atlas-iac/testing/tests/test_hermes_handoff_ephemeral.py

439 lines
16 KiB
Python
Raw Normal View History

"""Contracts for the strictly bounded, self-cleaning armed mode."""
hermes: add a fail-closed full-handoff acceptance harness 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.
2026-08-17 10:14:17 +00:00
from __future__ import annotations
import json
import subprocess
from pathlib import Path
hermes: add a fail-closed full-handoff acceptance harness 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.
2026-08-17 10:14:17 +00:00
import pytest
from testing.tests.test_hermes_handoff_support import (
FakeClock,
FakeSpawn,
load_handoff_module,
)
ephemeral = load_handoff_module("hermes_handoff_ephemeral")
fix(hermes): close the zero-evidence fail-open in absence checks 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>
2026-08-17 16:24:37 +00:00
arming = load_handoff_module("hermes_handoff_arming")
hermes: add a fail-closed full-handoff acceptance harness 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.
2026-08-17 10:14:17 +00:00
exec_module = load_handoff_module("hermes_handoff_exec")
policy = load_handoff_module("hermes_handoff_policy")
model = load_handoff_module("hermes_handoff_model")
TOKEN = "acceptance-20260817a"
HEAD = "8f005458282269ba5c07941814e4237f5d4cf3ac"
MAIN = "1" * 40
hermes: add a fail-closed full-handoff acceptance harness 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.
2026-08-17 10:14:17 +00:00
VANTAGE = exec_module.operator_vantage()
def request(**overrides):
hermes: add a fail-closed full-handoff acceptance harness 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.
2026-08-17 10:14:17 +00:00
fields = {
"repo": policy.EXPECTED_REPO,
"remote": policy.EXPECTED_REMOTE,
hermes: add a fail-closed full-handoff acceptance harness 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.
2026-08-17 10:14:17 +00:00
"token": TOKEN,
"confirmation": ephemeral.CONFIRMATION,
"base": policy.EXPECTED_BASE,
"expected_head": HEAD,
"expected_base_sha": MAIN,
hermes: add a fail-closed full-handoff acceptance harness 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.
2026-08-17 10:14:17 +00:00
}
fields.update(overrides)
return ephemeral.ArmRequest(**fields)
def completed(stdout: str = "", stderr: str = "", returncode: int = 0):
return subprocess.CompletedProcess(
args=[], returncode=returncode, stdout=stdout, stderr=stderr
)
hermes: add a fail-closed full-handoff acceptance harness 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.
2026-08-17 10:14:17 +00:00
def armed_runner(results: list, mode=policy.ARMED):
hermes: add a fail-closed full-handoff acceptance harness 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.
2026-08-17 10:14:17 +00:00
spawn = FakeSpawn(results)
runner = exec_module.Runner(
mode=mode,
clock=FakeClock(),
spawn=spawn,
environ={},
deadline_seconds=3600,
attestor=lambda command, _environment: exec_module.Attestation(
command, "a" * 64
),
hermes: add a fail-closed full-handoff acceptance harness 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.
2026-08-17 10:14:17 +00:00
)
return runner, spawn
hermes: add a fail-closed full-handoff acceptance harness 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.
2026-08-17 10:14:17 +00:00
def by_id(results: list) -> dict:
return {result.spec.id: result for result in results}
def pr_payload(state="open", number=42):
return {
"number": number,
"state": state,
"draft": True,
"merged": False,
"base": {"ref": "main", "sha": MAIN},
"head": {"ref": request().ref, "sha": HEAD},
}
def linked_worktree(tmp_path: Path) -> Path:
root = tmp_path / "worktree"
common = tmp_path / "repo.git"
gitdir = common / "worktrees" / "acceptance"
gitdir.mkdir(parents=True)
(root).mkdir()
(root / ".git").write_text(f"gitdir: {gitdir}\n", encoding="utf-8")
(gitdir / "commondir").write_text("../..\n", encoding="utf-8")
(gitdir / "HEAD").write_text(
fix(hermes): close the zero-evidence fail-open in absence checks 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>
2026-08-17 16:24:37 +00:00
f"ref: refs/heads/{arming.EXPECTED_HEAD_REF}\n", encoding="utf-8"
)
fix(hermes): close the zero-evidence fail-open in absence checks 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>
2026-08-17 16:24:37 +00:00
ref = common / "refs" / "heads" / arming.EXPECTED_HEAD_REF
ref.parent.mkdir(parents=True)
ref.write_text(f"{HEAD}\n", encoding="utf-8")
remote_ref = common / "refs" / "remotes" / "origin" / "main"
remote_ref.parent.mkdir(parents=True)
remote_ref.write_text(f"{MAIN}\n", encoding="utf-8")
(common / "config").write_text(
fix(hermes): close the zero-evidence fail-open in absence checks 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>
2026-08-17 16:24:37 +00:00
f'[remote "origin"]\n\turl = {arming.EXPECTED_REMOTE_URL}\n',
encoding="utf-8",
)
return root
hermes: add a fail-closed full-handoff acceptance harness 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.
