fix: register enabled extensions for agent on integration use/upgrade (#2949)

* fix: register enabled extensions for agent on integration install/upgrade

install and upgrade only set up the integration's own core commands; only
switch re-registered the enabled extensions' commands for the target agent.
A second integration added via install (or refreshed via upgrade) was
therefore silently missing the extension commands the existing agents
already had (e.g. the bundled agent-context extension).

Extract switch's registration into a shared _register_extensions_for_agent
helper and call it from install and upgrade too, so every installed agent
ends up with every enabled extension's commands — full parity with switch.

Closes #2886

* test: pin skills-mode secondary-agent registration; document #2948 limitation

Extension skill rendering is scoped to the active agent (init-options track a
single ai / ai_skills pair), so a skills-mode agent registered while not active
(e.g. Copilot --skills installed as a secondary integration) gets command files
rather than skills. install/upgrade match extension add here; only switch
renders skills, because it activates the target first.

Add a regression test pinning this behavior and document the limitation on the
shared helper. Per-agent skills parity is tracked separately in #2948.

* fix: don't re-render the active agent's skills when registering a non-active agent

register_enabled_extensions_for_agent runs an active-agent-scoped skills pass
(_register_extension_skills resolves the skills dir from init-options["ai"],
ignoring the passed agent). Routing install/upgrade of a secondary integration
through it re-rendered the *active* skills-mode agent's extension skills as a
side effect — resurrecting skill files the user had deliberately deleted. Gate
the skills pass on the target being the active agent; switch is unaffected
because it activates the target first.

Also harden the skills-mode install test (assert a core skill so --skills is
load-bearing, drop a vacuous registered_skills assertion) and add a regression
test. Surfaced by review of the PR; skills parity for non-active agents stays
tracked in #2948.

* refactor: share the extension-op scaffold and run (un)registration post-commit

Review cleanups, no behavior change on the success path:

- Extract the best-effort ExtensionManager scaffold (lazy import, instantiate,
  except -> _print_cli_warning) into _best_effort_extension_op. Both
  _register_extensions_for_agent and a new _unregister_extensions_for_agent
  delegate to it, removing the duplicate block left inline in switch.
- Invoke the best-effort extension registration AFTER the install/switch/upgrade
  try/except has committed, so a failure in it can never trigger the rollback
  (install and switch teardown on except).

* docs: clarify extension registration parity scope

* fix(integrations): defer extension registration until use

* fix(tests): remove redundant shutil import

* fix(integrations): backfill extensions for installed switch targets
This commit is contained in:
Pascal THUET
2026-06-23 00:48:55 +02:00
committed by GitHub
parent afe7657d2c
commit ce01877610
5 changed files with 437 additions and 71 deletions

View File

@@ -1678,16 +1678,12 @@ class ExtensionManager:
def register_enabled_extensions_for_agent(self, agent_name: str) -> None:
"""Register installed, enabled extensions for ``agent_name``.
This is intended to be called after switching integrations. Command
registration is scoped to the explicit ``agent_name`` argument, but some
behavior still depends on the current init-options state (for example,
skills-mode handling uses the active ``ai`` / ``ai_skills`` settings).
Callers should therefore pass the agent that has just been made active
in init-options; in normal use, ``agent_name`` is expected to match the
current ``ai`` value. This mirrors extension install behavior while
avoiding stale default-mode command directories when that active agent
is running in skills mode (notably Copilot ``--skills``).
Command-file registration is scoped to the explicit ``agent_name``
argument, so this method can be used after install, upgrade, or switch.
Extension skill rendering is still scoped to the active ``ai`` /
``ai_skills`` settings in init-options, so non-active skills-mode
targets receive command files here. Per-agent skills parity is tracked
separately in #2948.
"""
if not agent_name:
return
@@ -1744,31 +1740,46 @@ class ExtensionManager:
if new_registered != registered_commands:
updates["registered_commands"] = new_registered
try:
registered_skills = self._register_extension_skills(manifest, ext_dir)
except Exception as skills_err:
# Skills are a companion artifact. If command registration
# already succeeded, still persist it so later cleanup can
# find those command files.
from .. import _print_cli_warning
_print_cli_warning(
"register extension skills for",
"extension",
ext_id,
skills_err,
continuing=(
"Continuing with available registration results for this "
"extension and the remaining extensions."
),
)
else:
if registered_skills:
existing_skills = self._valid_name_list(
metadata.get("registered_skills", [])
# Extension *skills* are only ever rendered for the active agent:
# `_register_extension_skills` resolves the skills dir and
# frontmatter from init-options["ai"], ignoring ``agent_name``.
# When this method runs for a non-active agent — as install/upgrade
# now do for a secondary integration (#2886) — the skills pass would
# re-render the *active* agent's extension skills as a side effect,
# resurrecting skill files the user deliberately deleted. Skip it
# unless the target is the active agent; `switch` is unaffected
# because it activates the target before registering. (Rendering
# skills for a non-active target is tracked separately in #2948.)
if agent_name == active_agent:
try:
registered_skills = self._register_extension_skills(
manifest, ext_dir
)
merged_skills = list(dict.fromkeys(existing_skills + registered_skills))
updates["registered_skills"] = merged_skills
except Exception as skills_err:
# Skills are a companion artifact. If command registration
# already succeeded, still persist it so later cleanup can
# find those command files.
from .. import _print_cli_warning
_print_cli_warning(
"register extension skills for",
"extension",
ext_id,
skills_err,
continuing=(
"Continuing with available registration results for this "
"extension and the remaining extensions."
),
)
else:
if registered_skills:
existing_skills = self._valid_name_list(
metadata.get("registered_skills", [])
)
merged_skills = list(
dict.fromkeys(existing_skills + registered_skills)
)
updates["registered_skills"] = merged_skills
if updates:
self.registry.update(ext_id, updates)