You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Four agentic CLI/MCP versions updated in pkg/constants/version_constants.go. make build && make recompile (x2) succeeded: 260/260 workflows compiled, 0 errors, 219 warnings.
Tool
Old
New
Risk
Claude Code
2.1.214
2.1.216
Low
Copilot CLI
1.0.71
1.0.73
Medium
OpenAI Codex
0.144.5
0.144.6
Low
MCP Gateway (gh-aw-mcpg)
v0.4.1
v0.4.3
Low
No change: Copilot SDK 1.0.7 · Pi 0.80.10 · Playwright MCP 0.0.78 · Playwright CLI 0.1.17 · GitHub MCP Server v1.6.0 · Playwright Browser v1.61.1.
Claude Code 2.1.214 → 2.1.216
Published: 2.1.215 @ 2026-07-19, 2.1.216 @ 2026-07-20. No public repo; NPM + --help + changelog analysis.
Features
New CLI flag --forward-subagent-text: forwards subagent text/thinking blocks as assistant/user messages with parent_tool_use_id set (only with --print + --output-format=stream-json).
New sandbox.filesystem.disabled setting: skips filesystem isolation while keeping network egress control.
Fixes / Behavior
2.1.215: Claude no longer auto-runs /verify and /code-review skills — must be invoked explicitly (minor behavior change).
Perf: fixed quadratic message-normalization slowdown in long sessions.
Auto mode: fixed HTTP 401 classifier denials after mid-session token rotation.
auto-mode subcommand help: "Inspect" → "Inspect or reset auto mode classifier".
Impact: Low. Additive flag/setting; auto-skill behavior change does not affect gh-aw runs.
Help diff 2.1.215 → 2.1.216
+ --forward-subagent-text Forward subagent text and thinking blocks as
assistant/user messages with parent_tool_use_id set
(only works with --print and --output-format=stream-json)
- auto-mode Inspect auto mode classifier
+ auto-mode Inspect or reset auto mode classifier
Copilot CLI 1.0.71 → 1.0.73
Published: 1.0.72 @ 2026-07-20, 1.0.73 @ 2026-07-21. Prerelease 1.0.72-1 skipped. The bulk of changes landed in the large 1.0.72 release.
Features (1.0.72)
$ at prompt opens a shell in the session dir (opt-in via /settings shellShortcut on).
Sessions sidebar fully keyboard/mouse navigable.
/model --session (-s): session-scoped model / reasoning-effort / context-window changes.
Inline hex color swatches (renderHexColors setting, on by default).
Multi-turn subagents always enabled.
Fixes (1.0.72)
agentStop hook infinite-loop fix: turn ends after 8 consecutive blocks; hooks get stop_hook_active flag.
Security: copilot skill list strips terminal control chars from names/descriptions (ANSI escape injection).
/cd no longer carries command approvals across repos; /worktree reliability; /mcp delete now stops background process.
Fixes (1.0.73)
Anthropic subagents keep working when extra directories are configured.
Relative links in custom agent instructions now resolved from the agent file's location.
Breaking / behavior shifts
Command restructure: top-level plugins → plugin; skill command folded under plugin; billing/limits/plugins interactive subcommands reorganized. config/environment are now help topics (copilot help config).
--model auto hint removed from --model description.
Impact: Medium. gh-aw generates non-interactive flags; verified MCPs still load and compile succeeds. The command restructure and --effort choice change are worth watching. Per constant note: verify MCPs not blocked and /models works with PAT auth.
260 .lock.yml regenerated (make build && make recompile x2).
Note: wasm golden testdata still reference gh-aw-mcpg:v0.4.1 (fixtures are not regenerated by recompile; consistent with prior runs). Lock files and js/*.js not committed per workflow policy.
Overview
Four agentic CLI/MCP versions updated in
pkg/constants/version_constants.go.make build && make recompile(x2) succeeded: 260/260 workflows compiled, 0 errors, 219 warnings.No change: Copilot SDK 1.0.7 · Pi 0.80.10 · Playwright MCP 0.0.78 · Playwright CLI 0.1.17 · GitHub MCP Server v1.6.0 · Playwright Browser v1.61.1.
