Based on PR #15 (fix/hermes-result-decomposition-reliability); stacked
on the decomposed cli_lane modules.
- cli_lane_quota: soft-exclude a provider from NEW cli-auto work below
the remaining-quota threshold (both-below prefers more remaining;
fetch failure fails open with a metric).
- cli_lane_health: lane now writes provider health (G7) with classified
failure reasons splitting the capacity conflation (quota/auth/
rate-limit/transport) and cooldown hysteresis; re-admission only on
full cooldown expiry, passed quota reset, or fresh success (G4).
- cli_lane_routing: capacity-limited health now excludes a provider
(G3); cooldown/reset-aware re-admission.
- cli_lane_failover: explicit cli-codex-*/cli-claude-* assignees fail
closed as transient instead of switching providers (G5); fallback
depth stays bounded at two hosted providers (G1) with effort
preserved; Switchyard outages block transient, not capability (G9).
- cli_lane_metrics: route-decision/fallback counters, quota and
soft-exclusion gauges, pod-local scrape server (G6).
- cli_lane_provider: worker env drops ANTHROPIC_API_KEY, CLAUDE_API_KEY,
OPENAI_API_KEY, API_SERVER_KEY so no metered path exists (G10).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Add kanban routing keys (provider_quota_min_remaining_percent: 15,
provider_capacity_cooldown_seconds: 300, provider_auth_cooldown_seconds:
3600), a lane-metrics port/Service on 9011 with service-annotation
scraping, monitoring ingress for the new port, and the lane's quota
metrics URL env. Based on PR #15 (fix/hermes-result-decomposition-
reliability); stacked because this work rides on the decomposed lane.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Expand sudoers User_Alias chains so aliases, wildcards, netgroups, and
undefined names cannot smuggle authority to the Hermes account, and
require polkit grants to scope through exact literal identity
comparisons: computed strings, bracket lookups, subject aliasing,
operator-built values, and unconditional or wildcard grants fail closed.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Inflate every receive-pack object under strict pack, size, and checksum
bounds, resolve deltas against in-pack bases only, and scan the real
decompressed payloads for runtime-token forms, private keys, SSH key
material, and known provider token formats. Thin packs are rejected so
no pushed content escapes the scan, and upstream Git exchanges now run
under one absolute stream deadline.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Run every Gitea API and broker control exchange inside a killable helper
process whose connect, send, and read share one absolute wall-clock
deadline, and add a watchdog that force-closes streaming connections at
expiry. Redirects are rejected before authentication headers can move.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Replace the reimplemented gobwas/glob matcher with exact-literal rule
matching that fails closed on any special or malformed pattern, accepts
legacy zero priorities, and rejects duplicate primary-branch rules.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The coverage check read only line-rate, so a file could pass with weak
branch coverage. Load both Cobertura rates and fail any tracked file
below the 95% floor on either metric, failing closed when branch
evidence is absent. Prove enforcement end to end with a synthetic
fully-lined but branch-weak file failing run_profile.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Exercise the mailu sync retry, attribute, and skip branches, the
listener non-object JSON path, and the hygiene conftest skip; drop the
unreachable inverted-range clamp in the semgrep report (the line helper
already floors the end line) and pin that behavior with a test. Exclude
the mailu __main__ guards from measurement.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Cover the uncovered branches in the image release script (redirect
handler, policy-read credential and size bounds, count and shape
rejections, EOF manifest scan) and in the Harbor immutability helper
(scope-failure verification, absent robot duration, stale-read retries,
created-rule ID binding, bounded startup retries). Exclude the
__main__ guard like the sibling release script.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Repairs the blockers from the independent review of the previous head. The
role-aware verdict contract was correct but reachable only through one call
site and only on cards written with real newlines, so most real review cards
never used it.
* The role-blind call is no longer a weaker classifier that can reject before
the role-aware judge runs. Without a card there is no defensible
role-dependent judgement, so `unfinished_result_reason()` applies only the
card-independent checks. That closes the short-circuit at every call site,
including the one PR15 moves to `cli_lane_execution`, and it makes a
journalled terminal record accepted under one version of these semantics
re-validate under any other instead of being quarantined into a re-dispatch
of an already-accepted task.