2026-08-17 10:14:17 +00:00
@pytest.mark.parametrize(
"ref", ["main", "master", "refs/heads/main", "HEAD", "feature/x", ""]
hermes: add a fail-closed full-handoff acceptance harness 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.
2026-08-17 10:14:17 +00:00
)
def test_only_the_unique_ephemeral_ref_is_allowed(ref: str) -> None:
hermes: add a fail-closed full-handoff acceptance harness 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.
2026-08-17 10:14:17 +00:00
with pytest.raises(ephemeral.ArmingError):
ephemeral.assert_push_target_allowed(ref)
ephemeral.assert_push_target_allowed(request().ref)
hermes: add a fail-closed full-handoff acceptance harness 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.
2026-08-17 10:14:17 +00:00
@pytest.mark.parametrize(
"overrides",
hermes: add a fail-closed full-handoff acceptance harness 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.
2026-08-17 10:14:17 +00:00
[
{"confirmation": "wrong"},
{"repo": "atlas/other"},
{"remote": "upstream"},
{"base": "develop"},
{"token": "short"},
{"expected_head": "bad"},
{"expected_base_sha": "bad"},
hermes: add a fail-closed full-handoff acceptance harness 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.
2026-08-17 10:14:17 +00:00
],
)
def test_preflight_rejects_bad_inputs_before_any_runner_exists(
tmp_path: Path, overrides: dict
) -> None:
with pytest.raises(ephemeral.ArmingError):
ephemeral.preflight(
request(**overrides), policy.EXPECTED_REPO, linked_worktree(tmp_path)
)
hermes: add a fail-closed full-handoff acceptance harness 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.
2026-08-17 10:14:17 +00:00
def test_preflight_attests_the_actual_linked_worktree(tmp_path: Path) -> None:
ephemeral.preflight(request(), policy.EXPECTED_REPO, linked_worktree(tmp_path))
root = linked_worktree(tmp_path / "wrong")
(Path((root / ".git").read_text().split(": ", 1)[1].strip()) / "HEAD").write_text(
"ref: refs/heads/other\n", encoding="utf-8"
)
with pytest.raises(ephemeral.ArmingError, match="existing PR #19"):
ephemeral.preflight(request(), policy.EXPECTED_REPO, root)
def test_worktree_attestation_rejects_each_malformed_local_source(
tmp_path: Path,
) -> None:
roots = []
empty = tmp_path / "empty"
empty.mkdir()
roots.append(empty)
bad_marker = linked_worktree(tmp_path / "marker")
(bad_marker / ".git").write_text("not-a-gitdir\n", encoding="utf-8")
roots.append(bad_marker)
wrong_sha = linked_worktree(tmp_path / "sha")
gitdir = Path((wrong_sha / ".git").read_text().split(": ", 1)[1].strip())
common = (gitdir / (gitdir / "commondir").read_text().strip()).resolve()
fix(hermes): close the zero-evidence fail-open in absence checks 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>
2026-08-17 16:24:37 +00:00
(common / "refs" / "heads" / arming.EXPECTED_HEAD_REF).write_text("0" * 40)
roots.append(wrong_sha)
wrong_main = linked_worktree(tmp_path / "main")
gitdir = Path((wrong_main / ".git").read_text().split(": ", 1)[1].strip())
common = (gitdir / (gitdir / "commondir").read_text().strip()).resolve()
(common / "refs" / "remotes" / "origin" / "main").write_text("0" * 40)
roots.append(wrong_main)
bad_config = linked_worktree(tmp_path / "config")
gitdir = Path((bad_config / ".git").read_text().split(": ", 1)[1].strip())
common = (gitdir / (gitdir / "commondir").read_text().strip()).resolve()
(common / "config").write_text("[broken\n", encoding="utf-8")
roots.append(bad_config)
wrong_remote = linked_worktree(tmp_path / "remote")
gitdir = Path((wrong_remote / ".git").read_text().split(": ", 1)[1].strip())
common = (gitdir / (gitdir / "commondir").read_text().strip()).resolve()
(common / "config").write_text('[remote "origin"]\nurl=https://example.invalid/x\n')
roots.append(wrong_remote)
for root in roots:
with pytest.raises((ephemeral.ArmingError, FileNotFoundError)):
ephemeral.preflight(request(), policy.EXPECTED_REPO, root)
oversized = tmp_path / "oversized"
oversized.write_text("x" * 5, encoding="utf-8")
with pytest.raises(ephemeral.ArmingError):
fix(hermes): close the zero-evidence fail-open in absence checks 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>
2026-08-17 16:24:37 +00:00
ephemeral.bounded_read(oversized, 4)
hermes: add a fail-closed full-handoff acceptance harness 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.