Claude Code 2.1.214 → 2.1.216
Published: 2.1.215 @ 2026-07-19, 2.1.216 @ 2026-07-20. No public repo; NPM +
--help+ changelog analysis.Features
--forward-subagent-text: forwards subagent text/thinking blocks as assistant/user messages withparent_tool_use_idset (only with--print+--output-format=stream-json).sandbox.filesystem.disabledsetting: skips filesystem isolation while keeping network egress control.Fixes / Behavior
/verifyand/code-reviewskills — must be invoked explicitly (minor behavior change).auto-modesubcommand help: "Inspect" → "Inspect or reset auto mode classifier".Impact: Low. Additive flag/setting; auto-skill behavior change does not affect gh-aw runs.
Help diff 2.1.215 → 2.1.216
Copilot CLI 1.0.71 → 1.0.73
Published: 1.0.72 @ 2026-07-20, 1.0.73 @ 2026-07-21. Prerelease
1.0.72-1skipped. The bulk of changes landed in the large 1.0.72 release.Features (1.0.72)
$at prompt opens a shell in the session dir (opt-in via/settings shellShortcut on)./model --session(-s): session-scoped model / reasoning-effort / context-window changes.copilot plugins install --skill <file|URL|dir>(+--scope project).renderHexColorssetting, on by default).Fixes (1.0.72)
agentStophook infinite-loop fix: turn ends after 8 consecutive blocks; hooks getstop_hook_activeflag.copilot skill liststrips terminal control chars from names/descriptions (ANSI escape injection)./cdno longer carries command approvals across repos;/worktreereliability;/mcp deletenow stops background process.Fixes (1.0.73)
Breaking / behavior shifts
plugins→plugin;skillcommand folded underplugin;billing/limits/pluginsinteractive subcommands reorganized.config/environmentare now help topics (copilot help config).--effortchoices dropped"minimal"(now none/low/medium/high/xhigh/max).--model autohint removed from--modeldescription.Impact: Medium. gh-aw generates non-interactive flags; verified MCPs still load and compile succeeds. The command restructure and
--effortchoice change are worth watching. Per constant note: verify MCPs not blocked and/modelsworks with PAT auth.Refs: https://github.com/github/copilot-cli/releases
OpenAI Codex 0.144.5 → 0.144.6
Latest stable; alpha
0.145.0-alpha.28skipped.Fixes
No
--helpdiff. Impact: Low (bundled model metadata only).Refs: https://github.com/openai/codex/releases/tag/rust-v0.144.6 · openai/codex@rust-v0.144.5...rust-v0.144.6
MCP Gateway (gh-aw-mcpg) v0.4.1 → v0.4.3
Published: v0.4.2 @ 2026-07-19, v0.4.3 @ 2026-07-20. Used as default sandbox.agent container.
Container digest resolved and pinned:
ghcr.io/github/gh-aw-mcpg:v0.4.3@sha256:3c744710ea275cd5ee65db92a1099e0d980754bd9fafda9ce67704c67004dc83. Updated.github/aw/actions-lock.json+ syncedpkg/actionpins/data/action_pins.jsonandpkg/workflow/data/action_pins.json.Impact: Low. Internal quality/security-guard work only.
Refs: https://github.com/github/gh-aw-mcpg/releases/tag/v0.4.3 · https://github.com/github/gh-aw-mcpg/releases/tag/v0.4.2
Changes
pkg/constants/version_constants.go: 4 version constants updated..github/aw/actions-lock.json+ 2 syncedaction_pins.json: MCPG v0.4.3 container pin..lock.ymlregenerated (make build && make recompilex2).Note: wasm golden testdata still reference
gh-aw-mcpg:v0.4.1(fixtures are not regenerated byrecompile; consistent with prior runs). Lock files andjs/*.jsnot committed per workflow policy.References: §29806407418
Warning
Firewall blocked 2 domains
The following domains were blocked by the firewall during workflow execution:
awmgmcpgmodels.devSee Network Configuration for more information.