* The lane resolves the role from the card and passes it, and records it in
the Kanban metadata. The verdict contract binds only where the lane can buy
another turn: in single-shot mode a rejection discards the worker's real
result, so review cards keep the relaxation without gaining any rejection
single-shot mode did not already have.
* Card scope expands the literal \n escapes the board stores in one-line
bodies, so explicit `Hermes-Task-Role` / `Hermes-Expected-Output` directives
are honoured on the 6 of 78 live cards that carry no real newline, and the
read-only, verdict and mutation heuristics stop being cut apart by them.
* Card scope now ends at the first non-card H2 and at the runner's controller
evidence, which is emitted under its own heading. Goal-controller rejection
history can no longer sit inside the card, and an upstream heading rename
fails closed instead of admitting history into role resolution.
* The inference recognises the SHIP/BLOCK-shaped deliverables real cards
actually use: 21 of 78 live cards resolve to review, up from 10, with no
implementation card misclassified. Cards asking for a findings list rather
than a verdict deliberately stay on the model judge, since the verdict is
the review contract's only gate.
* A declared review role no longer outranks mutation evidence: a report that
changed files falls back to the implementation regime.
* Judge reasons go through the agent runtime's canonical redactor, extended
for the two shapes it deliberately passes through and this lane handles -
`scheme://user:secret@host` and a credential named in prose - while a 40-hex
commit SHA survives as evidence.
Regressions cover the recovered t_dbdcd739 incident, a verbatim snapshot of
every live card on three boards with a hand-labelled expected role, the
upgrade and single-shot properties against the previous gate, the upstream
context-heading contract, and the end-to-end `execute_claim` shape that used
to burn every goal turn.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Document what an operator needs to know about the repairs: why the ordinal
workspace claim is ReadWriteMany and why colocation is only a preference, that
`lease_failed` is retryable rather than terminal and what a persistent deferred
park means, why a retry publishes an attempt-scoped ref and what to expect from
the extra drafts, and why the best-effort pool is deliberately absent from the
hermes Kustomization health checks.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Independent review t_5975c06a blocked this branch on a P1: a Kanban write that
failed while a lease expired left a `lease_failed` row that was invisible to
every pass, immortal to garbage collection, and fatal to the coordinator. It
poisoned `reconcile()` forever with a conflicting-duplicate primary key,
produced a spurious capability `block_task` from `dispatch()`, and -- because
startup maintenance ran unguarded before the port bound, against a store on a
PVC -- crash-looped the coordinator with no automatic recovery.
`lease_failed` is now a retryable state that every maintenance pass drains, and
a row only reaches a terminal state on authoritative evidence about its exact
Kanban run, so nothing is collected before its outcome is known and nothing is
silently dropped. Each row, task, and board is processed in isolation, and a
coordinator-side fault is never converted into a Kanban mutation. Startup runs
through the same guarded cycle as the steady-state loop.
The wire protocol and the durable store are now separate modules, and the
maintenance passes moved out of the coordinator, so each file stays under the
managed line ceiling with room for the recovery logic.
Also closes three consequential handoff risks the same review raised:
* mediator-N pinned itself hard to worker-N while sharing a ReadWriteOnce
claim, so a drain or preemption that moved only the lower-priority worker
deadlocked the ordinal on Multi-Attach until an operator deleted a Pod. The
shared workspace is now ReadWriteMany (as the hermes-chat tenant workspaces
already are on the same class), colocation is a preference, and the mediator
shares the worker's preemption priority, so each Pod reschedules on its own.
* the broker permits only branch creation, so a retry that added commits could
never submit and the run's work was discarded with the failure. Submission
now targets a fresh attempt- or content-scoped ref in the same reviewed
namespace -- never an update -- and is idempotent under replay. A refused
submission downgrades the result and says why instead of unwinding the run.