2026-08-17 10:14:17 +00:00
def test_guard_failure_stops_before_runner_calls(monkeypatch) -> None:
hermes: add a fail-closed full-handoff acceptance harness 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.
2026-08-17 10:14:17 +00:00
monkeypatch.setattr(ephemeral, "assert_push_target_allowed", lambda _ref: None)
runner, spawn = armed_runner([])
results = by_id(ephemeral.run_armed(runner, VANTAGE, request()))
assert results["ephemeral.protected-branch-refusal"].status == model.FAIL
assert all(
result.status in {model.FAIL, model.NOT_RUN} for result in results.values()
)
assert spawn.calls == []
hermes: add a fail-closed full-handoff acceptance harness 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.
2026-08-17 10:14:17 +00:00
def test_full_armed_run_validates_and_removes_exact_artifacts() -> None:
ref_line = f"{HEAD}\trefs/heads/{request().ref}\n"
hermes: add a fail-closed full-handoff acceptance harness 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.
2026-08-17 10:14:17 +00:00
runner, spawn = armed_runner(
[
completed(),
completed(stdout=f"{HEAD}\n"),
completed(),
completed(stdout=ref_line),
completed(stdout=json.dumps(pr_payload())),
completed(stdout=json.dumps([pr_payload()])),
completed(stdout=json.dumps(pr_payload(state="closed"))),
completed(),
completed(),
completed(stdout=json.dumps(pr_payload(state="closed"))),
hermes: add a fail-closed full-handoff acceptance harness 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.
2026-08-17 10:14:17 +00:00
]
)
results = by_id(ephemeral.run_armed(runner, VANTAGE, request()))
assert all(result.status == model.PASS for result in results.values())
push = next(call for call in spawn.calls if call["argv"][1] == "push")
assert push["argv"][-1] == f"HEAD:refs/heads/{request().ref}"
hermes: add a fail-closed full-handoff acceptance harness 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.
2026-08-17 10:14:17 +00:00
assert all(result.spec.mandatory for result in results.values())
def test_existing_or_unverifiable_ref_never_reaches_a_write() -> None:
for first in (completed(stdout="abc\trefs/heads/x\n"), completed(returncode=128)):
runner, spawn = armed_runner([first])
results = by_id(ephemeral.run_armed(runner, VANTAGE, request()))
assert results["ephemeral.feature-branch-push"].status == model.FAIL
assert not any(
call["argv"][1] in {"push", "POST", "PATCH", "DELETE"}
for call in spawn.calls
)
hermes: add a fail-closed full-handoff acceptance harness 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.
2026-08-17 10:14:17 +00:00
def test_failed_or_ambiguous_create_is_discovered_and_cleanup_still_runs() -> None:
ref_line = f"{HEAD}\trefs/heads/{request().ref}\n"
hermes: add a fail-closed full-handoff acceptance harness 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.
2026-08-17 10:14:17 +00:00
runner, _ = armed_runner(
[
completed(),
completed(stdout=f"{HEAD}\n"),
completed(),
completed(stdout=ref_line),
completed(returncode=1),
completed(stdout=json.dumps([pr_payload()])),
completed(stdout=json.dumps(pr_payload(state="closed"))),
completed(),
completed(),
completed(stdout=json.dumps(pr_payload(state="closed"))),
hermes: add a fail-closed full-handoff acceptance harness 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.
