Skip to content

:hedgehog: PostHog is the leading platform for building self-driving products. Our developer tools – AI observability, analytics, session replay, flags, experiments, error tracking, logs, and more – capture all the context agents need to diagnose problems, uncover opportunities, and ship fixes. Steer it all from Slack, web, desktop, or the MCP.

fix(tasks): disable skills during repository selection (#128138)

Georgiy Tarasov45b2133d219e
diff --git a/packages/agent/packages/agent-contracts/src/domain-types.ts b/packages/agent/packages/agent-contracts/src/domain-types.ts
index d0f67c455cf8..67ab3a9e69f7 100644
--- a/packages/agent/packages/agent-contracts/src/domain-types.ts
+++ b/packages/agent/packages/agent-contracts/src/domain-types.ts
@@ -371,6 +371,7 @@ const taskRunStateFields = {
codex_max_subagents: optionalField(z.number().int().positive()),
codex_shared_prompt_cache: optionalField(z.boolean()),
disable_repo_hooks: optionalField(z.boolean()),
+ disable_skills: optionalField(z.boolean()),
initial_permission_mode: optionalField(executionModeSchema),
initial_prompt_override: optionalField(z.string()),
pending_followup_messages: optionalField(
diff --git a/packages/agent/packages/agent/src/server/agent-server.test.ts b/packages/agent/packages/agent/src/server/agent-server.test.ts
index 26660cace27a..9d35ecdc0218 100644
--- a/packages/agent/packages/agent/src/server/agent-server.test.ts
+++ b/packages/agent/packages/agent/src/server/agent-server.test.ts
@@ -1009,6 +1009,41 @@ describe("AgentServer HTTP Mode", () => {
30000,
);
+ it.each([true, false])(
+ "applies the caller's skill setting (%s)",
+ async (disableSkills) => {
+ mswServer.use(
+ http.get(
+ "http://localhost:8000/api/projects/:projectId/tasks/:taskId/runs/:runId/",
+ () =>
+ HttpResponse.json(
+ createTaskRun({
+ id: "test-run-id",
+ state: { disable_skills: disableSkills },
+ }),
+ ),
+ ),
+ );
+ mockedClaudeSdk.query.mockClear();
+ const s = createServer({
+ runtimeAdapter: "claude",
+ reasoningEffort: "high",
+ });
+ await s.start();
+ const request = mockedClaudeSdk.query.mock.lastCall?.[0] as unknown as {
+ options: { skills?: string[]; effort?: string; tools?: unknown };
+ };
+ expect(request?.options?.skills).toEqual(
+ disableSkills ? [] : undefined,
+ );
+ expect(request?.options?.effort).toBe("high");
+ expect(request?.options?.tools).toEqual({
+ type: "preset",
+ preset: "claude_code",
+ });
+ },
+ );
+
it("keeps the run-state system prompt after a transient run fetch failure", async () => {
let attempts = 0;
mswServer.use(
diff --git a/packages/agent/packages/agent/src/server/agent-server.ts b/packages/agent/packages/agent/src/server/agent-server.ts
index a330a3cb31c1..450492d87910 100644
--- a/packages/agent/packages/agent/src/server/agent-server.ts
+++ b/packages/agent/packages/agent/src/server/agent-server.ts
@@ -2496,7 +2496,7 @@ export class AgentServer {
this.config.fastMode !== undefined && {
fastMode: this.config.fastMode,
}),
- ...this.buildClaudeCodeSessionMeta(runtimeAdapter),
+ ...this.buildClaudeCodeSessionMeta(runtimeAdapter, launch.disableSkills),
};
await this.measureInitialization("repository_ready", () =>
@@ -4570,12 +4570,13 @@ export class AgentServer {
*/
private buildClaudeCodeSessionMeta(
runtimeAdapter: Adapter,
+ disableSkills = false,
): { claudeCode: { options: Record<string, unknown> } } | undefined {
const plugins = this.config.claudeCode?.plugins;
const effort =
runtimeAdapter === "claude" ? this.config.reasoningEffort : undefined;
