The local goal judge scored every worker report against "did the reviewed
implementation reach a shippable state". A read-only reviewer that returned a
completed BLOCK verdict with findings was therefore resumed turn after turn with
an instruction to repair code it was forbidden to touch (observed live on
t_dbdcd739), burning subscription capacity and risking an unbounded loop.
Completion is now judged against the action the card assigned:
* Cards declare their role explicitly with Hermes-Task-Role / Hermes-Expected-
Output metadata. Pre-contract cards fall back to a narrow inference that needs
a read-only scope, a requested verdict, no requested mutation deliverable, and
a report that changed no files.
* Role resolution reads only the card itself. Prior attempts, parent results,
cross-task history and comments appended to the worker context can no longer
reassign the role.
* Review, audit and diagnostic cards finalize deterministically on a truthful
SHIP or BLOCK verdict with evidence, and fail closed on a missing,
unrecognized or self-contradictory verdict, on a BLOCK without findings, and
on a verdict without evidence. Every rejection reason carries the read-only
guard, so a resumed review is never told to edit the reviewed implementation.
* Implementation cards keep the fail-closed model judge unchanged, including the
unfinished-work heuristic and judge-unavailable rejection.
* All judge reasons are bounded, single-line and secret-redacted before they
reach Kanban metadata, comments and continuation prompts.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Decides whether the Hermes platform handoff is fit to release, and refuses
to round an absence of evidence up to a pass.
The harness is read-only by default and classifies 71 checks PASS / FAIL /
NOT_RUN / NOT_APPLICABLE. Any mandatory FAIL or NOT_RUN is NO_GO, and so is a
harness-level problem: an unreachable vantage, a catalog entry whose evidence
no longer exists, an expired deadline, or an evaluator that raised.
Evidence comes from two vantages that cannot cover for each other: an external
read-only operator kubeconfig, and the Hermes agent probing itself from inside
its own pod. Before any check runs, the harness asks each vantage who it is and
stops if they are the same principal, because dual-vantage evidence from one
identity is a restatement rather than a corroboration. `--as` is rejected for
every operator-side command and reachable only as the inner command of a
`kubectl exec`, so impersonation can never stand in for a real self-probe. A
deny check needs a live refused request, not only an authorization review.
Two safety properties are structural rather than conventional, enforced where
an argv becomes a subprocess: the default mode mutates nothing (mutating verbs
require a server dry run; there is deliberately no live TokenRequest probe,
because a successful one would mint a real credential), and no probe can pull a
credential value into a report (no vault/sops/curl, secrets readable only with
-o name, environment probes list names, shell only through frozen reviewed
templates). Captures are bounded before they are screened, and the rendered
report is re-screened before it is written.
Mutation lives behind a separate arming flag with an exact confirmation phrase,
a caller-supplied unique ref, a preflight that refuses a protected push target
before any network call, and a cleanup whose verification is itself mandatory.
A default run reports those four checks NOT_RUN.
The catalog is declarative so a reviewer reads what is asserted rather than how
it is plumbed, and so structural properties can be proven over every entry
before a run. Catalog drift surfaces as NOT_RUN, which stops the release.
docs/hermes_full_handoff_acceptance.md carries the merge order for PRs #14-#18
on top of the merged #13 baseline, the image build and Flux rollout, the
rollback point for each step, the go/no-go checklist, and the limits that are
asserted rather than exercised.
Validation: 295 handoff tests pass with 100% line coverage on all 15 new
modules; the full unit suite is 647 passed with two failures that reproduce
unchanged on origin/main; Ruff, py_compile, kustomize render, and a diff
credential screen are clean; a live read-only run against Atlas returns NO_GO
for the pre-merge cluster with no unscreened fields in the report.
The sticky-block gate added in the previous commit classifies a task from
the `created` event payload that upstream `create_task` writes. That
producer is code we do not own, so trusting it silently was the gap: if
upstream renamed the key, dropped it, or stopped deriving it from
`initial_status`, the image would still build and ship a consumer that
mis-classifies every task it reads.
Anchor the producer contract at build time, before the regression suite
runs, with three assert-only preconditions: the `initial_status="blocked"`
park resolves `task_status` to `"blocked"`, every non-park creation
resolves it to something else, and the `created` event carries that same
variable under `"status"`. None of them rewrite the producer.
Textual anchors cannot see dataflow, so add the runtime net the reviewer
asked for. The suite now drives the real API: create + claim an ordinary
task, trip the circuit breaker once at failure_limit=1 so it parks with a
`gave_up` event (leaving its own `created` event as the most recent
create/block/unblock row), then recompute at failure_limit=2 and require
promotion to ready. That case is red under an unconditional-true created
predicate and red under producer drift that labels every created event
blocked, while the explicit block/unblock, dependency-promotion and
circuit-breaker-at-current-limit cases stay green. Non-blocked and
malformed created payloads are pinned as controls, and the gate now
rejects non-dict payloads rather than trusting `.get`.
Also make the live placement correction durable: titan-04 is cordoned
after repeated kernel undervoltage and kubelet failure and titan-19 was
probe/Longhorn unstable under worker load, so both join the hard NotIn
list; titan-05 is healthy but sits at 3592m/3600m requested CPU, so the
main hermes container gives back 50m (350m -> 300m) to schedule there.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Three narrowly scoped Hermes reliability fixes backed by live evidence
from the Cassandra/titan-iac proof run.
Worker concurrency. Three simultaneous direct CLI workers on the 4-core
hermes-agent node drove load to ~45 and made the hermes and oauth2-proxy
containers fail their probes, leaving the pod 8/10 Ready; two workers
stayed at 10/10. Cap HERMES_CLI_LANE_CONCURRENCY at 2 and lower the
cli-lane-runner CPU limit from 3 to 2 so the dashboard and auth sidecars
keep a guaranteed share of the node. Requests are unchanged: the pod
still asks for 745m total, so placement does not move.
Service links. Kubernetes injects a service-link variable pair for every
service in the namespace, and hermes-claude-broker produces
HERMES_CLAUDE_BROKER_PORT=tcp://10.43.31.76:9006 — a value the broker
parses as an int. That contaminated worker and test environments even
though the deployment already addresses every service by DNS name. Set
enableServiceLinks: false on the hermes-agent pod spec.
Blocked-task scheduling. create_task(initial_status="blocked") records a
created event carrying status=blocked but never a blocked event, while
_has_sticky_block() only inspects blocked/unblocked events. recompute_ready()
considers blocked tasks, so an explicitly parked task with no incomplete
parent auto-promoted on the next dispatcher cycle. Teach _has_sticky_block()
to also recognize a created event whose payload status is blocked, which
covers tasks created before this image patch without adding a persisted
field. Dependency-driven promotion and the circuit-breaker failure-limit
guard are untouched; unblock_task() still releases either kind of block.
hermes-kanban-blocked-regression.py runs against the real upstream
kanban_db API during the image build, so the build fails if any of these
semantics regress.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
