wip(rm-03): rescue checkpoint of in-flight queue-guard work

CHECKPOINT ONLY — not a deliverable, not reviewed, not complete.

Committed by mos-remediation as TRANSPORT after the authoring seat (coder-mos1) became
unable to act: its context window was exceeded at 102.6%/372k AND its overflow-recovery
path failed ("Turn prefix summarization failed: servers overloaded"). The seat could not
process instructions or commit its own work, leaving 28 modified files durable only on
local disk.

Authorship preserved as coder-mos1 — this is transport, not authorship. The work is the
seat's; committing it is the only way to make it survive the rotation that must follow.

Scope note, unresolved: this diff spans the queue-guard tools (9 files under
framework/tools/git) AND 5 framework guides plus 11 agent templates. RM-03 was scoped to
ci-queue-wait.sh and its tests. Whether the doc/template edits are consequential to the
tri-state change or drive-by was queried and never answered — the seat was bricked before
it could reply. That question stays open and must be resolved before any of this is
reviewed; the wider edits may need splitting out.

Direct instance of D-31: commit-then-rotate works, rotate-without-commit loses everything.
This commit is contained in:
2026-08-01 13:31:44 -05:00
parent f4fd5967fc
commit b238e0a8b1
28 changed files with 995 additions and 205 deletions
@@ -13,7 +13,7 @@
# Covers:
# (a) 404 branch-absent -> exit 0, "queue clear" message.
# (b) 200 existing branch + a terminal CI state -> unchanged behavior.
# (c) genuine API error (500) -> still fail-closed (nonzero exit).
# (c) genuine API error (500) -> loud, audited CANNOT_ASSERT; degraded exit 0.
set -euo pipefail
@@ -62,7 +62,7 @@ case "$mode" in
200) code=200; body='{"commit":{"id":"deadbeefcafef00d0123456789abcdef01234567"}}' ;;
500) code=500; body='{"message":"internal server error"}' ;;
no-status) code=200; body='{}' ;;
terminal-success) code=200; body='{"state":"success"}' ;;
terminal-success) code=200; body='{"state":"success","statuses":[{"status":"success"}]}' ;;
*)
echo "curl stub: unknown mode=$mode" >&2
exit 2
@@ -91,6 +91,7 @@ run_ci_queue_wait() {
export MOSAIC_CREDENTIALS_FILE="$WORK_DIR/no-credentials.json"
export GITEA_TOKEN="stub-token"
export GITEA_URL="https://git.example.test"
export MOSAIC_CI_QUEUE_AUDIT_LOG="$WORK_DIR/audit/ci-queue-wait.jsonl"
"$SCRIPT_DIR/ci-queue-wait.sh" -B "$branch" --purpose push -t 5 -i 1
)
}
@@ -131,19 +132,26 @@ elif [[ "$out_b" == *"queue clear"* ]]; then
fail=1
fi
# (c) genuine API error (500) -> still fail-closed, exit nonzero.
# (c) genuine API error (500) -> CANNOT_ASSERT is loud and audited, but does not brick delivery.
set +e
out_c=$(MOSAIC_STUB_BRANCH_MODE=500 run_ci_queue_wait "feat/some-branch" 2>&1)
status_c=$?
set -e
if [[ "$status_c" -eq 0 ]]; then
echo "FAIL(c): expected a nonzero exit for a genuine 500 API error, got 0" >&2
if [[ "$status_c" -ne 0 ]]; then
echo "FAIL(c): expected degraded exit 0 for provider unavailability, got $status_c" >&2
echo "$out_c" >&2
fail=1
elif [[ "$out_c" != *"CANNOT_ASSERT"* ]]; then
echo "FAIL(c): expected a loud CANNOT_ASSERT diagnostic" >&2
echo "$out_c" >&2
fail=1
elif [[ "$out_c" == *"queue clear"* ]]; then
echo "FAIL(c): a genuine API error must not be reported as queue-clear" >&2
echo "$out_c" >&2
fail=1
elif [[ ! -s "$WORK_DIR/audit/ci-queue-wait.jsonl" ]]; then
echo "FAIL(c): expected a durable CANNOT_ASSERT audit record" >&2
fail=1
fi
if [[ "$fail" -eq 0 ]]; then