- if (!plugins?.length && !effort) {
+ if (!plugins?.length && !effort && !disableSkills) {
return undefined;
}
@@ -4586,6 +4587,9 @@ export class AgentServer {
if (effort) {
options.effort = effort;
}
+ if (disableSkills) {
+ options.skills = [];
+ }
return { claudeCode: { options } };
}
diff --git a/packages/agent/packages/agent/src/server/launch/run-launch.ts b/packages/agent/packages/agent/src/server/launch/run-launch.ts
index 07041fdfbb0c..6c03aede61a3 100644
--- a/packages/agent/packages/agent/src/server/launch/run-launch.ts
+++ b/packages/agent/packages/agent/src/server/launch/run-launch.ts
@@ -36,6 +36,7 @@ export interface RunLaunch {
readonly endRunWhenDone: boolean;
readonly callerEndsRun: boolean;
readonly disableRepoHooks: boolean;
+ readonly disableSkills: boolean;
readonly systemPrompt: ClaudeCodeConfig["systemPrompt"];
readonly systemPromptValid: boolean;
readonly tokenUsage: unknown;
@@ -71,6 +72,7 @@ export function readRunLaunch(sources: RunLaunchSources): RunLaunch {
endRunWhenDone: runState?.end_run_when_done === true,
callerEndsRun: runState?.caller_ends_run === true,
disableRepoHooks: runState?.disable_repo_hooks === true,
+ disableSkills: runState?.disable_skills === true,
systemPrompt: systemPrompt.success ? systemPrompt.data : undefined,
systemPromptValid: systemPrompt.success,
tokenUsage: runState?.token_usage,
diff --git a/products/tasks/backend/facade/api.py b/products/tasks/backend/facade/api.py
index ece277439179..2f7558ae03ac 100644
--- a/products/tasks/backend/facade/api.py
+++ b/products/tasks/backend/facade/api.py
@@ -578,6 +578,7 @@ _TASK_RUN_PUBLIC_STATE_KEYS = frozenset(
"context_window",
"custom_image_id",
"disable_repo_hooks",
+ "disable_skills",
"fast_mode",
# An imported transcript's provenance and mirror cursor: the viewer marks the run as
# imported, and the legacy mirror reads its own cursor back through this DTO.
@@ -3029,6 +3030,7 @@ _PROTECTED_RUN_STATE_KEYS = frozenset(
"scout_trial_judge",
"scout_trial_private",
"include_live_context",
+ "disable_skills",
"analytics_query_context",
"sandbox_oauth_token_ids",
# Sandbox run-start lineage and fan-out count. A child sandbox that removed its parent keys
diff --git a/products/tasks/backend/logic/repo_selection/agent.py b/products/tasks/backend/logic/repo_selection/agent.py
index 070a1652af66..f93fea52736d 100644
--- a/products/tasks/backend/logic/repo_selection/agent.py
+++ b/products/tasks/backend/logic/repo_selection/agent.py
@@ -760,6 +760,7 @@ async def select_repository(
# Read-only PostHog scopes so the agent can call `execute-sql` against `system.integration_repository_cache`.
posthog_mcp_scopes="read_only",
github_read_access=True,
+ disable_skills=True,
sandbox_resources=SandboxResources(cpu_cores=2, memory_gb=8),
# Optional agent runtime/model override (e.g. the Codex runtime + gpt-5.5). All-None keeps
# the agent-server default; threaded through from the caller's per-step resolution.
diff --git a/products/tasks/backend/logic/services/custom_prompt_internals.py b/products/tasks/backend/logic/services/custom_prompt_internals.py
index 753fc0020d22..a380a55752d2 100644
--- a/products/tasks/backend/logic/services/custom_prompt_internals.py
+++ b/products/tasks/backend/logic/services/custom_prompt_internals.py
@@ -145,6 +145,7 @@ class CustomPromptSandboxContext:
"""Let runs that start from the same prompt prefix read each other's prompt cache. Codex-only. The
agent server asks the AI gateway to drop Codex's per-session cache key on every request of the run,
sub-agents included."""
