retries. Screenshots are validated locally (PNG/JPG/JPEG, ≤100MB) before any network call, then uploaded via presigned URL with SHA-256 checksum headers.

Submissions carry cli-version, os, node-version, and cli-mode metadata. The CLI does not attach AWS account IDs, credentials, project names, or telemetry IDs.

What's in the diff

src/cli/commands/feedback/ — Commander registration, action, TTY consent prompt
src/cli/operations/feedback/ — submission orchestrator, payload builder, presigned-URL + S3 upload + form-ingestion client
src/cli/tui/screens/feedback/ — Ink wizard with useFeedbackFlow hook
src/cli/cli.ts, src/cli/tui/copy.ts — wiring
README.md, docs/commands.md, docs/feedback.md — customer-facing docs

29 unit tests across operations, command, consent-prompt, and TUI hook.

Test plan

npm test -- feedback — 29/29 pass
eslint clean on all new files
tsc --noEmit clean on all new files
Manual prod submission verified end-to-end (message + screenshot, both submit and decline paths)
CLI consent prompt: y/yes/Y/YES accepts; n/no/bare-Enter declines; non-TTY refuses
TUI wizard: all phase transitions, Esc back-stepping, message preservation across consent decline
Plain mode and --json mode for success, declined, no-tty, validation, network error
Bug bash plan attached for testers (5 testers × 6 min)

github-actions · 2026-05-20T17:22:57Z

Package Tarball

aws-agentcore-0.14.1.tgz

How to install

gh release download pr-1321-tarball --repo aws/agentcore-cli --pattern "*.tgz" --dir /tmp/pr-tarball
npm install -g /tmp/pr-tarball/aws-agentcore-0.14.1.tgz

agentcore-devx-automation · 2026-05-20T17:25:24Z

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

agentcore-cli-automation

Thanks for the careful work on consent flow, validation, and docs — the wizard structure and cli/tui mode tagging are nicely done. A few things I'd like to see addressed before merging:

Missing telemetry instrumentation. Per src/cli/telemetry/README.md, every CLI command should emit a command_run metric, but feedback doesn't. Other recent commands (e.g. deploy, remove, create) all use withCommandRunTelemetry / runCliCommand.
Screenshot S3 key is fabricated client-side independent of the presigned URL. The form references a key the client invented rather than the actual upload location.
TUI validates message and screenshot only after consent. Bad input forces the user to redo consent.

Inline comments below.

github-actions · 2026-05-20T17:25:51Z

Coverage Report

Status	Category	Percentage	Covered / Total
🔵	Lines	44.3%	9531 / 21510
🔵	Statements	43.55%	10114 / 23223
🔵	Functions	41.01%	1655 / 4035
🔵	Branches	40.84%	6162 / 15086

Generated in workflow #3182 for commit 91cb5da by the Vitest Coverage Report Action

agentcore-devx-automation · 2026-05-20T17:41:43Z

Claude Security Review: the review run failed before completing. See the run for details.

agentcore-devx-automation · 2026-05-20T17:57:54Z

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

Hweinstock

implementation looks good, some questions on the testing, but we can revisit as a follow-up to bug bash today.

Adds a new top-level `agentcore feedback` command that lets users send feedback (and an optional screenshot) directly from the CLI: agentcore feedback "your message" [--screenshot path] [--json] agentcore feedback # multi-step wizard Both modes display the AWS Customer Agreement and require interactive y/N consent before submitting. Bare Enter defaults to No, and non-TTY stdin is refused so the consent decision is always explicit. The wizard walks through Message → optional Screenshot → Consent → Submit, with Esc stepping back and message text preserved across declines/retries. Screenshots are validated locally (PNG/JPG/JPEG, ≤100MB) before any network call, then uploaded via presigned URL with SHA-256 checksum headers. Submissions are tagged with cli-version, os, node-version, and cli-mode metadata to help triage. The CLI does not attach AWS account IDs, credentials, project names, or telemetry IDs. Includes README + docs/commands.md updates and a new docs/feedback.md covering syntax, consent text, and what not to include.

Three changes from automated review: 1. Telemetry: register `feedback` in command-run schema with `mode` + `has_screenshot` attrs, instrument the CLI handler with client.withCommandRun (CANCELLED on declined consent), and the TUI submit path with withCommandRunTelemetry. 2. Screenshot S3 key: parse the actual object key from the presigned URL path instead of fabricating one client-side. The fabricated key could drift from Aperture's bucket layout (UTC/local date, prefix changes) and silently produce form references to non-existent objects. 3. TUI input validation: validate message in setMessage and screenshot in setScreenshot, surface the error inline on the input phase via state.inputError. Previously bad input was only caught after the user walked through consent, forcing them to re-do consent on retry. Exposes validateFeedbackMessage and validateScreenshotPath from the operations module so the hook can reuse the same validators as the submission orchestrator.

agentcore-devx-automation · 2026-05-20T19:50:58Z

Claude Security Review: the review run failed before completing. See the run for details.

