Repository navigation
[deps-release-notes] copilot-cli — upstream release action items #2041
Copy link
Copy link
Closed as not planned
Labels
automationdependenciesPull requests that update a dependency filePull requests that update a dependency file
Description
Rolling upstream release action items — copilot-cli
This is the single canonical tracking issue for action items arising from
new releases of the copilot-cli dependency. The update-awf-version workflow
appends a new comment to this issue for each version bump going forward, so
the most recent activity lives in the comments below. This body is a
consolidated history of everything filed so far.
Latest pinned version covered: 1.0.81
Consolidated history (earliest → latest)
1.0.48 → 1.0.56 (was #803)
- Security: secret scanning covers commit messages + PR descriptions (v1.0.51): Redacts secrets before publishing.
- Security:
permissions.disableBypassPermissionsMode(v1.0.55): Prevents enabling allow-all/yolo mode; relevant for sandboxed pipelines. --session-id=<id>flag (v1.0.51): Resume/start sessions with a specific UUID — useful for pipeline correlation.postToolUsehookadditionalContextinjected as system message (v1.0.49): Instead of being discarded.preMcpToolCallhook (v1.0.51): Control outgoing MCP request metadata.deferred-tool-loadingfrontmatter for custom agents (v1.0.52): Reduces initial token overhead.- MCP tools with both
content+structuredContentsurface both (v1.0.56): Review ado-aw MCP tools to ensure dual output is intentional. - Context window tier selection (v1.0.52): Default ~200K vs 1M tokens enforced end-to-end.
- Claude Opus 4.8 support (v1.0.55).
- Deprecation:
oauth.clientId/oauth.callbackPort→oauthClientId/auth.redirectPort(v1.0.52).
1.0.48 → 1.0.59 (was #843)
- Breaking:
preToolUsehook errors now deny the tool call (v1.0.57): Previously hook errors silently allowed execution; now they block the call. - Breaking: Remote JSON RPC enabled by default (v1.0.58): Verify this doesn't open unexpected network surface in AWF-sandboxed environments.
- Breaking:
--plugin-dirskills take precedence over personal-home skills (v1.0.55): Skill resolution order: project > plugin-dir > personal > custom. - ADO-only repos: built-in GitHub MCP server exposes only
web_search(v1.0.57): Directly relevant to ado-aw; consider whether the GitHub MCP extension should account for ADO-repository detection.
1.0.48 → 1.0.60 (was #881)
- Alpine Linux (musl libc) support (v1.0.49): Now runs on Alpine-based containers.
- Rubber Duck agent enabled by default (v1.0.58): Check whether this adds unexpected turn overhead in automated runs.
1.0.60 → 1.0.61 (was #951)
- Auto-load MCP servers from
.github/mcp.json(v1.0.61): ado-aw uses--additional-mcp-config; consider guarding against auto-loading interference. - Claude Fable 5 model support (v1.0.61): Add to validated models in
src/engine.rs/docs/engine.md. - mTLS + private-CA for OTLP telemetry (v1.0.61): Added
http/protobufOTLP export via OTel env vars. - Bug fix:
grep/globsingle-path fix (v1.0.61): Previously missed results with single path argument. - Bug fix: shell validation false positives (v1.0.61): Fixed blocking of harmless commands with words like "kill" in string literals/heredocs.
1.0.60 → 1.0.62 (was #1002)
- Breaking:
write_bashinteractive input removed (v1.0.62): Shell commands now use lightweight process spawning; interactive input no longer supported. ado-aw exposeswrite_bashas a first-class tool — assess whether tool surface/docs need updating. - Configure subagent model, reasoning effort, context tier (v1.0.62): Consider surfacing
reasoning_effort/context-tier as front-matter fields. - Kerberos/SPNEGO proxy auth (v1.0.62): Auto-auth through corporate forward proxies; relevant for enterprise ADO environments.
1.0.62 → 1.0.63 (was #1057)
deferToolsMCP server config option (v1.0.63): Keeps server tools available even when tool search is enabled. ado-aw could use this for safe-outputs and GitHub MCP servers to prevent them from being filtered out.
1.0.62 → 1.0.64 (was #1183)
- New OTel cache + reasoning token fields (v1.0.64):
gen_ai.usage.cache_read.input_tokens,gen_ai.usage.cache_creation.input_tokens,gen_ai.usage.reasoning.output_tokensper GenAI semconv (old underscore-separated names gone). Updatesrc/agent_stats.rsto parse new attribute names. - Static OAuth client overrides for MCP server auth (v1.0.64): ado-aw could expose in MCP server config front-matter.