2026-08-17 10:14:17 +00:00
]
)
results = by_id(ephemeral.run_armed(runner, VANTAGE, request()))
assert results["ephemeral.draft-pull-request"].status == model.PASS
hermes: add a fail-closed full-handoff acceptance harness 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.
2026-08-17 10:14:17 +00:00
assert results["ephemeral.cleanup-verified"].status == model.PASS
def test_pull_request_parsers_reject_every_malformed_or_ambiguous_shape() -> None:
assert ephemeral._pr_problems([], request()) == ["pull request is not an object"]
malformed = pr_payload()
malformed.update({"number": True, "draft": "yes", "base": None})
malformed["head"] = {"ref": "other", "sha": "bad"}
problems = ephemeral._pr_problems(malformed, request())
assert {"number is malformed", "draft does not match", "base is malformed"} <= set(
problems
)
fix(hermes): close the zero-evidence fail-open in absence checks 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>
2026-08-17 16:24:37 +00:00
assert ephemeral._created_number(
exec_module.Outcome((), "x", returncode=1), request()
) == (0, "pull-request creation failed: exit status 1")
for body in ("not-json", "{}"):
number, error = ephemeral._created_number(
exec_module.Outcome((), "x", 0, body), request()
)
fix(hermes): close the zero-evidence fail-open in absence checks 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>
2026-08-17 16:24:37 +00:00
assert number == 0 and error
assert ephemeral._created_number(
exec_module.Outcome((), "x", 0, json.dumps(pr_payload())), request()
) == (42, "")
for rows in ([None], [{"head": {"ref": "other"}, "base": {}}]):
numbers, error = ephemeral._match_pulls(rows, request())
assert numbers == [] and error
invalid = pr_payload()
invalid["draft"] = False
fix(hermes): close the zero-evidence fail-open in absence checks 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>
2026-08-17 16:24:37 +00:00
assert ephemeral._match_pulls([invalid], request())[1]
assert "discovered 0" in ephemeral._match_pulls([], request())[1]
hermes: add a fail-closed full-handoff acceptance harness 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.
2026-08-17 10:14:17 +00:00
@pytest.mark.parametrize(
"delete_result,remaining,closed",
[
(completed(returncode=1), completed(), pr_payload(state="closed")),
(completed(), completed(returncode=128), pr_payload(state="closed")),
(completed(), completed(), {}),
(completed(), completed(), pr_payload(state="open")),
],
)
def test_cleanup_fails_closed_on_partial_or_malformed_evidence(
delete_result, remaining, closed
) -> None:
ref_line = f"{HEAD}\trefs/heads/{request().ref}\n"
hermes: add a fail-closed full-handoff acceptance harness 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.
2026-08-17 10:14:17 +00:00
runner, _ = armed_runner(
[
completed(),
completed(stdout=f"{HEAD}\n"),
completed(),
completed(stdout=ref_line),
completed(),
completed(stdout=json.dumps([pr_payload()])),
completed(stdout=json.dumps(pr_payload(state="closed"))),
delete_result,
remaining,
completed(stdout=json.dumps(closed)),
hermes: add a fail-closed full-handoff acceptance harness 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.
2026-08-17 10:14:17 +00:00
]
)
cleanup = by_id(ephemeral.run_armed(runner, VANTAGE, request()))[
"ephemeral.cleanup-verified"
]
hermes: add a fail-closed full-handoff acceptance harness 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.
2026-08-17 10:14:17 +00:00
assert cleanup.status == model.FAIL
def test_cleanup_fails_when_close_or_verification_commands_are_bad_json_or_failures() -> (
None
):
api = "/api/v1/repos/atlas/titan-iac"
for scripted in (
[completed(returncode=1), completed(), completed(), completed(returncode=1)],
[completed(stdout="bad"), completed(), completed(), completed(stdout="bad")],
hermes: add a fail-closed full-handoff acceptance harness 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.