Three issues from bug bash: - 2.5: Esc/Ctrl+C in the wizard didn't terminate the process. The onExit callback unmounted the screen but Ink's stdin raw-mode listeners kept Node alive. Fix: await render()'s waitUntilExit() and call process.exit(0) once the wizard unmounts. Matches invoke/command.tsx's pattern. - 3.2: Pointing --screenshot at a directory ("/tmp") produced "Screenshot must be one of: .png, .jpg, .jpeg" because we extension- checked first. Fix: stat the path first; surface a directory-specific error before the extension check. - 3.6: Quoted paths starting with "~/" (e.g. ~/Desktop/foo.png) hit ENOENT because Node's fs APIs don't expand tildes — only the shell does, and quoting suppresses that. Fix: expand a leading "~" or "~/..." to os.homedir() before stat/read. Added unit tests for both new validator branches.

agentcore-devx-automation · 2026-05-20T20:06:02Z

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

…onal Per tester feedback: the explicit Y/N prompt felt like an extra step for something that's already optional. Drop the screenshot-prompt phase entirely — the wizard now goes Message → Screenshot → Consent. The Screenshot phase is the path picker itself, with allowEmpty=true on PathInput. Submit with a path to attach; Enter on empty input or Esc to skip. Help text reflects this: "Enter select / skip · Esc skip". Verified end-to-end via TUI harness: live submission id 96e2874e-3a61-4f65-8140-c144393d841e through the new flow.

agentcore-devx-automation · 2026-05-20T20:25:31Z

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

Hweinstock

LGTM! very minor nits and a question.

Per Hweinstock's review: - handleFeedback / useFeedbackFlow: replace the local `err instanceof Error ? err : new Error(String(err))` pattern with `toError(err)` from src/lib/errors/types.ts. - FeedbackValidationError: extend `ValidationError` from src/lib/errors/types.ts so the shared error classification helpers treat it as a user error, which means telemetry's classifyError(err)/isUserError(err) will set error_source: 'user' on failure metrics for invalid screenshots / empty messages.

agentcore-devx-automation · 2026-05-20T22:54:05Z

Claude Security Review: the review run failed before completing. See the run for details.

Hweinstock

Awesome!

…works (#1327) The /security-review slash command runs `git diff origin/HEAD...` as its first action to enumerate the PR's changes. actions/checkout doesn't set up the remote's symbolic HEAD ref, so that command fails with "ambiguous argument 'origin/HEAD...': unknown revision". Claude then loops trying variants until --max-turns 30 trips and the action exits 1. Set origin/HEAD to the PR's base ref right after checkout so the skill's git invocations resolve correctly. Run #26186045056 on PR #1321 was the trigger - 30 turns spent on shell-error recovery, no findings posted.

tejaskash · 2026-05-20T17:21:30Z

Summary

Adds a new top-level agentcore feedback command:

One-shot:agentcore feedback "msg" [--screenshot path] [--json]
Wizard:agentcore feedback (no args) → multi-step Ink TUI

Both modes display the AWS Customer Agreement and require interactive y/N consent before submitting. Bare Enter defaults to No; non-TTY stdin is refused so the consent decision is always explicit.

The wizard walks Message → optional Screenshot → Consent → Submit, with Esc stepping back and message text preserved across declines/retries. Screenshots are validated locally (PNG/JPG/JPEG, ≤100MB) before any network call, then uploaded via presigned URL with SHA-256 checksum headers.

Submissions carry cli-version, os, node-version, and cli-mode metadata. The CLI does not attach AWS account IDs, credentials, project names, or telemetry IDs.

What's in the diff

src/cli/commands/feedback/ — Commander registration, action, TTY consent prompt
src/cli/operations/feedback/ — submission orchestrator, payload builder, presigned-URL + S3 upload + form-ingestion client
src/cli/tui/screens/feedback/ — Ink wizard with useFeedbackFlow hook
src/cli/cli.ts, src/cli/tui/copy.ts — wiring
README.md, docs/commands.md, docs/feedback.md — customer-facing docs

29 unit tests across operations, command, consent-prompt, and TUI hook.

Test plan

npm test -- feedback — 29/29 pass
eslint clean on all new files
tsc --noEmit clean on all new files
Manual prod submission verified end-to-end (message + screenshot, both submit and decline paths)
CLI consent prompt: y/yes/Y/YES accepts; n/no/bare-Enter declines; non-TTY refuses
TUI wizard: all phase transitions, Esc back-stepping, message preservation across consent decline
Plain mode and --json mode for success, declined, no-tty, validation, network error
Bug bash plan attached for testers (5 testers × 6 min)

github-actions · 2026-05-20T17:22:57Z

Package Tarball

aws-agentcore-0.14.1.tgz

How to install

gh release download pr-1321-tarball --repo aws/agentcore-cli --pattern "*.tgz" --dir /tmp/pr-tarball
npm install -g /tmp/pr-tarball/aws-agentcore-0.14.1.tgz

agentcore-devx-automation · 2026-05-20T17:25:24Z

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

agentcore-cli-automation

Thanks for the careful work on consent flow, validation, and docs — the wizard structure and cli/tui mode tagging are nicely done. A few things I'd like to see addressed before merging:

Missing telemetry instrumentation. Per src/cli/telemetry/README.md, every CLI command should emit a command_run metric, but feedback doesn't. Other recent commands (e.g. deploy, remove, create) all use withCommandRunTelemetry / runCliCommand.
Screenshot S3 key is fabricated client-side independent of the presigned URL. The form references a key the client invented rather than the actual upload location.
TUI validates message and screenshot only after consent. Bad input forces the user to redo consent.

Inline comments below.

github-actions · 2026-05-20T17:25:51Z

Coverage Report

Status	Category	Percentage	Covered / Total
🔵	Lines	44.3%	9531 / 21510
🔵	Statements	43.55%	10114 / 23223
🔵	Functions	41.01%	1655 / 4035
🔵	Branches	40.84%	6162 / 15086

Generated in workflow #3182 for commit 91cb5da by the Vitest Coverage Report Action

agentcore-devx-automation · 2026-05-20T17:41:43Z

Claude Security Review: the review run failed before completing. See the run for details.

agentcore-devx-automation · 2026-05-20T17:57:54Z

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

Hweinstock

implementation looks good, some questions on the testing, but we can revisit as a follow-up to bug bash today.