+ disable_skills: bool = False
disable_repo_hooks: bool = False
"""Stop hooks from the cloned repo's ``.claude/settings*.json`` from running. A repo's Stop hook
can block the turn that must return the caller's JSON. The agent still reads the repo's CLAUDE.md.
@@ -363,6 +364,8 @@ async def _create_task_and_trigger(
extra_run_state["codex_max_subagents"] = context.codex_max_subagents
if context.codex_shared_prompt_cache:
extra_run_state["codex_shared_prompt_cache"] = True
+ if context.disable_skills:
+ extra_run_state["disable_skills"] = True
if context.disable_repo_hooks:
extra_run_state["disable_repo_hooks"] = True
if analytics_query_context is not None:
diff --git a/products/tasks/backend/logic/services/store_skills.py b/products/tasks/backend/logic/services/store_skills.py
index dff665b5b17e..6d828f72983e 100644
--- a/products/tasks/backend/logic/services/store_skills.py
+++ b/products/tasks/backend/logic/services/store_skills.py
@@ -88,6 +88,7 @@ def refresh_store_skills_state(task_run: TaskRun, user: User, *, reason: str) ->
store_skills = (
[]
if (task_run.state or {}).get("include_live_context") is False
+ or (task_run.state or {}).get("disable_skills") is True
else resolve_store_skills(task_run.task.team, user, run_id=run_id)
)
if store_skills is None:
diff --git a/products/tasks/backend/logic/services/test_store_skills.py b/products/tasks/backend/logic/services/test_store_skills.py
index 7fddebb5b3e4..ce4928bbf7a4 100644
--- a/products/tasks/backend/logic/services/test_store_skills.py
+++ b/products/tasks/backend/logic/services/test_store_skills.py
@@ -16,10 +16,12 @@ class TestRefreshStoreSkillsState(BaseTest):
(None, None, "unchanged"),
(False, True, "empty"),
(False, None, "empty"),
+ (None, True, "empty", True),
+ (None, None, "empty", True),
]
)
def test_refresh_respects_live_context_setting(
- self, include_live_context: bool | None, flag_value: bool | None, expected: str
+ self, include_live_context: bool | None, flag_value: bool | None, expected: str, disable_skills: bool = False
) -> None:
LLMSkill.objects.create(
team=self.team,
@@ -32,6 +34,7 @@ class TestRefreshStoreSkillsState(BaseTest):
)
task = Task.objects.create(team=self.team, title="Review evidence", created_by=self.user)
state: dict[str, object] = {STORE_SKILLS_STATE_KEY: [{"name": "stale-live-skill"}]}
+ state["disable_skills"] = disable_skills
if include_live_context is not None:
state["include_live_context"] = include_live_context
run = TaskRun.objects.create(task=task, team=self.team, state=state)
diff --git a/products/tasks/backend/temporal/process_task/activities/get_task_processing_context.py b/products/tasks/backend/temporal/process_task/activities/get_task_processing_context.py
index f1d5a60d496d..f3d1d34ca785 100644
--- a/products/tasks/backend/temporal/process_task/activities/get_task_processing_context.py
+++ b/products/tasks/backend/temporal/process_task/activities/get_task_processing_context.py
@@ -1073,7 +1073,7 @@ def get_task_processing_context(input: GetTaskProcessingContextInput) -> TaskPro