- Autopilot mode auto-handles
ask_user, elicitation, sampling, permission prompts (v1.0.64): Reliability improvement for unattended pipeline runs. - Deprecation:
report_intenttool removed (v1.0.64):INTERNAL_TOOL_NAMESexclusion insrc/agent_stats.rscontaining"execute_tool report_intent"is now dead code; can be cleaned up.
1.0.64 → 1.0.67 (was #1272)
- Claude Sonnet 5 model support (v1.0.67): Add to
src/engine.rsmodel allowlist + docs. - Subagent sessions inherit parent tool restrictions (v1.0.67): Security-relevant behavior change; verify against expected security model.
- MCP OAuth against Microsoft Entra vanity domain servers (v1.0.67): Fixes
AADSTSauth failures for Entra vanity domains; relevant for enterprise Azure AD MCP tools. copilot skillsubcommand (v1.0.65): New/skillalias for listing/adding/removing skills; may be relevant for agency plugin scaffolding.- Deprecation: Claude Opus 4.6 Fast → Claude Opus 4.8 Fast (v1.0.66): Update any ado-aw references to
claude-opus-4.6-fast.
1.0.64 → 1.0.68 (was #1305)
- Breaking: session limits require minimum 30 AI credits (v1.0.67): Pipelines with
session-limit:below 30 AI credits will be rejected at runtime. Review compiler defaults. - Claude Opus 4.8 Fast model support (v1.0.66): Add
claude-opus-4.8-fasttosrc/engine.rsallowlist. - kimi-k2.7-code model support (v1.0.68): Add to model allowlist.
1.0.70 → 1.0.81 (was #2041) — canonical
- Breaking (v1.0.79): Sandbox settings keys renamed from
sandbox.gitAuth/sandbox.ghAuthtosandbox.auth.git/sandbox.auth.gh, with no migration path. Old keys are silently ignored, and SDK requests using the old keys are rejected as invalid. Any ado-aw-generated engine config or docs referencing the old sandbox setting names need to be updated. See the v1.0.79 release. - Breaking (v1.0.79):
--worktreedefault behavior changed —worktreeBaseRefnow defaults toHEADinstead of the remote default branch. Could change output for any ado-aw workflow relying on worktree creation defaults. See the v1.0.79 release. - Security (v1.0.78): The
/allow-all autosafety-judge model is no longer user-configurable — hardens against downgrade to a weaker safety judge. See the v1.0.78 release. - Security (v1.0.78): Managed-settings fetch failure behavior changed from fail-closed to fail-open — worth confirming this doesn't weaken ado-aw's sandboxed-agent posture when managed settings can't be reached. See the v1.0.78 release.
- Security (v1.0.72): Sandbox macOS keychain access now defaults to off, reducing default credential exposure inside the sandboxed agent. See the v1.0.72 release.
- v1.0.81: Support for MCP protocol version 2026-07-28, OpenTelemetry trace-context propagation (
traceparent/tracestate) into hooks, and per-agent usage metrics via--usage-output-file— could improve ado-aw's OTel-based agent stats correlation (src/agent_stats.rs). See the v1.0.81 release. - v1.0.81: Windows Entra ID broker-based MCP auth (no-prompt sign-in) — relevant if ado-aw ever needs Windows-hosted agent auth flows.
- v1.0.81: Plan mode now hard-blocks mutating tool calls — a safety property ado-aw could rely on/document for plan-only agent configurations.
- v1.0.76/v1.0.77: Enterprise admins can enforce a restrictive sandbox floor and managed sandbox policy via macOS/Windows native MDM — relevant to ado-aw's OneBranch sandboxing story for enterprise consumers. See the v1.0.76 release and v1.0.77 release.
- v1.0.72: Opt-in git/gh auth inside the OS sandbox — worth evaluating against ado-aw's own credential-isolation model (
ado-proxy).