Adds a new top-level `agentcore feedback` command that lets users send feedback (and an optional screenshot) directly from the CLI: agentcore feedback "your message" [--screenshot path] [--json] agentcore feedback # multi-step wizard Both modes display the AWS Customer Agreement and require interactive y/N consent before submitting. Bare Enter defaults to No, and non-TTY stdin is refused so the consent decision is always explicit. The wizard walks through Message → optional Screenshot → Consent → Submit, with Esc stepping back and message text preserved across declines/retries. Screenshots are validated locally (PNG/JPG/JPEG, ≤100MB) before any network call, then uploaded via presigned URL with SHA-256 checksum headers. Submissions are tagged with cli-version, os, node-version, and cli-mode metadata to help triage. The CLI does not attach AWS account IDs, credentials, project names, or telemetry IDs. Includes README + docs/commands.md updates and a new docs/feedback.md covering syntax, consent text, and what not to include.

Three changes from automated review: 1. Telemetry: register `feedback` in command-run schema with `mode` + `has_screenshot` attrs, instrument the CLI handler with client.withCommandRun (CANCELLED on declined consent), and the TUI submit path with withCommandRunTelemetry. 2. Screenshot S3 key: parse the actual object key from the presigned URL path instead of fabricating one client-side. The fabricated key could drift from Aperture's bucket layout (UTC/local date, prefix changes) and silently produce form references to non-existent objects. 3. TUI input validation: validate message in setMessage and screenshot in setScreenshot, surface the error inline on the input phase via state.inputError. Previously bad input was only caught after the user walked through consent, forcing them to re-do consent on retry. Exposes validateFeedbackMessage and validateScreenshotPath from the operations module so the hook can reuse the same validators as the submission orchestrator.

agentcore-devx-automation · 2026-05-20T19:50:58Z

Claude Security Review: the review run failed before completing. See the run for details.

Three issues from bug bash: - 2.5: Esc/Ctrl+C in the wizard didn't terminate the process. The onExit callback unmounted the screen but Ink's stdin raw-mode listeners kept Node alive. Fix: await render()'s waitUntilExit() and call process.exit(0) once the wizard unmounts. Matches invoke/command.tsx's pattern. - 3.2: Pointing --screenshot at a directory ("/tmp") produced "Screenshot must be one of: .png, .jpg, .jpeg" because we extension- checked first. Fix: stat the path first; surface a directory-specific error before the extension check. - 3.6: Quoted paths starting with "~/" (e.g. ~/Desktop/foo.png) hit ENOENT because Node's fs APIs don't expand tildes — only the shell does, and quoting suppresses that. Fix: expand a leading "~" or "~/..." to os.homedir() before stat/read. Added unit tests for both new validator branches.

agentcore-devx-automation · 2026-05-20T20:06:02Z

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

…onal Per tester feedback: the explicit Y/N prompt felt like an extra step for something that's already optional. Drop the screenshot-prompt phase entirely — the wizard now goes Message → Screenshot → Consent. The Screenshot phase is the path picker itself, with allowEmpty=true on PathInput. Submit with a path to attach; Enter on empty input or Esc to skip. Help text reflects this: "Enter select / skip · Esc skip". Verified end-to-end via TUI harness: live submission id 96e2874e-3a61-4f65-8140-c144393d841e through the new flow.

agentcore-devx-automation · 2026-05-20T20:25:31Z

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

Hweinstock

LGTM! very minor nits and a question.

Per Hweinstock's review: - handleFeedback / useFeedbackFlow: replace the local `err instanceof Error ? err : new Error(String(err))` pattern with `toError(err)` from src/lib/errors/types.ts. - FeedbackValidationError: extend `ValidationError` from src/lib/errors/types.ts so the shared error classification helpers treat it as a user error, which means telemetry's classifyError(err)/isUserError(err) will set error_source: 'user' on failure metrics for invalid screenshots / empty messages.

agentcore-devx-automation · 2026-05-20T22:54:05Z

Claude Security Review: the review run failed before completing. See the run for details.

Hweinstock

Awesome!

…works (#1327) The /security-review slash command runs `git diff origin/HEAD...` as its first action to enumerate the PR's changes. actions/checkout doesn't set up the remote's symbolic HEAD ref, so that command fails with "ambiguous argument 'origin/HEAD...': unknown revision". Claude then loops trying variants until --max-turns 30 trips and the action exits 1. Set origin/HEAD to the PR's base ref right after checkout so the skill's git invocations resolve correctly. Run #26186045056 on PR #1321 was the trigger - 30 turns spent on shell-error recovery, no findings posted.

tejaskash · 2026-05-20T17:21:30Z

Summary

Adds a new top-level agentcore feedback command:

One-shot:agentcore feedback "msg" [--screenshot path] [--json]
Wizard:agentcore feedback (no args) → multi-step Ink TUI

Both modes display the AWS Customer Agreement and require interactive y/N consent before submitting. Bare Enter defaults to No; non-TTY stdin is refused so the consent decision is always explicit.

The wizard walks Message → optional Screenshot → Consent → Submit, with Esc stepping back and message text preserved across declines/retries. Screenshots are validated locally (PNG/JPG/JPEG, ≤100MB) before any network call, then uploaded via presigned URL with SHA-256 checksum headers.

Submissions carry cli-version, os, node-version, and cli-mode metadata. The CLI does not attach AWS account IDs, credentials, project names, or telemetry IDs.

