Policy Hot-Reload¶

Document status: Implemented (option A, signing key). Revocation is deliberately out of scope; see the limit stated below
Applies to: cMCP Runtime gateway (PolicyStore, startup)
Related config: policy_reload_interval_seconds

Summary¶

Hot-reload was never missing. It was implemented in PolicyStore.reload_if_stale, wired into PolicyEvaluator, and documented as a supported knob — and it could not swap a policy in any production configuration. The sections below record why, with the measurements, because the shape of that mistake is worth keeping.

Now: policy can change at runtime when the gateway pins a signing key rather than an artifact hash (option A, built). The guaranteed-inert configuration is refused at startup, a failing reload no longer re-reads the bundle on every request, and a signed bundle whose version increases is installed without a restart. Revocation is not implemented and that limit is stated rather than implied.

The rest of this document is the analysis that got there, kept because the diagnosis matters more than the fix: a status file said "not yet" while the code said "implemented and inert", and nothing failed.

What is actually there¶

PolicyStore (in policy/bundle.py) holds the active bundle behind an RLock. PolicyEvaluator.evaluate calls reload_if_stale() on the way in. Once the interval has elapsed, the store re-reads the bundle from disk and swaps it in if the hash changed, keeping the current bundle if anything raises.

That is a reasonable poll-and-swap design. The problem is the trust anchor it re-uses.

The defect: the pin and the reload contradict each other¶

At startup the gateway requires CMCP_POLICY_HASH unless CMCP_DEV_MODE=1 (POLICY-001/#137: without a pinned hash, a tampered bundle loads silently). That pinned hash is handed to PolicyStore as expected_hash and then re-used on every reload:

new_bundle = load_policy_bundle(self._bundle_path, self._expected_hash)

load_policy_bundle raises PolicyHashMismatch when the bundle on disk does not match the hash it was given. So the reload path asks a question that answers itself: "has the bundle changed, and does it still hash to the value it had before it changed?" A bundle that has genuinely changed always fails. A bundle that has not changed always passes and swaps nothing.

The line that would install a new bundle:

if new_bundle.bundle_hash != self._bundle.bundle_hash:

is unreachable whenever expected_hash is set, because the call above it has already raised.

Measured, not inferred. With expected_hash set to the startup hash, an operator edit from allow-all to deny-all produces:

startup bundle hash : sha256:fdb2e39e...
new on-disk hash    : sha256:e1426d71...
reload_if_stale()   : False
active bundle hash  : sha256:fdb2e39e...   <-- unchanged

The gateway keeps enforcing allow-all. The operator's deny-all edit does not take effect, and the only signal is a WARNING.

Why no test caught it¶

Every reload test in tests/unit/test_policy_bundle.py constructs PolicyStore without expected_hash:

store = PolicyStore(bundle=old_bundle, bundle_path=str(bundle_dir),
                    reload_interval_seconds=1)

That is the dev-mode configuration, where reload does work. test_policy_store_bundle_swap_on_hash_change passes and proves the swap logic is correct — in the one configuration production never uses. The pinned-hash case, which is the only case a deployment runs, is untested.

Second defect: it re-reads the bundle on every request, forever¶

reload_if_stale advances _last_reload_at only on the success path. The failure path returns without touching it, so the staleness check stays true and the next call tries again immediately. Since the reload in production always fails, the interval stops being an interval.

Measured: 50 policy evaluations after the first interval elapsed produce 50 full bundle re-reads from disk, each one reading every policy file and computing a SHA-256 over the canonical bundle, on the request hot path.

So the production effect of setting policy_reload_interval_seconds: 60 is not "policy updates every minute". It is "after one minute, every tool call performs full bundle I/O and hashing, and the policy never changes". That is a self-inflicted load amplifier on the enforcement path, reached by following the documented configuration.

Third, smaller: the error text is from the wrong lifecycle¶

The swallowed exception logs Policy bundle hash mismatch: gateway will not start during steady-state operation. The gateway is already running and is not going to stop. except Exception also catches genuine faults (unreadable file, malformed Cedar, permissions) and files them all under the same warning, so the log cannot distinguish "operator edited the policy and the pin now disagrees" from "the disk is failing".

The real design question¶

Hot-reload and hash pinning are not accidentally in tension — they want different things:

• Pinning a hash says the policy is exactly this artifact, decided before the process started, and nothing may change it afterwards. That is what makes a policy bundle attestable: the hash goes into the TRACE claim, and a verifier can check that the gateway enforced the bundle it said it did.
• Hot-reload says the policy may change while the process runs.

A design that wants both must answer: when a new bundle arrives, what authorises it? A hash fixed at startup cannot, by construction. Something else has to.

There is a second question that follows immediately and matters just as much for cMCP specifically: what does a reload do to evidence? A TRACE claim names the policy bundle hash the call was evaluated under. If the bundle can change mid-process, then claims from one process carry different bundle hashes, and every consumer of those claims needs to cope with that. Any option below has to say what the claim records and how a verifier reconstructs which policy was live for a given call.

Options¶

A. Pin a signing key, not a bundle hash¶

The bundle carries a signature over its own manifest; the gateway pins the public key allowed to sign policy. A reload verifies the new bundle's signature rather than its hash.

• Trust moves from "this exact artifact" to "any artifact this authority approves", which is what actually makes runtime change safe.
• Fits the existing manifest (author_identity, commit_sha are already there, unsigned) and the direction the rest of the stack has taken.
• Cost: key management, revocation, and a rollback story — a validly signed older bundle is a downgrade attack unless the manifest carries a version that must increase monotonically.
• Evidence: the claim records the bundle hash and the signer plus the bundle version, so a verifier can check both what ran and who authorised it.

