- serial-linux-selfhosted now fetches depth 2 and passes
DSH_ARCHIVE_BASE_REF=github.event.before, running the same
frozen-archive comparison as serial-linux instead of diffing the
new manifest against itself.
- Runbook (both languages): documents the deliberate dependabot
exception (queued-on-hosted during failover is expected, not a
failed switch); corrects the emergency-capacity bootstrap to
exclude .runner/.credentials when cloning a runner directory; and
replaces the 'admin-only' variable claim with the accurate
trust-model statement — repository variables are writer-manageable,
which in this private fork-disabled repo with an all-workflows
runner group is routing among members, not an escalation.
Static gate green locally: 32 passed, 0 failed.
- All four failover selectors (three workers + the verdict job) and the
paired env/cache expressions now exclude dependabot[bot]: under
failover, dependency-supplied code keeps queueing for the hosted pool
instead of executing on the persistent VM. A delayed Dependabot PR
during an outage is an acceptable cost; dependency code on the
privileged host is not.
- Runbook (both languages): records the shipped failover bounds
(coverage 8, snapshots 12, sized for six instances) and documents
that the verdict job follows the selector too — operators previously
had no explanation for a verdict queued after all workers passed.
- Local static gate green: 32 passed, 0 failed (translation pairing
519 pairs consistent).
Review fix (ds-review-bot on #623): the per-session-modes section still
described first-turn stamping — superseded by the header baseline — and its
formula lacked the middle rung. The formula now reads own post-seed knob
events ?? inherited SessionHeader baseline ?? composition default, and the
paragraph describes the synchronous capture into the creation-time header
with no first-turn timing window. Chinese counterpart updated in the same
commit.
The pairing gate requires link target #9 to be byte-identical between
the language sides; my earlier 'fix' pointed the zh side at the zh
runbook and broke the contract. Reverted to the shared target and
re-recorded the pairing hash.
The local PTY readiness poll held its inferred_idle fallback for exactly
one pollIntervalMs after a prompt marker, so a bash foreground handoff
that lands on the silence boundary only wins the exact stdin_read
attribution when the kernel publishes it inside that single poll. On a
slow or loaded host it does not, and the attribution flips.
handoffGraceMs replaces the hardcoded one-poll window as a validated,
deployment-owned config field defaulting to 500ms, rejected at load when
it cannot contain one readiness poll. Real-shell tests that interrupt a
send now assert the session is usable again rather than which readiness
tier observed the handoff, because no fixed grace removes the race.
`config.maxWallMs` is only checked for positivity, and it is handed to
`setTimeout`, which clamps any delay above 2^31-1 ms to 1 ms. A deployment
configuring a 25-day wall ceiling therefore gets the opposite of what it asked
for: every run times out on the first tick. The runtime now range-checks the
field at load against MAX_TIMER_DELAY_MS from dsh-timeout and throws, so the
misconfiguration fails loud where it is self-contained instead of silently
inverting the budget.
`computeMs` needs no matching bound: it is compared against measured event-loop
utilization rather than fed to a timer.
The test asserts both the rejection and that the boundary value itself loads.
The darwin-parity job failed intermittently on the SIGINT test with the
operation buffer holding only the echoed command line, never RAW_READY. The
harness sets idleSilenceMs to 250, so when a cold python3 start stays silent
past that bound the send settles as inferred_idle; PtySendOperation.append then
drops all later output, and the marker reaches only the scrollback.
Give the harness per-test idleSilenceMs/timeoutMs overrides and let this
scenario raise both above interpreter startup latency, so the readiness marker
lands inside the send it belongs to. waitForOutput's own deadline and the test
timeout grow to match the new bounds.
The product timings are unchanged; the pty Agent Note records why a test that
waits on an operation must outlast the child's startup.
The required windows job moves from windows-2025 to ubuntu-latest, running
checksum-verified Windows Node under Wine at Linux-job wall clock (2m46s
warm vs 7-9min); master's serial-windows native-kernel reference is
untouched, and a new master-only wine-apt-cache job seeds the apt cache
every pull request restores. The experiment workflow folds into ci.yml,
the Agent Note moves to implemented with measured results, and the two CI
topology notes update to the shipped facts.
Review follow-up on the symmetric-cache conversion: provisioning and
caching are separable concerns, so the remaining 8 corepack enable
sites (the three enterprise restore-only PR jobs, the required Windows
job, serial-macos/serial-windows, the cache-less ci.yml jobs, and
sandbox.yml) now provision pnpm via pnpm/action-setup@v4 too — only
the provisioning line moves; every deliberate cache shape (enterprise
restore-only pairing, serial-linux producer, Windows cache-less) is
byte-preserved. corepack is gone from CI entirely.
Agent Note retitled to the decision's real shape (provisioning policy,
with caching as per-job policy on top), EN+ZH, pair re-recorded;
composite-action and stop-at-cache-workflows alternatives recorded.
Review fix (ds-review-bot on #623): the ACP scenario runs at deployment
strength (the automation protocol has no session-scoped switch), so the
assembled-app path could not detect the delegation bypass itself. The new
keyless subagent-inheritance headless scenario closes that on the
semantic-checkpoint precedent: a seeded parent log carrying a real
sandbox/mode: read-only switch under a workspace-write deployment default
is resumed through the Loader-booted cli-demo app via a resume fixture
plugin and delegates through the real subagent tool; the child's real
write is denied by the real dsh-fs-sandbox fence (physical ENOENT
assertion), its persisted header carries the inherited baseline, and both
logs pin as expected outputs. Verified red: disabling the driver's capture
makes the scenario fail on the disk assertion (the child writes under the
deployment default).
The spare tier is retired. Steady-state pool load is one serial standby
job per master push, so six always-on instances already are the
failover capacity; pre-registered offline runners are a silently
expiring guarantee (GitHub garbage-collects them after 30 days
offline). Incident-time extra capacity is a one-minute org-token
registration, now documented in the runbook.
Two review findings on the turndown swap, both verified empirically:
- Unclosed-tag nesting makes the synchronous turndown/domino walk
superlinear (measured: depth 512 ~0.15s, 2k ~2s, 20k ~5s), during
which the cooperative fetchTimeoutMs timer cannot fire. renderBody
now preflights nesting depth with a linear tag scan and passes
bodies past 512 levels through raw; the try/catch stays for markup
the scan cannot see (comment-hidden tags), simulated in tests via a
converter throw.
- Markdown escaping can expand converted HTML ~2x (100k underscores
render as 200k chars), so provider body caps no longer bounded the
model-visible result. formatFetchOutput now caps the complete output
(header + body + footer) under new fetchMaxOutputChars config
(default 200000 = 2x the local provider's default body cap), reusing
the truncation notice.
README EN+ZH, config catalog, Agent Note EN+ZH updated; the new
web-fetch fixture is migrated to the packed layout master now
requires; tool-web coverage stays 100% per-file.