# sandbox needs no extra request on its boot path, and best-effort: a store failure must not
# stop the run, it only leaves the sandbox without store skills for this session.
try:
- if not include_live_context:
+ if not include_live_context or state.get("disable_skills") is True:
store_skills: list[dict[str, Any]] | None = []
state[STORE_SKILLS_STATE_KEY] = []
else:
diff --git a/products/tasks/backend/temporal/process_task/activities/tests/test_get_task_processing_context.py b/products/tasks/backend/temporal/process_task/activities/tests/test_get_task_processing_context.py
index 1e6ddaf5e499..0d63763b28e3 100644
--- a/products/tasks/backend/temporal/process_task/activities/tests/test_get_task_processing_context.py
+++ b/products/tasks/backend/temporal/process_task/activities/tests/test_get_task_processing_context.py
@@ -799,17 +799,27 @@ class TestGetTaskProcessingContextActivity:
@pytest.mark.django_db(transaction=True)
@pytest.mark.parametrize(
- "flag_value,include_live_context,expected_state",
+ "flag_value,include_live_context,expected_state,disable_skills",
[
- (True, None, "resolved"),
- (False, None, []),
- (None, None, "untouched"), # a flag-service outage must not clear stubs a resumed sandbox still has
- (True, False, []),
- (None, False, []),
+ (True, None, "resolved", False),
+ (False, None, [], False),
+ (None, None, "untouched", False), # a flag-service outage must not clear stubs a resumed sandbox still has
+ (True, False, [], False),
+ (None, False, [], False),
+ (True, None, [], True),
+ (None, None, [], True),
],
)
def test_store_skills_state_follows_the_sandbox_flag(
- self, activity_environment, test_task, user, mocker, flag_value, include_live_context, expected_state
+ self,
+ activity_environment,
+ test_task,
+ user,
+ mocker,
+ flag_value,
+ include_live_context,
+ expected_state,
+ disable_skills,
):
LLMSkill.objects.create(
team=test_task.team,
@@ -821,7 +831,7 @@ class TestGetTaskProcessingContextActivity:
created_by=user,
)
task_run = test_task.create_run(
- extra_state={"include_live_context": include_live_context} if include_live_context is not None else {}
+ extra_state={"include_live_context": include_live_context, "disable_skills": disable_skills}
)
TaskRun.update_state_atomic(task_run.id, updates={STORE_SKILLS_STATE_KEY: [{"name": "from-last-session"}]})
input_data = GetTaskProcessingContextInput(run_id=str(task_run.id))
@@ -839,7 +849,7 @@ class TestGetTaskProcessingContextActivity:
result = async_to_sync(activity_environment.run)(get_task_processing_context, input_data)
assert result.context_layer_enabled is (include_live_context is not False)
- if include_live_context is False:
+ if include_live_context is False or disable_skills:
assert result.state is not None
assert result.state[STORE_SKILLS_STATE_KEY] == []
for name in ("run_gateway_env_vars", "mcp_exec_skills_env_vars", "get_git_identity_env_vars"):
diff --git a/products/tasks/backend/temporal/process_task/tests/test_utils.py b/products/tasks/backend/temporal/process_task/tests/test_utils.py
index 227280ae2b19..0ae72baab0ee 100644
--- a/products/tasks/backend/temporal/process_task/tests/test_utils.py
+++ b/products/tasks/backend/temporal/process_task/tests/test_utils.py
@@ -1907,23 +1907,26 @@ class TestMcpExecSkillsEnvVars(SimpleTestCase):
("eval_flag_on", "eval", None, True, True),
("desktop_keeps_bundled_skills_even_with_flag_on", None, None, True, False),
("web_phai_flag_off", "posthog_ai", None, False, False),
+ ("caller_disables_skills", None, "signal_report", False, True, True),
("web_phai_task_without_interaction_origin_flag_on", None, "posthog_ai", True, True),
("web_phai_task_without_interaction_origin_flag_off", None, "posthog_ai", False, False),
]
)
- def test_strips_bundled_skills_only_for_learn_capable_runs_with_the_flag_on(
+ def test_strips_bundled_skills_for_caller_or_learn_capable_runs(
self,
_name: str,
interaction_origin: str | None,
origin_product: str | None,
flag_enabled: bool,
expect_stripped: bool,
+ disable_skills: bool = False,
) -> None:
ctx = MagicMock(
interaction_origin=interaction_origin,
origin_product=origin_product,
organization_id="org-1",
distinct_id="user-1",
+ state={"disable_skills": disable_skills},
)
with patch(
"products.tasks.backend.temporal.process_task.utils.is_mcp_exec_skills_enabled", return_value=flag_enabled
@@ -1931,7 +1934,7 @@ class TestMcpExecSkillsEnvVars(SimpleTestCase):
env = mcp_exec_skills_env_vars(ctx)
assert env == ({"POSTHOG_CODE_DISABLE_BUNDLED_SKILLS": "1"} if expect_stripped else {})
- if interaction_origin is None and origin_product is None:
+ if disable_skills or (interaction_origin is None and origin_product is None):
flag_check.assert_not_called()
else:
flag_check.assert_called_once_with("org-1", "user-1")
diff --git a/products/tasks/backend/temporal/process_task/utils.py b/products/tasks/backend/temporal/process_task/utils.py
index 6470dec813bc..9dcfa27476ad 100644
--- a/products/tasks/backend/temporal/process_task/utils.py
+++ b/products/tasks/backend/temporal/process_task/utils.py
@@ -811,6 +811,8 @@ def mcp_exec_skills_env_vars(ctx: TaskProcessingContext) -> dict[str, str]:
Desktop runs keep their bundled skills: the MCP server excludes the `posthog-code`
consumer from `learn`, so stripping them there would leave the agent with no skills.