Consolidated by the Deps Release-Notes Consolidator workflow. Superseded per-release issues were closed and point here.> Generated by Deps Release-Notes Consolidator · auto · 76.3 AIC · ⌖ 11.6 AIC · ⊞ 10.6K · ◷
Reactions are currently unavailable
Metadata
Metadata
Assignees
Labels
automationdependenciesPull requests that update a dependency filePull requests that update a dependency file
Activity
github-actions commentedon Sep 5, 2026
copilot-cli1.0.70→1.0.83Releases analyzed
Breaking changes
sandbox.gitAuth/sandbox.ghAuthsettings keys renamed tosandbox.auth.git/sandbox.auth.gh, no migration (v1.0.79): old keys are silently ignored in settings files, and SDK requests sending the old keys are now rejected as invalid rather than ignored.--worktreenow defaults to starting from HEAD instead of the remote default branch (v1.0.79), controllable via the newworktreeBaseRefsetting./sandbox. Worth checking whether any ado-aw sandboxed agent workflow binds local ports.copilot's own sandbox mode is ever exercised inside it.PLUGINS_DASHBOARDopt-out and the legacy skills picker (v1.0.81):/skills, bare/mcp, and/mcp showalways open the dashboard now;/pluginscommand removed (moved to/plugin,/mcp,/skills).Security fixes
permissions.disableBypassPermissionsModeunrecognized values now enforced asdisableinstead of rejected (v1.0.81): hardens enterprise-managed policy handling for the bypass/allow-all mode ado-aw's sandboxed agent must not enable.forceRemoteSettingsRefreshnow fails closed (v1.0.82 chain via v1.0.81 notes): cached managed-settings policy is never served as a fetch-failure fallback when this is set; until a fresh policy is fetched, non-default MCP servers are blocked and bypass-permissions mode cannot be enabled — a stricter default posture worth being aware of for network-isolated ado-aw runs.Notable features for ado-aw to adopt
--session-id-adjacent correlation and per-agent usage metrics in--usage-output-fileJSON (v1.0.81): could improve OTel-based agent stats parsing (src/agent_stats.rs).traceparent/tracestate) (v1.0.81): potential integration point for ado-aw's audit/trace tooling.models.listnow includes service-publishedinfoMessages/warningMessagesper model (v1.0.81): could surface deprecation/availability warnings to ado-aw authors viacatalog/lint.--enable-mcp-serverflag to re-enable MCP servers disabled in settings for a single run (v1.0.80): could be useful for MCPG diagnostics.usage_updatenotifications (v1.0.78): more granular data ado-aw's audit analyzers could ingest.Deprecations
sandbox.gitAuth/sandbox.ghAuth→sandbox.auth.git/sandbox.auth.gh(v1.0.79) — see Breaking changes above; no migration path, old keys silently ignored./pluginscommand removed in favor of/plugin,/mcp,/skills(v1.0.81).Warning
Firewall blocked 1 domain
The following domain was blocked by the firewall during workflow execution:
index.crates.ioTo allow these domains, add them to the
network.allowedlist in your workflow frontmatter:See Network Configuration for more information.
github-actions commentedon Sep 6, 2026
copilot-cli1.0.81→1.0.83Releases analyzed
Notable features for ado-aw to adopt
model, tried in order until one is available, withmodel-policy: required(v1.0.83): A model-fallback list feature for custom agents. ado-aw'sengine:front matter currently only supports a singlemodel:value — this could be surfaced as a fallback list to improve reliability when a preferred model is temporarily unavailable.~/.npmrc;sandbox.allowDevToolAccesscan disable this (v1.0.83): Worth double-checking against ado-aw's own credential-isolation assumptions in the AWF sandbox — files like~/.npmrccan carry registry auth tokens, and this widens what sandboxed file-read tools can see by default.ghcommands now authenticate as the account configured for the repository instead of the Copilot CLI login (v1.0.83): Behavior change for any sandboxedghinvocation inside the agent; worth confirming this doesn't conflict with ado-aw's own GitHub MCP / credential-isolation model.No breaking changes, security-labeled fixes, or deprecations were called out in the v1.0.82/v1.0.83 release notes beyond the items already logged in the prior
1.0.81comment on this issue.Warning
Firewall blocked 1 domain
The following domain was blocked by the firewall during workflow execution:
index.crates.ioTo allow these domains, add them to the
network.allowedlist in your workflow frontmatter:See Network Configuration for more information.