What's in the diff

src/cli/commands/feedback/ — Commander registration, action, TTY consent prompt
src/cli/operations/feedback/ — submission orchestrator, payload builder, presigned-URL + S3 upload + form-ingestion client
src/cli/tui/screens/feedback/ — Ink wizard with useFeedbackFlow hook
src/cli/cli.ts, src/cli/tui/copy.ts — wiring
README.md, docs/commands.md, docs/feedback.md — customer-facing docs

29 unit tests across operations, command, consent-prompt, and TUI hook.

Test plan

npm test -- feedback — 29/29 pass
eslint clean on all new files
tsc --noEmit clean on all new files
Manual prod submission verified end-to-end (message + screenshot, both submit and decline paths)
CLI consent prompt: y/yes/Y/YES accepts; n/no/bare-Enter declines; non-TTY refuses
TUI wizard: all phase transitions, Esc back-stepping, message preservation across consent decline
Plain mode and --json mode for success, declined, no-tty, validation, network error
Bug bash plan attached for testers (5 testers × 6 min)

github-actions · 2026-05-20T17:22:57Z

Package Tarball

aws-agentcore-0.14.1.tgz

How to install

gh release download pr-1321-tarball --repo aws/agentcore-cli --pattern "*.tgz" --dir /tmp/pr-tarball
npm install -g /tmp/pr-tarball/aws-agentcore-0.14.1.tgz

agentcore-devx-automation · 2026-05-20T17:25:24Z

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

agentcore-cli-automation

Thanks for the careful work on consent flow, validation, and docs — the wizard structure and cli/tui mode tagging are nicely done. A few things I'd like to see addressed before merging:

Missing telemetry instrumentation. Per src/cli/telemetry/README.md, every CLI command should emit a command_run metric, but feedback doesn't. Other recent commands (e.g. deploy, remove, create) all use withCommandRunTelemetry / runCliCommand.
Screenshot S3 key is fabricated client-side independent of the presigned URL. The form references a key the client invented rather than the actual upload location.
TUI validates message and screenshot only after consent. Bad input forces the user to redo consent.

Inline comments below.

github-actions · 2026-05-20T17:25:51Z

Coverage Report

Status	Category	Percentage	Covered / Total
🔵	Lines	44.3%	9531 / 21510
🔵	Statements	43.55%	10114 / 23223
🔵	Functions	41.01%	1655 / 4035
🔵	Branches	40.84%	6162 / 15086

Generated in workflow #3182 for commit 91cb5da by the Vitest Coverage Report Action

agentcore-devx-automation · 2026-05-20T17:41:43Z

Claude Security Review: the review run failed before completing. See the run for details.

agentcore-devx-automation · 2026-05-20T17:57:54Z

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

Hweinstock

implementation looks good, some questions on the testing, but we can revisit as a follow-up to bug bash today.

Adds a new top-level `agentcore feedback` command that lets users send feedback (and an optional screenshot) directly from the CLI: agentcore feedback "your message" [--screenshot path] [--json] agentcore feedback # multi-step wizard Both modes display the AWS Customer Agreement and require interactive y/N consent before submitting. Bare Enter defaults to No, and non-TTY stdin is refused so the consent decision is always explicit. The wizard walks through Message → optional Screenshot → Consent → Submit, with Esc stepping back and message text preserved across declines/retries. Screenshots are validated locally (PNG/JPG/JPEG, ≤100MB) before any network call, then uploaded via presigned URL with SHA-256 checksum headers. Submissions are tagged with cli-version, os, node-version, and cli-mode metadata to help triage. The CLI does not attach AWS account IDs, credentials, project names, or telemetry IDs. Includes README + docs/commands.md updates and a new docs/feedback.md covering syntax, consent text, and what not to include.

Three changes from automated review: 1. Telemetry: register `feedback` in command-run schema with `mode` + `has_screenshot` attrs, instrument the CLI handler with client.withCommandRun (CANCELLED on declined consent), and the TUI submit path with withCommandRunTelemetry. 2. Screenshot S3 key: parse the actual object key from the presigned URL path instead of fabricating one client-side. The fabricated key could drift from Aperture's bucket layout (UTC/local date, prefix changes) and silently produce form references to non-existent objects. 3. TUI input validation: validate message in setMessage and screenshot in setScreenshot, surface the error inline on the input phase via state.inputError. Previously bad input was only caught after the user walked through consent, forcing them to re-do consent on retry. Exposes validateFeedbackMessage and validateScreenshotPath from the operations module so the hook can reuse the same validators as the submission orchestrator.

agentcore-devx-automation · 2026-05-20T19:50:58Z

Claude Security Review: the review run failed before completing. See the run for details.

Three issues from bug bash: - 2.5: Esc/Ctrl+C in the wizard didn't terminate the process. The onExit callback unmounted the screen but Ink's stdin raw-mode listeners kept Node alive. Fix: await render()'s waitUntilExit() and call process.exit(0) once the wizard unmounts. Matches invoke/command.tsx's pattern. - 3.2: Pointing --screenshot at a directory ("/tmp") produced "Screenshot must be one of: .png, .jpg, .jpeg" because we extension- checked first. Fix: stat the path first; surface a directory-specific error before the extension check. - 3.6: Quoted paths starting with "~/" (e.g. ~/Desktop/foo.png) hit ENOENT because Node's fs APIs don't expand tildes — only the shell does, and quoting suppresses that. Fix: expand a leading "~" or "~/..." to os.homedir() before stat/read. Added unit tests for both new validator branches.

agentcore-devx-automation · 2026-05-20T20:06:02Z

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

…onal Per tester feedback: the explicit Y/N prompt felt like an extra step for something that's already optional. Drop the screenshot-prompt phase entirely — the wizard now goes Message → Screenshot → Consent. The Screenshot phase is the path picker itself, with allowEmpty=true on PathInput. Submit with a path to attach; Enter on empty input or Esc to skip. Help text reflects this: "Enter select / skip · Esc skip". Verified end-to-end via TUI harness: live submission id 96e2874e-3a61-4f65-8140-c144393d841e through the new flow.

agentcore-devx-automation · 2026-05-20T20:25:31Z

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

Hweinstock

LGTM! very minor nits and a question.