2026-08-17 10:14:17 +00:00
[
completed(stdout=json.dumps(pr_payload())),
completed(),
completed(),
completed(stdout=json.dumps(pr_payload(state="closed"))),
],
):
runner, _ = armed_runner(scripted)
result = ephemeral._cleanup(runner, VANTAGE, request(), api, [42])
assert result.status == model.FAIL
def test_read_only_runner_cannot_reach_push() -> None:
runner, spawn = armed_runner(
[completed(), completed(stdout=f"{HEAD}\n"), completed(), completed()],
mode=policy.READ_ONLY,
hermes: add a fail-closed full-handoff acceptance harness 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.
2026-08-17 10:14:17 +00:00
)
results = by_id(ephemeral.run_armed(runner, VANTAGE, request()))
assert results["ephemeral.feature-branch-push"].status == model.FAIL
assert not any(call["argv"][1] == "push" for call in spawn.calls)
fix(hermes): close the zero-evidence fail-open in absence checks 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>
2026-08-17 16:24:37 +00:00
def push_prelude(ref_line: str) -> list:
"""The four scripted results that get an armed run to a verified push."""
return [
completed(),
completed(stdout=f"{HEAD}\n"),
completed(),
completed(stdout=ref_line),
]
def test_pull_request_discovery_reads_every_page() -> None:
"""A full first page must not end discovery: the match can be on page two."""
ref_line = f"{HEAD}\trefs/heads/{request().ref}\n"
filler = [
{
"number": index,
"state": "open",
"draft": True,
"merged": False,
"base": {"ref": "main", "sha": MAIN},
"head": {"ref": "feature/other", "sha": HEAD},
}
for index in range(1, ephemeral.DISCOVERY_PAGE_SIZE + 1)
]
runner, spawn = armed_runner(
[
*push_prelude(ref_line),
completed(stdout=json.dumps(pr_payload())),
completed(stdout=json.dumps(filler)),
completed(stdout=json.dumps([pr_payload()])),
completed(stdout=json.dumps(pr_payload(state="closed"))),
completed(),
completed(),
completed(stdout=json.dumps(pr_payload(state="closed"))),
]
)
results = by_id(ephemeral.run_armed(runner, VANTAGE, request()))
assert results["ephemeral.draft-pull-request"].status == model.PASS
pages = [call["argv"][2] for call in spawn.calls if "pulls?state=all" in str(call)]
assert any("page=1" in page for page in pages)
assert any("page=2" in page for page in pages)
def test_uncertain_discovery_still_closes_the_created_pull_and_names_residue() -> None:
"""The create response names the pull, so an unusable index cannot orphan it."""
ref_line = f"{HEAD}\trefs/heads/{request().ref}\n"
runner, spawn = armed_runner(
[
*push_prelude(ref_line),
completed(stdout=json.dumps(pr_payload())),
completed(returncode=1),
completed(stdout=json.dumps(pr_payload(state="closed"))),
completed(),
completed(),
completed(stdout=json.dumps(pr_payload(state="closed"))),
]
)
results = by_id(ephemeral.run_armed(runner, VANTAGE, request()))
draft = results["ephemeral.draft-pull-request"]
assert draft.status == model.FAIL
assert draft.evidence["pull_requests"] == [42]
assert draft.evidence["residue_ref"] == request().ref
assert any("pulls/42" in line for line in draft.evidence["manual_cleanup"])
assert any("branches/" in line for line in draft.evidence["manual_cleanup"])
# The pull the harness created is closed, not left behind.
assert any(
call["argv"][1] == "PATCH" and call["argv"][2].endswith("/pulls/42")
for call in spawn.calls
)
assert results["ephemeral.cleanup-verified"].status == model.FAIL
def test_a_wholly_unidentifiable_create_names_the_ref_for_manual_cleanup() -> None:
"""When neither source names a pull, the operator gets the exact ref to sweep."""
ref_line = f"{HEAD}\trefs/heads/{request().ref}\n"
runner, _ = armed_runner(
[
*push_prelude(ref_line),
completed(stdout="not-json"),
completed(returncode=1),
completed(),
completed(),
]
)
draft = by_id(ephemeral.run_armed(runner, VANTAGE, request()))[
"ephemeral.draft-pull-request"
]
assert draft.status == model.FAIL
assert draft.evidence["pull_requests"] == []
assert draft.evidence["residue_ref"] == request().ref
assert any(
request().ref in line and "close by hand" in line
for line in draft.evidence["manual_cleanup"]
)