github-actions commentedon Sep 12, 2026
copilot-cli1.0.70→1.0.83Releases analyzed
Breaking changes
sandbox.gitAuth/sandbox.ghAuthsettings keys renamed tosandbox.auth.git/sandbox.auth.ghwith no migration (v1.0.79) — old keys are silently ignored in settings files, and SDK requests that still send them are rejected as invalid. If ado-aw ever generates or documentssettings.json/SDK-request payloads for the Copilot CLI engine (e.g. sandbox git/gh auth config), those references need updating to the new keys.slirp4netns,nsenter,iptables,ip6tables,iptables-restore,ip6tables-restoreonPATH; proxy-mode egress restriction additionally requiresutil-linux 2.35+and/dev/net/tun(v1.0.83) — the ado-aw agent sandbox image/runner (OneBranch build agents invoking thecopilotCLI) must have these binaries available or sandboxed shell commands will fail to launch after upgrading to 1.0.83.127.0.0.1(v1.0.83) — any ado-aw agent workflow whose bash tool binds a local port (e.g. for a test server) will needAllow local networkenabled in/sandbox, or an equivalent CLI/config flag if ado-aw wires sandbox settings.COPILOT_CLI_VERSION-pinned engine config; worth confirming this doesn't weaken assumptions indocs/engine.md.Security fixes
copilot skill liststrips terminal control characters from skill names/descriptions, preventing a crafted skill from injecting ANSI escape sequences into listing output (v1.0.72).allowWrite/allowReadboundaries.userPromptSubmittedhook edge cases that could corrupt a session.ghcommands now authenticate as the account configured for the repository instead of the Copilot CLI login (v1.0.83) — reduces the risk of using the wrong credential scope forghoperations inside the sandbox.Notable features for ado-aw to adopt
azure-devopsor other MCP tool integrations need standards-based OAuth client registration.ado-proxy/API-proxy TLS story if the engine now handles client-cert auth natively.--with-tokentocopilot loginto read an auth token from stdin (v1.0.81) — could simplify non-interactiveCOPILOT_CLI_VERSION-pinned login flows in the Agent stage compared to existing token-injection mechanisms.--usage-output-fileJSON output (v1.0.81) — could enrichsrc/agent_stats.rs's OTel-based token/usage parsing with a structured per-agent breakdown if ado-aw adopts this output format.Deprecations
/theme's deprecation-notice display logic was itself bug-fixed in v1.0.79, not a new deprecation.)github-actions commentedon Sep 16, 2026
copilot-cli1.0.70→1.0.85Releases analyzed
Note: this comment covers only the increment beyond
1.0.81, since prior bumps up to that version were already recorded above.Breaking changes
PATH(slirp4netns,nsenter,iptables,ip6tables,iptables-restore,ip6tables-restore) (v1.0.83) — sandboxed commands will fail to launch without these. Proxy-mode egress restriction additionally requiresutil-linux 2.35+and/dev/net/tunaccess. Worth verifying these are present in the AWF/OneBranch agent image ado-aw compiles into, since ado-aw invokescopilotinside that sandbox.127.0.0.1(v1.0.83) — this could break any ado-aw agent workflow that starts a local server and expects to curllocalhostfrom the same sandboxed session; the release notes point at "Allow local network" in/sandboxas the opt-back-in.Notable features for ado-aw to adopt
docs/mcp.md.copilot initnow removes the.githubdirectory it created when the run exits without writing an instructions file (v1.0.85) — behavior change relevant if ado-aw's owninitflow or docs referencecopilot initside effects./vim(v1.0.85) — purely an interactive-CLI UX feature, not applicable to ado-aw's non-interactive agent invocations.Deprecations
1.0.81→1.0.85) beyond UX/CLI surface changes already covered above; the remaining items were interactive-terminal UX fixes, model-picker/session-management polish, and internal refactors with no consumer-visible effect on ado-aw's non-interactivecopilotinvocation.Warning
Firewall blocked 1 domain
The following domain was blocked by the firewall during workflow execution:
index.crates.ioTo allow these domains, add them to the
network.allowedlist in your workflow frontmatter:See Network Configuration for more information.
github-actions commentedon Sep 20, 2026
copilot-cli1.0.70→1.0.86Releases analyzed
Breaking changes
allowDevToolCachessandbox setting renamed toallowDevToolAccess(v1.0.79): "BREAKING: the sandbox settingallowDevToolCachesis renamedallowDevToolAccess... The old key is no longer read and is ignored silently, so an existingfalseopt-out reverts to the default (on)." Any ado-aw-generated sandbox config or docs referencingallowDevToolCaches(seedocs/network.md/ sandbox-related front matter) must be updated to the new key, since a silent revert-to-default could widen dev-tool access unexpectedly for pipelines that had opted out./pluginscommand andPLUGINS_DASHBOARDopt-out removed (v1.0.81): "Removed/plugins; its resources moved to/plugin,/mcpand/skills... Removed thePLUGINS_DASHBOARDopt-out and the legacy skills picker it kept alive." Not directly agentic-workflow-facing (interactive-only surfaces), but any ado-aw docs/scripts referencing these commands should be checked.copilot's own sandboxing (as opposed to relying solely on AWF's isolation), these new binary/kernel-capability requirements should be verified against the sandbox image.Security fixes
agentStophook infinite-loop guard (v1.0.72): anagentStophook that always blocks now force-ends the turn after 8 consecutive blocks, with astop_hook_activeflag for hook authors to detect forced continuation.userPromptSubmittedhook responses.agency/plugins/ado-aw/skills/.Notable features for ado-aw to adopt
traceparent/tracestate, letting hook-emitted spans correlate with the agent's trace — could improveado-aw audit's OTel-based agent stats (src/agent_stats.rs) if ado-aw ever adds custom hooks.--usage-output-file(v1.0.81): richer usage breakdown thatado-aw audit's OTel/usage analyzer could surface per-subagent.--with-tokenreads an auth token from stdin (v1.0.81): a non-interactive credential-injection path that may be relevant to how ado-aw wires the Copilot engine's auth in sandboxed pipeline steps.Deprecations
Warning
Firewall blocked 1 domain
The following domain was blocked by the firewall during workflow execution:
index.crates.ioTo allow these domains, add them to the
network.allowedlist in your workflow frontmatter:See Network Configuration for more information.