Per Hweinstock's review: - handleFeedback / useFeedbackFlow: replace the local `err instanceof Error ? err : new Error(String(err))` pattern with `toError(err)` from src/lib/errors/types.ts. - FeedbackValidationError: extend `ValidationError` from src/lib/errors/types.ts so the shared error classification helpers treat it as a user error, which means telemetry's classifyError(err)/isUserError(err) will set error_source: 'user' on failure metrics for invalid screenshots / empty messages.

agentcore-devx-automation · 2026-05-20T22:54:05Z

Claude Security Review: the review run failed before completing. See the run for details.

Hweinstock

Awesome!

…works (#1327) The /security-review slash command runs `git diff origin/HEAD...` as its first action to enumerate the PR's changes. actions/checkout doesn't set up the remote's symbolic HEAD ref, so that command fails with "ambiguous argument 'origin/HEAD...': unknown revision". Claude then loops trying variants until --max-turns 30 trips and the action exits 1. Set origin/HEAD to the PR's base ref right after checkout so the skill's git invocations resolve correctly. Run #26186045056 on PR #1321 was the trigger - 30 turns spent on shell-error recovery, no findings posted.

tejaskash · 2026-05-20T17:21:30Z

Summary

Adds a new top-level agentcore feedback command:

One-shot:agentcore feedback "msg" [--screenshot path] [--json]
Wizard:agentcore feedback (no args) → multi-step Ink TUI

Both modes display the AWS Customer Agreement and require interactive y/N consent before submitting. Bare Enter defaults to No; non-TTY stdin is refused so the consent decision is always explicit.

The wizard walks Message → optional Screenshot → Consent → Submit, with Esc stepping back and message text preserved across declines/retries. Screenshots are validated locally (PNG/JPG/JPEG, ≤100MB) before any network call, then uploaded via presigned URL with SHA-256 checksum headers.

Submissions carry cli-version, os, node-version, and cli-mode metadata. The CLI does not attach AWS account IDs, credentials, project names, or telemetry IDs.

What's in the diff

src/cli/commands/feedback/ — Commander registration, action, TTY consent prompt
src/cli/operations/feedback/ — submission orchestrator, payload builder, presigned-URL + S3 upload + form-ingestion client
src/cli/tui/screens/feedback/ — Ink wizard with useFeedbackFlow hook
src/cli/cli.ts, src/cli/tui/copy.ts — wiring
README.md, docs/commands.md, docs/feedback.md — customer-facing docs

29 unit tests across operations, command, consent-prompt, and TUI hook.

Test plan

npm test -- feedback — 29/29 pass
eslint clean on all new files
tsc --noEmit clean on all new files
Manual prod submission verified end-to-end (message + screenshot, both submit and decline paths)
CLI consent prompt: y/yes/Y/YES accepts; n/no/bare-Enter declines; non-TTY refuses
TUI wizard: all phase transitions, Esc back-stepping, message preservation across consent decline
Plain mode and --json mode for success, declined, no-tty, validation, network error
Bug bash plan attached for testers (5 testers × 6 min)

github-actions · 2026-05-20T17:22:57Z

Package Tarball

aws-agentcore-0.14.1.tgz

How to install

gh release download pr-1321-tarball --repo aws/agentcore-cli --pattern "*.tgz" --dir /tmp/pr-tarball
npm install -g /tmp/pr-tarball/aws-agentcore-0.14.1.tgz

agentcore-devx-automation · 2026-05-20T17:25:24Z

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

agentcore-cli-automation

Thanks for the careful work on consent flow, validation, and docs — the wizard structure and cli/tui mode tagging are nicely done. A few things I'd like to see addressed before merging:

Missing telemetry instrumentation. Per src/cli/telemetry/README.md, every CLI command should emit a command_run metric, but feedback doesn't. Other recent commands (e.g. deploy, remove, create) all use withCommandRunTelemetry / runCliCommand.
Screenshot S3 key is fabricated client-side independent of the presigned URL. The form references a key the client invented rather than the actual upload location.
TUI validates message and screenshot only after consent. Bad input forces the user to redo consent.

Inline comments below.

github-actions · 2026-05-20T17:25:51Z

Coverage Report

Status	Category	Percentage	Covered / Total
🔵	Lines	44.3%	9531 / 21510
🔵	Statements	43.55%	10114 / 23223
🔵	Functions	41.01%	1655 / 4035
🔵	Branches	40.84%	6162 / 15086

Generated in workflow #3182 for commit 91cb5da by the Vitest Coverage Report Action

agentcore-devx-automation · 2026-05-20T17:41:43Z

Claude Security Review: the review run failed before completing. See the run for details.

agentcore-devx-automation · 2026-05-20T17:57:54Z

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

Hweinstock

implementation looks good, some questions on the testing, but we can revisit as a follow-up to bug bash today.