"""
+ if (ctx.state or {}).get("disable_skills") is True:
+ return {ENV_DISABLE_BUNDLED_SKILLS: "1"}
if _resolve_mcp_consumer(ctx.interaction_origin, origin_product=ctx.origin_product) == "posthog-code":
return {}
if not is_mcp_exec_skills_enabled(ctx.organization_id, ctx.distinct_id):
diff --git a/products/tasks/backend/tests/test_multi_turn_session.py b/products/tasks/backend/tests/test_multi_turn_session.py
index 630efd3876d4..f65c4e81c971 100644
--- a/products/tasks/backend/tests/test_multi_turn_session.py
+++ b/products/tasks/backend/tests/test_multi_turn_session.py
@@ -2007,11 +2007,11 @@ class TestCreateTaskAndTriggerForwardsContext:
assert [written[entry["storage_path"]] for entry in manifest.values()] == [b"Review the PR.", "Ü".encode()]
@pytest.mark.asyncio
- async def test_codex_options_reach_the_state_the_agent_server_reads(self) -> None:
+ async def test_caller_options_reach_the_state_the_agent_server_reads(self) -> None:
# The agent server sees only allowlisted state keys, so a dropped key silently keeps the Codex default.
team, user = await sync_to_async(self._setup_team_and_user)()
context = CustomPromptSandboxContext(
- team_id=team.id, user_id=user.id, codex_max_subagents=9, codex_shared_prompt_cache=True
+ team_id=team.id, user_id=user.id, codex_max_subagents=9, codex_shared_prompt_cache=True, disable_skills=True
)
with patch("products.tasks.backend.temporal.client.execute_task_processing_workflow"):
@@ -2021,6 +2021,7 @@ class TestCreateTaskAndTriggerForwardsContext:
public_state = _public_task_run_state(persisted.state)
assert public_state["codex_max_subagents"] == 9
assert public_state["codex_shared_prompt_cache"] is True
+ assert public_state["disable_skills"] is True
@pytest.mark.asyncio
@pytest.mark.parametrize(("output_schema", "expected_flag"), [({"type": "object"}, True), (None, None)])
diff --git a/products/tasks/backend/tests/test_repo_selection_prompt.py b/products/tasks/backend/tests/test_repo_selection_prompt.py
index e0cf35cd62d8..43d0941f814c 100644
--- a/products/tasks/backend/tests/test_repo_selection_prompt.py
+++ b/products/tasks/backend/tests/test_repo_selection_prompt.py
@@ -120,6 +120,7 @@ def test_select_repository_renders_team_rules_and_visibility_into_prompt() -> No
candidate_repos=["acme/a", "acme/b"],
)
+ assert start.call_args.kwargs["context"].disable_skills is True
assert selected.repository == "acme/b"
prompt = start.call_args.kwargs["prompt"]
assert "Team routing rules" in prompt