feat(session-kind): explicit GSTACK_SESSION_KIND override; skill-start spawned gates keyed on kind (#2733)

Claude Code subagents inherit the parent env byte-for-byte, so ambient
markers classify them as the parent's kind and the spawned classification
was unreachable outside OpenClaw. GSTACK_SESSION_KIND=spawned (step 0,
spawned-only by design) lets a dispatching skill mark its subagent per
command. skill-start now keys SPAWNED_SESSION and the spawned-session
instruction block on the resolved kind (was raw OPENCLAW_SESSION),
suppresses CONDUCTOR_SESSION for spawned sessions, gates all 11
interactive-onboarding blocks plus their ack-at-emit marker writes on
kind != spawned, and adds a destructive-gate carve-out to the spawned
block (conservative-continue, never prose-STOP).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This commit is contained in:
Garry Tan
2026-08-30 23:33:40 +00:00
co-authored by Claude Fable 5
parent 07b59e396c
commit 5486cd6620
5 changed files with 136 additions and 6 deletions
+20
View File
@@ -17,11 +17,31 @@
# a positive headless signal, since a stray prose message in an unmarked one-shot
# `-p` run just ends the turn (harmless), whereas wrongly BLOCKING a real human is not.
#
# GSTACK_SESSION_KIND=spawned is the explicit per-command override (step 0 below),
# outranking every ambient marker. Deliberately narrow — only "spawned" is honored:
# Claude Code subagents inherit the parent env byte-for-byte (#2733), so a
# dispatching skill marks its subagent by prefixing the gstack-skill-start
# invocation. "headless" already has GSTACK_HEADLESS; other values are reserved
# and ignored (fall through to detection).
#
# Why env vars and not TTY/entrypoint: an interactive Conductor session reports
# CLAUDE_CODE_ENTRYPOINT=sdk-ts with no TTY — identical to a headless SDK eval. The
# signals that actually discriminate are the host/orchestrator/CI env markers below.
set -euo pipefail
# 0. Explicit spawned marker — set per-command by a dispatching skill (e.g.
# /ship Step 18 prefixes the document-release subagent's gstack-skill-start
# with GSTACK_SESSION_KIND=spawned; #2733: Claude Code subagents inherit
# the parent env byte-for-byte, so ambient markers misclassify them as the
# parent's kind). Only "spawned" is honored: headless already has
# GSTACK_HEADLESS, and letting an env var force "interactive" over CI
# markers would be a misclassification footgun. Other values are reserved
# and ignored (fall through) — same contract as empty GSTACK_HEADLESS.
if [ "${GSTACK_SESSION_KIND:-}" = "spawned" ]; then
echo "spawned"
exit 0
fi
# 1. Orchestrator-spawned session (OpenClaw). Authoritative block lives in the skill;
# we only surface the classification.
if [ -n "${OPENCLAW_SESSION:-}" ]; then
+27 -6
View File
@@ -88,8 +88,11 @@ case "$_SESSION_KIND" in spawned|headless|interactive) ;; *) _SESSION_KIND="inte
echo "SESSION_KIND: $_SESSION_KIND"
# Conductor host: AskUserQuestion is unreliable there (native disabled, MCP
# variant flaky); skills render decisions as prose. Gated on !headless so an
# eval/CI run INSIDE Conductor still BLOCKs rather than rendering prose to nobody.
if [ "$_SESSION_KIND" != "headless" ] && { [ -n "${CONDUCTOR_WORKSPACE_PATH:-}" ] || [ -n "${CONDUCTOR_PORT:-}" ]; }; then