Adds a new top-level `agentcore feedback` command that lets users send feedback (and an optional screenshot) directly from the CLI: agentcore feedback "your message" [--screenshot path] [--json] agentcore feedback # multi-step wizard Both modes display the AWS Customer Agreement and require interactive y/N consent before submitting. Bare Enter defaults to No, and non-TTY stdin is refused so the consent decision is always explicit. The wizard walks through Message → optional Screenshot → Consent → Submit, with Esc stepping back and message text preserved across declines/retries. Screenshots are validated locally (PNG/JPG/JPEG, ≤100MB) before any network call, then uploaded via presigned URL with SHA-256 checksum headers. Submissions are tagged with cli-version, os, node-version, and cli-mode metadata to help triage. The CLI does not attach AWS account IDs, credentials, project names, or telemetry IDs. Includes README + docs/commands.md updates and a new docs/feedback.md covering syntax, consent text, and what not to include.

Three changes from automated review: 1. Telemetry: register `feedback` in command-run schema with `mode` + `has_screenshot` attrs, instrument the CLI handler with client.withCommandRun (CANCELLED on declined consent), and the TUI submit path with withCommandRunTelemetry. 2. Screenshot S3 key: parse the actual object key from the presigned URL path instead of fabricating one client-side. The fabricated key could drift from Aperture's bucket layout (UTC/local date, prefix changes) and silently produce form references to non-existent objects. 3. TUI input validation: validate message in setMessage and screenshot in setScreenshot, surface the error inline on the input phase via state.inputError. Previously bad input was only caught after the user walked through consent, forcing them to re-do consent on retry. Exposes validateFeedbackMessage and validateScreenshotPath from the operations module so the hook can reuse the same validators as the submission orchestrator.

agentcore-devx-automation · 2026-05-20T19:50:58Z

Claude Security Review: the review run failed before completing. See the run for details.

Three issues from bug bash: - 2.5: Esc/Ctrl+C in the wizard didn't terminate the process. The onExit callback unmounted the screen but Ink's stdin raw-mode listeners kept Node alive. Fix: await render()'s waitUntilExit() and call process.exit(0) once the wizard unmounts. Matches invoke/command.tsx's pattern. - 3.2: Pointing --screenshot at a directory ("/tmp") produced "Screenshot must be one of: .png, .jpg, .jpeg" because we extension- checked first. Fix: stat the path first; surface a directory-specific error before the extension check. - 3.6: Quoted paths starting with "~/" (e.g. ~/Desktop/foo.png) hit ENOENT because Node's fs APIs don't expand tildes — only the shell does, and quoting suppresses that. Fix: expand a leading "~" or "~/..." to os.homedir() before stat/read. Added unit tests for both new validator branches.

agentcore-devx-automation · 2026-05-20T20:06:02Z

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

…onal Per tester feedback: the explicit Y/N prompt felt like an extra step for something that's already optional. Drop the screenshot-prompt phase entirely — the wizard now goes Message → Screenshot → Consent. The Screenshot phase is the path picker itself, with allowEmpty=true on PathInput. Submit with a path to attach; Enter on empty input or Esc to skip. Help text reflects this: "Enter select / skip · Esc skip". Verified end-to-end via TUI harness: live submission id 96e2874e-3a61-4f65-8140-c144393d841e through the new flow.

agentcore-devx-automation · 2026-05-20T20:25:31Z

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

Hweinstock

LGTM! very minor nits and a question.

Per Hweinstock's review: - handleFeedback / useFeedbackFlow: replace the local `err instanceof Error ? err : new Error(String(err))` pattern with `toError(err)` from src/lib/errors/types.ts. - FeedbackValidationError: extend `ValidationError` from src/lib/errors/types.ts so the shared error classification helpers treat it as a user error, which means telemetry's classifyError(err)/isUserError(err) will set error_source: 'user' on failure metrics for invalid screenshots / empty messages.

agentcore-devx-automation · 2026-05-20T22:54:05Z

Claude Security Review: the review run failed before completing. See the run for details.

Hweinstock

Awesome!

…works (#1327) The /security-review slash command runs `git diff origin/HEAD...` as its first action to enumerate the PR's changes. actions/checkout doesn't set up the remote's symbolic HEAD ref, so that command fails with "ambiguous argument 'origin/HEAD...': unknown revision". Claude then loops trying variants until --max-turns 30 trips and the action exits 1. Set origin/HEAD to the PR's base ref right after checkout so the skill's git invocations resolve correctly. Run #26186045056 on PR #1321 was the trigger - 30 turns spent on shell-error recovery, no findings posted.

tejaskash · 2026-05-20T17:21:30Z

Summary

Adds a new top-level agentcore feedback command:

One-shot:agentcore feedback "msg" [--screenshot path] [--json]
Wizard:agentcore feedback (no args) → multi-step Ink TUI

Both modes display the AWS Customer Agreement and require interactive y/N consent before submitting. Bare Enter defaults to No; non-TTY stdin is refused so the consent decision is always explicit.

The wizard walks Message → optional Screenshot → Consent → Submit, with Esc stepping back and message text preserved across declines/retries. Screenshots are validated locally (PNG/JPG/JPEG, ≤100MB) before any network call, then uploaded via presigned URL with SHA-256 checksum headers.

Submissions carry cli-version, os, node-version, and cli-mode metadata. The CLI does not attach AWS account IDs, credentials, project names, or telemetry IDs.

What's in the diff

src/cli/commands/feedback/ — Commander registration, action, TTY consent prompt
src/cli/operations/feedback/ — submission orchestrator, payload builder, presigned-URL + S3 upload + form-ingestion client
src/cli/tui/screens/feedback/ — Ink wizard with useFeedbackFlow hook
src/cli/cli.ts, src/cli/tui/copy.ts — wiring
README.md, docs/commands.md, docs/feedback.md — customer-facing docs

29 unit tests across operations, command, consent-prompt, and TUI hook.