* the provider CLIs were reinstalled into an emptyDir on every Pod start inside
the 10m Flux health window for the whole hermes app. They now install once
per pinned version onto a durable volume, re-verified against the real
binaries and time-bounded, and the best-effort pool no longer gates the
health of the app its dependents wait on.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Three fenced worker Pods claim Hermes Kanban runs through a coordinator that
owns every state transition, with per-ordinal HMAC authority, a mediated
broker-only SCM path, and durable per-ordinal workspaces.
Content is the reviewed head of PR #18 (689bcb6e) with PR 16's and PR 19's
contributions removed: they were merged in only to validate co-existence and are
not prerequisites, so this branch no longer carries them as ancestors. Only PR 14
and PR 15 remain, because the broker boundary and the cli_lane_* decomposition
are load-bearing for two of the fixed P0 boundaries.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The distributed worker pool stacks on exactly two open pull requests, in this
order: PR 14 supplies the broker-only SCM boundary the mediators route through
and the hermes-scm-boundary-v2 ConfigMap they mount, and PR 15 supplies the
cli_lane_* decomposition -- including canonical_run_id and the eligibility
predicate on claim_ready -- that the coordinator depends on. Neither can be
dropped without breaking a fixed P0 boundary, so both are carried here as
prerequisites and this branch must not merge before them.
PR 16 (agent image release lane) and PR 19 (full-handoff acceptance harness)
are NOT prerequisites and are deliberately absent, so reviewing this branch no
longer means approving them.
PR 14 and PR 15 conflict with each other in nine paths. Each is resolved to the
resolution already reviewed on this branch at 4d4cf1bd.
Refresh #12 and #18 to their live Gitea heads, record #19's replaced head
exactly, and replace the pairwise conflict set with a freshly computed one
over all 28 combinations. #18+#19 is now clean because #19 no longer
duplicates the repo-wide hygiene legacy-exception mechanism that #18 also
carries; #12+#19 is new because #12 advanced. Name the exact overlapping
files per conflicting pair and how to resolve each additively.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
evaluate_names_absent returned PASS when its step exited 0 with no output,
so five mandatory checks - the ones asserting that provider API keys, forge
credentials, a cluster-admin binding, and shared coordinator state are
absent - could report a pass on no evidence and turn a NO_GO into a GO.
Both name rules now resolve their step through one guard in _line_step, so
zero observations are NOT_RUN. Regressions pin all five real catalog specs
plus both reachable silence paths: a POSIX pipeline whose status comes from
its last stage, and a drifted kubectl -o jsonpath. The pool claim projection
emits one <volume>=<claim> line per template volume so a volume without a
PVC still counts as an observation rather than reading as drift.
Also closes the review's reachable hardening and evidence defects:
- pin Gitea paths to atlas/titan-iac on an exact segment boundary and
reject relative segments, including percent-encoded ones
- forbid impersonation structurally in every mode and vantage; the inner
command of kubectl exec is re-checked rather than exempted, and
validate_catalog no longer guards only the operator vantage
- drop flux and helm from the binary allowlist; they had no pinned release
digest, so no allowlisted binary can now be admitted that the executor
would refuse to attest
- remove the inert --concurrency and --expect-telegram-sessions flags and
the dead concurrency bound; Telegram continuity stays mandatory
- read the ephemeral pull index page by page, treat the create response as
an authoritative source for the pull number, close every number either
source names, and surface residue_ref plus exact manual_cleanup commands
when creation is uncertain
- keep executable_path and executable_sha256 on unrecorded bulk-evidence
steps so withholding bytes never withholds binary attestation
- revert the repo-wide hygiene legacy-exception mechanism; the contract
change here is purely additive and the four pre-existing over-cap files
are left to the canonical contract change in PR #14/#15
- correct the runbook ruff format scope so the documented command passes
Split hermes_handoff_arming.py out of hermes_handoff_ephemeral.py to keep
both modules under the 500-line cap. All 16 handoff modules hold at least
95% line and branch coverage; the mutation gate is 13/13.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
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>