github-actions commentedon Sep 22, 2026
copilot-cli1.0.70→1.0.87Releases analyzed
(Releases v1.0.71–v1.0.81 and v1.0.84 either did not exist as public tags or were not returned by the release listing; the above are the releases actually present in the
1.0.70 < v <= 1.0.87range.)Breaking changes
copilot login --hostnow works, and--max-autopilot-continuesno longer accepts scientific notation. ado-aw does not pass--max-autopilot-continuesor parsecopilot --helpoutput programmatically, so no immediate action, but worth a quick check if any prompt/skill documentation quotes old CLI help text verbatim.copilot pluginssubcommand surface reorganized (v1.0.85):copilot plugins install --skill→copilot skill add; cross-kind--kind/--scope/--mcp/--skillflags removed fromcopilot plugins;copilot plugins list --jsonnow emits a flat plugin array instead of{ plugins, errors };copilot plugins listis now an alias ofcopilot plugin list(plugins only, not MCP/skills/instructions/LSP). ado-aw doesn't invoke anycopilot plugins/copilot plugin/copilot mcp/copilot skillsubcommands in generated pipeline steps, so no action needed for the compiler itself — flagging in case any doc snippet inprompts/ordocs/references the oldcopilot plugins install --skillform.127.0.0.1. ado-aw runs the agent inside AWF's own network-isolation sandbox (not relying on Copilot CLI's own sandbox for network policy), so this shouldn't change AWF-mediated network behavior, but is worth confirming if any workflow relies on the agent binding a local port for internal tooling.slirp4netns,nsenter,iptables,ip6tables,iptables-restore,ip6tables-restoreon PATH (v1.0.83) for proxy-mode network egress restriction. Relevant if the AWF/agent container image doesn't already provide these; worth a compatibility check against the imagesado-awpulls for the agent step.Security fixes
ado-proxyaddresses.allowManagedHooksOnlypolicy now also blocks extension-registered hook callbacks (preToolUse,postToolUse,postToolUseFailure) that previously bypassed managed-only lockdown (v1.0.85).Notable features for ado-aw to adopt
include-custom-instructions: truecustom-agent frontmatter flag (v1.0.86) lets a custom agent opt into repository instruction files (AGENTS.md, copilot-instructions.md, CLAUDE.md). Could be surfaced as a front-matter passthrough for ado-aw agent files that want repo-level instructions automatically merged.modelvalues, tried in order, withmodel-policy: required(v1.0.83) — a fallback-model list could be a usefulengine.modelextension for ado-aw'sengine:front-matter.--upstream-proxyAWF integration requiring client certs.engine.model: autobehavior in ado-aw pipelines.slowConnectionThresholdMsper-MCP-server configuration (v1.0.87) — could be exposed via ado-aw's MCP server config surface for tuning slow user-defined MCP servers.Deprecations
--no-auto-loginand--no-sandboxremain supported but stay hidden from--help(v1.0.85) — unchanged behavior, just no longer documented in help output.copilot plugins list/install --skill/cross-kind flags superseded bycopilot plugin/copilot skill/copilot mcpequivalents (see Breaking changes above) — the old spellings still work as aliases per the release notes, but new documentation should prefer the new subcommand forms.Warning
Firewall blocked 1 domain
The following domain was blocked by the firewall during workflow execution:
index.crates.ioTo allow these domains, add them to the
network.allowedlist in your workflow frontmatter:See Network Configuration for more information.
github-actions commentedon Sep 26, 2026
Consolidated into the canonical rolling issue #2232 ( [deps-release-notes] copilot-cli — upstream release action items ). Its action items have been preserved there.