Test plan

npm test -- feedback — 29/29 pass
eslint clean on all new files
tsc --noEmit clean on all new files
Manual prod submission verified end-to-end (message + screenshot, both submit and decline paths)
CLI consent prompt: y/yes/Y/YES accepts; n/no/bare-Enter declines; non-TTY refuses
TUI wizard: all phase transitions, Esc back-stepping, message preservation across consent decline
Plain mode and --json mode for success, declined, no-tty, validation, network error
Bug bash plan attached for testers (5 testers × 6 min)

github-actions · 2026-05-20T17:22:57Z

Package Tarball

aws-agentcore-0.14.1.tgz

How to install

gh release download pr-1321-tarball --repo aws/agentcore-cli --pattern "*.tgz" --dir /tmp/pr-tarball
npm install -g /tmp/pr-tarball/aws-agentcore-0.14.1.tgz

agentcore-devx-automation · 2026-05-20T17:25:24Z

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

agentcore-cli-automation

Thanks for the careful work on consent flow, validation, and docs — the wizard structure and cli/tui mode tagging are nicely done. A few things I'd like to see addressed before merging:

Missing telemetry instrumentation. Per src/cli/telemetry/README.md, every CLI command should emit a command_run metric, but feedback doesn't. Other recent commands (e.g. deploy, remove, create) all use withCommandRunTelemetry / runCliCommand.
Screenshot S3 key is fabricated client-side independent of the presigned URL. The form references a key the client invented rather than the actual upload location.
TUI validates message and screenshot only after consent. Bad input forces the user to redo consent.

Inline comments below.

github-actions · 2026-05-20T17:25:51Z

Coverage Report

Status	Category	Percentage	Covered / Total
🔵	Lines	44.3%	9531 / 21510
🔵	Statements	43.55%	10114 / 23223
🔵	Functions	41.01%	1655 / 4035
🔵	Branches	40.84%	6162 / 15086

Generated in workflow #3182 for commit 91cb5da by the Vitest Coverage Report Action

agentcore-devx-automation · 2026-05-20T17:41:43Z

Claude Security Review: the review run failed before completing. See the run for details.

agentcore-devx-automation · 2026-05-20T17:57:54Z

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

Hweinstock

implementation looks good, some questions on the testing, but we can revisit as a follow-up to bug bash today.

Adds a new top-level `agentcore feedback` command that lets users send feedback (and an optional screenshot) directly from the CLI: agentcore feedback "your message" [--screenshot path] [--json] agentcore feedback # multi-step wizard Both modes display the AWS Customer Agreement and require interactive y/N consent before submitting. Bare Enter defaults to No, and non-TTY stdin is refused so the consent decision is always explicit. The wizard walks through Message → optional Screenshot → Consent → Submit, with Esc stepping back and message text preserved across declines/retries. Screenshots are validated locally (PNG/JPG/JPEG, ≤100MB) before any network call, then uploaded via presigned URL with SHA-256 checksum headers. Submissions are tagged with cli-version, os, node-version, and cli-mode metadata to help triage. The CLI does not attach AWS account IDs, credentials, project names, or telemetry IDs. Includes README + docs/commands.md updates and a new docs/feedback.md covering syntax, consent text, and what not to include.

Three changes from automated review: 1. Telemetry: register `feedback` in command-run schema with `mode` + `has_screenshot` attrs, instrument the CLI handler with client.withCommandRun (CANCELLED on declined consent), and the TUI submit path with withCommandRunTelemetry. 2. Screenshot S3 key: parse the actual object key from the presigned URL path instead of fabricating one client-side. The fabricated key could drift from Aperture's bucket layout (UTC/local date, prefix changes) and silently produce form references to non-existent objects. 3. TUI input validation: validate message in setMessage and screenshot in setScreenshot, surface the error inline on the input phase via state.inputError. Previously bad input was only caught after the user walked through consent, forcing them to re-do consent on retry. Exposes validateFeedbackMessage and validateScreenshotPath from the operations module so the hook can reuse the same validators as the submission orchestrator.

agentcore-devx-automation · 2026-05-20T19:50:58Z

Claude Security Review: the review run failed before completing. See the run for details.

Three issues from bug bash: - 2.5: Esc/Ctrl+C in the wizard didn't terminate the process. The onExit callback unmounted the screen but Ink's stdin raw-mode listeners kept Node alive. Fix: await render()'s waitUntilExit() and call process.exit(0) once the wizard unmounts. Matches invoke/command.tsx's pattern. - 3.2: Pointing --screenshot at a directory ("/tmp") produced "Screenshot must be one of: .png, .jpg, .jpeg" because we extension- checked first. Fix: stat the path first; surface a directory-specific error before the extension check. - 3.6: Quoted paths starting with "~/" (e.g. ~/Desktop/foo.png) hit ENOENT because Node's fs APIs don't expand tildes — only the shell does, and quoting suppresses that. Fix: expand a leading "~" or "~/..." to os.homedir() before stat/read. Added unit tests for both new validator branches.

agentcore-devx-automation · 2026-05-20T20:06:02Z

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

…onal Per tester feedback: the explicit Y/N prompt felt like an extra step for something that's already optional. Drop the screenshot-prompt phase entirely — the wizard now goes Message → Screenshot → Consent. The Screenshot phase is the path picker itself, with allowEmpty=true on PathInput. Submit with a path to attach; Enter on empty input or Esc to skip. Help text reflects this: "Enter select / skip · Esc skip". Verified end-to-end via TUI harness: live submission id 96e2874e-3a61-4f65-8140-c144393d841e through the new flow.

agentcore-devx-automation · 2026-05-20T20:25:31Z

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

Hweinstock

LGTM! very minor nits and a question.