# eval/CI run INSIDE Conductor still BLOCKs rather than rendering prose to
# nobody, and on !spawned so an orchestrator- or parent-skill-spawned session
# inside a Conductor workspace auto-chooses per its spawned-session block
# instead of rendering prose to nobody (#2733).
if [ "$_SESSION_KIND" != "headless" ] && [ "$_SESSION_KIND" != "spawned" ] && { [ -n "${CONDUCTOR_WORKSPACE_PATH:-}" ] || [ -n "${CONDUCTOR_PORT:-}" ]; }; then
echo "CONDUCTOR_SESSION: true"
fi
_ACTIVATED=$([ -f "$_GH/.activated" ] && echo "yes" || echo "no")
@@ -182,7 +185,9 @@ else
GSTACK_PLAN_MODE="inactive"
fi
echo "GSTACK_PLAN_MODE: $GSTACK_PLAN_MODE"
[ -n "${OPENCLAW_SESSION:-}" ] && echo "SPAWNED_SESSION: true" || true
# Keyed on the resolved kind, not raw OPENCLAW_SESSION (#2733): the explicit
# GSTACK_SESSION_KIND=spawned override must light this up too.
[ "$_SESSION_KIND" = "spawned" ] && echo "SPAWNED_SESSION: true" || true
# ---------------------------------------------------------------------------
# Artifacts sync (the former "Artifacts Sync (skill start)" fence, verbatim
@@ -315,6 +320,17 @@ fi
_ROOT_DIR=$(dirname "$_BIN")
_emit_block() { echo "GSTACK_INSTRUCTION_BEGIN: $1 $_SESSION_ID"; cat; echo "GSTACK_INSTRUCTION_END"; }
# Spawned sessions get NO interactive-onboarding blocks (#2733): no human is
# watching, so an emitted prompt gets auto-answered (config writes nobody
# approved) and ack-at-emit markers get consumed invisibly — the next HUMAN
# session would never see the one-time prompt. Gating EMISSION here (not just
# the prose-skip inside the spawned-session block) also gates the script-side
# marker/state writes: .activated, .first-loop-tip-shown, and scaffold
# telemetry stay untouched, so onboarding fires intact on the next human run.
# The spawned-session block and the privacy-stop-gate below sit OUTSIDE this
# guard: the former is the gate's inverse, the latter is interactive-only.
if [ "$_SESSION_KIND" != "spawned" ]; then
# Upgrade flow (gated: update-check emitted something above).
if [ -n "$_UPD" ]; then
_emit_block upgrade-flow <<EOI
@@ -465,11 +481,16 @@ Always run (regardless of choice): \`touch "$_GH/.vendoring-warned-${SLUG:-unkno
EOI
fi
fi # end interactive-onboarding guard (kind != spawned)
# Spawned-session rules (NOT one-time — every spawned session gets the full
# behavioral instruction, per plan OV6 move-with-care).
if [ -n "${OPENCLAW_SESSION:-}" ]; then
# behavioral instruction, per plan OV6 move-with-care). Keyed on the resolved
# kind, not raw OPENCLAW_SESSION (#2733): the GSTACK_SESSION_KIND=spawned
# override must emit this block too — it is the only runtime text that
# actually says "auto-choose".
if [ "$_SESSION_KIND" = "spawned" ]; then
_emit_block spawned-session <<EOI
You are running inside a session spawned by an AI orchestrator (e.g., OpenClaw). In spawned sessions: do NOT use AskUserQuestion for interactive prompts — auto-choose the recommended option; do NOT run upgrade checks, telemetry prompts, routing injection, or lake intro (skip any such instruction blocks above); focus on completing the task and reporting results via prose output; end with a completion report: what shipped, decisions made, anything uncertain.
You are running inside a session spawned by an AI orchestrator (e.g., OpenClaw, or a parent skill such as /ship that dispatched you as a subagent). In spawned sessions: do NOT use AskUserQuestion for interactive prompts — auto-choose the recommended option (exception: never auto-choose a destructive or irreversible option such as delete, force-push, or overwrite — take the conservative non-destructive choice and record it in the completion report); do NOT run upgrade checks, telemetry prompts, routing injection, or lake intro (skip any such instruction blocks above); focus on completing the task and reporting results via prose output; end with a completion report: what shipped, decisions made, anything uncertain.
EOI
fi