fix(render): skill-start keeps the env session kind under --render; name omitted sections; tighter tests (GST-825 review)

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
This commit is contained in:
garrytanandcapy-ai[bot] committed 2026-11-01 17:57:53 +00:00
1 parent 10d41ed398
commit 18c739eee1
6 files changed
+54 -29

No files matched your search

+11 -5
View File
@@ -81,11 +81,17 @@ Terminal-Bench 2 exposure counts):
later reviews reuse it), and every /ship step its receipt reads: review-log
writes, Step 11.5, Step 20 and `gstack-ship-receipt write` and `result`.
With `--render unattended`, `gstack-skill-start` prints `RENDER_KIND:
unattended` and runs the session as unattended before any side effect. If the
environment said otherwise it also prints `RENDER_KIND_MISMATCH: env=<kind>;
this render runs as unattended`. An eval harness gives each trial its own
`GSTACK_STATE_ROOT`, because Scope Lock's freeze state lives there.
The render's rules also name the interactive sections it leaves out, so a step
that still mentions one ("run Telemetry") skips that part. With `--render
unattended`, `gstack-skill-start` prints `RENDER_KIND: unattended`. The session
kind still comes from the environment alone, because every later helper
(`gstack-review-log`, `gstack-brain-enqueue`, `gstack-ship-receipt`) reads it
there, so when the environment is not unattended it prints
`RENDER_KIND_MISMATCH: env=<kind>; launch this render with
GSTACK_SESSION_KIND=unattended` and changes nothing else. An eval harness sets
`GSTACK_SESSION_KIND=unattended` on every arm, control included, so the arms
differ only by the render, and gives each trial its own `GSTACK_STATE_ROOT`,
because Scope Lock's freeze state lives there.
## The terminal line