Per Hweinstock's review: - handleFeedback / useFeedbackFlow: replace the local `err instanceof Error ? err : new Error(String(err))` pattern with `toError(err)` from src/lib/errors/types.ts. - FeedbackValidationError: extend `ValidationError` from src/lib/errors/types.ts so the shared error classification helpers treat it as a user error, which means telemetry's classifyError(err)/isUserError(err) will set error_source: 'user' on failure metrics for invalid screenshots / empty messages.

agentcore-devx-automation · 2026-05-20T22:54:05Z

Claude Security Review: the review run failed before completing. See the run for details.

Hweinstock

Awesome!

…works (#1327) The /security-review slash command runs `git diff origin/HEAD...` as its first action to enumerate the PR's changes. actions/checkout doesn't set up the remote's symbolic HEAD ref, so that command fails with "ambiguous argument 'origin/HEAD...': unknown revision". Claude then loops trying variants until --max-turns 30 trips and the action exits 1. Set origin/HEAD to the PR's base ref right after checkout so the skill's git invocations resolve correctly. Run #26186045056 on PR #1321 was the trigger - 30 turns spent on shell-error recovery, no findings posted.

tejaskash temporarily deployed to e2e-testing May 20, 2026 17:21 — with GitHub Actions Inactive

github-actionsBot added agentcore-harness-reviewing AgentCore Harness review in progress size/xl PR size: XL labels May 20, 2026

agentcore-devx-automationBot added the claude-security-reviewing Claude Code /security-review in progress label May 20, 2026

tejaskash marked this pull request as ready for review May 20, 2026 17:23

tejaskash requested a review from a teamMay 20, 2026 17:23

agentcore-cli-automation suggested changes May 20, 2026
View reviewed changes

Comment threadsrc/cli/commands/feedback/command.tsx
Comment threadsrc/cli/operations/feedback/submit-feedback.ts Outdated
Comment threadsrc/cli/tui/screens/feedback/useFeedbackFlow.ts Outdated

agentcore-devx-automationBot removed the claude-security-reviewing Claude Code /security-review in progress label May 20, 2026

github-actionsBot removed the agentcore-harness-reviewing AgentCore Harness review in progress label May 20, 2026

github-actionsBot added size/xl PR size: XL and removed size/xl PR size: XL labels May 20, 2026

tejaskash temporarily deployed to e2e-testing May 20, 2026 17:39 — with GitHub Actions Inactive

agentcore-devx-automationBot added the claude-security-reviewing Claude Code /security-review in progress label May 20, 2026

agentcore-devx-automationBot removed the claude-security-reviewing Claude Code /security-review in progress label May 20, 2026

tejaskash force-pushed the feature/feedback-command branch from ab65255 to 94f3a55Compare May 20, 2026 17:55

github-actionsBot added size/xl PR size: XL and removed size/xl PR size: XL labels May 20, 2026

tejaskash temporarily deployed to e2e-testing May 20, 2026 17:55 — with GitHub Actions Inactive

agentcore-devx-automationBot added the claude-security-reviewing Claude Code /security-review in progress label May 20, 2026

agentcore-devx-automationBot removed the claude-security-reviewing Claude Code /security-review in progress label May 20, 2026

Hweinstock previously approved these changes May 20, 2026
View reviewed changes

tejaskash added 2 commits May 20, 2026 14:25

tejaskash dismissed Hweinstock’s stale review via ffbb0feMay 20, 2026 18:25

tejaskash force-pushed the feature/feedback-command branch from 94f3a55 to ffbb0feCompare May 20, 2026 18:25

agentcore-devx-automationBot removed the claude-security-reviewing Claude Code /security-review in progress label May 20, 2026

tejaskash mentioned this pull request May 20, 2026
ci(security-review): fix origin/HEAD missing so /security-review's git diff works #1327
Merged
2 tasks

github-actionsBot added size/xl PR size: XL and removed size/xl PR size: XL labels May 20, 2026

tejaskash temporarily deployed to e2e-testing May 20, 2026 20:02 — with GitHub Actions Inactive

agentcore-devx-automationBot added the claude-security-reviewing Claude Code /security-review in progress label May 20, 2026

agentcore-devx-automationBot removed the claude-security-reviewing Claude Code /security-review in progress label May 20, 2026

github-actionsBot removed the size/xl PR size: XL label May 20, 2026

tejaskash temporarily deployed to e2e-testing May 20, 2026 20:19 — with GitHub Actions Inactive

github-actionsBot added the size/xl PR size: XL label May 20, 2026

agentcore-devx-automationBot added the claude-security-reviewing Claude Code /security-review in progress label May 20, 2026

agentcore-devx-automationBot removed the claude-security-reviewing Claude Code /security-review in progress label May 20, 2026

Hweinstock previously approved these changes May 20, 2026
View reviewed changes

tejaskash dismissed Hweinstock’s stale review via 91cb5daMay 20, 2026 22:51

github-actionsBot added size/xl PR size: XL and removed size/xl PR size: XL labels May 20, 2026

tejaskash temporarily deployed to e2e-testing May 20, 2026 22:51 — with GitHub Actions Inactive

agentcore-devx-automationBot added the claude-security-reviewing Claude Code /security-review in progress label May 20, 2026

agentcore-devx-automationBot removed the claude-security-reviewing Claude Code /security-review in progress label May 20, 2026

Hweinstock approved these changes May 21, 2026
View reviewed changes

tejaskash merged commit 32c475d into mainMay 21, 2026
32 of 33 checks passed

tejaskash deleted the feature/feedback-command branch May 21, 2026 13:35