B. Pin a set of acceptable hashes¶

CMCP_POLICY_HASH becomes a list. A reload accepts any bundle whose hash is in the allowlist.

• Smallest change from what exists, and keeps the "exactly these artifacts" model.
• No new cryptography and nothing to revoke.
• Cost: every policy change still needs the operator to restart to extend the allowlist, so it does not deliver hot-reload — it only lets a fleet roll between a known set of policies without restart. Useful for staged rollout and fast rollback; not an answer to "we need to tighten a policy right now".

C. Re-read the pin from a trusted source at reload time¶

The expected hash comes from somewhere the gateway can re-consult — a separate hash file, a control plane, a transparency log — instead of a startup env var.

• Genuine hot-reload, and the authority for a policy change stays outside the gateway.
• Composes with the transparency direction: the pin could be an entry a verifier can independently look up.
• Cost: introduces a runtime dependency on that source and its own trust question. If the hash file sits next to the bundle and is writable by whoever writes the bundle, it authorises nothing.

D. Operator-triggered reload that supplies the new hash¶

No polling. An admin action (signal, authenticated endpoint) hands the gateway the new expected hash and it reloads once.

• Nothing changes under a running request without a human or a deploy pipeline asking, which is the most predictable behaviour and the easiest to audit.
• Removes the hot-path reload_if_stale() call from evaluate entirely, and with it the load amplifier.
• Cost: a new authenticated control surface on the gateway, which is attack surface on the enforcement component. Less convenient than polling.

E. Keep it development-only, and make that honest¶

Accept that a pinned, attestable policy and runtime mutation do not belong in the same deployment. Reload is permitted only when CMCP_DEV_MODE=1; configuring policy_reload_interval_seconds > 0 together with a pinned hash is a startup config error rather than a warning at request time.

• Honest, costs nothing to build, and removes the failure mode entirely.
• Matches how the rest of the runtime treats this class of thing: fail at construction on a configuration that could never work.
• Cost: cMCP keeps telling operators that a policy change needs an enclave restart, which for a confidential gateway means an attestation cycle. That is a real operational burden and the reason this item is on the list at all.

Decision (2026-08-10): option A, pin a signing key¶

Direction chosen: A. Runtime policy change gets its authority from a pinned signing key rather than a pinned artifact hash. Nothing about it is implemented yet; what follows is what the decision commits us to, so the implementation issue starts from a settled model rather than reopening the choice.

What it means concretely:

• The bundle manifest gains a signature over its own canonical contents. The manifest already carries author_identity and commit_sha, both currently unsigned assertions; signing turns them into claims someone is accountable for.
• The gateway pins a public key (a new setting, not CMCP_POLICY_HASH). A reload verifies the new bundle's signature against it.
• The manifest's version must increase monotonically across reloads, and a bundle whose version does not is refused. Without this a validly signed older bundle is a downgrade attack: an attacker who can write the bundle directory replays yesterday's more permissive policy, and every signature still checks out.
• CMCP_POLICY_HASH keeps its current meaning and stays the right choice for a deployment that wants exactly one policy for the life of the process. It is a pin on an artifact, so it remains incompatible with reload — the two are alternatives, not layers, and configuring both is refused at startup (see below).
• Evidence: a TRACE claim records the bundle hash it evaluated under plus the signer identity and the bundle version, so a verifier can answer both "what policy ran" and "who authorised it" for a process whose policy changed mid-life.

As built¶

export CMCP_POLICY_SIGNING_KEY=<raw Ed25519 public key, base64url or hex>
# and in cmcp-config.yaml
policy_reload_interval_seconds: 60

Why the version check is not optional¶

Without it the signing-key model is a downgrade attack. Anyone who can write the bundle directory replays yesterday's more permissive bundle: the authority really signed it, the signature really verifies, and the gateway installs a policy the operator already retired. Monotonicity is what makes "signed by the authority" mean "the authority's current intent". test_a_replayed_older_signed_bundle_is_refused constructs exactly that attack and fails if the check is removed, which was verified by removing it.

The limit: no revocation¶

A compromised signing key stops being trusted by changing CMCP_POLICY_SIGNING_KEY and restarting. There is no revocation list, no key set, and no expiry.

Stated plainly because it bounds what this buys: faster policy change, not faster key change. A deployment whose threat model includes a compromised policy signing key needing revocation inside a fleet-restart window is not served by this, and should keep pinning a hash. Adding revocation later is compatible with what is built — it constrains which keys are acceptable, and does not change the signature or the version rule.

In-flight sessions¶

A session admitted under the previous bundle is evaluated against the new one from its next call. This is what an operator tightening a policy during an incident expects, and it is the reason reload was wanted at all. The cost is that a long-running session can see its effective permissions narrow with no signal; a session-facing notification was considered and is not built.

Two things that were not optional, and are now done¶

Both stood on their own, independent of the direction, and both have landed:

1. A configuration that cannot work does not start. CMCP_POLICY_HASH together with policy_reload_interval_seconds > 0 now aborts startup (POLICY_RELOAD_PINNED_HASH). Under option A the two remain alternatives rather than layers, so this refusal is durable rather than a stopgap: a pin on an exact artifact and a policy that may change are contradictory whatever the reload path is authorised by. Dev mode pins no hash and is unaffected, which is the one configuration where reload actually works and is tested as such.
2. The interval is honoured on failure. _last_reload_at is stamped before the attempt, so an exception cannot skip it. A failing reload now costs one attempt per interval instead of one full bundle read plus hash per request. Worth doing regardless of what happens to hot-reload, because the same shape reappears in anything else that polls on the enforcement path.

The test gap generalised past this feature and was closed with it: a code path whose only tests construct it in the configuration production never uses is a path with no tests. The reload tests now cover pinned and unpinned side by side, one asserting that a pinned hash cannot install a changed bundle and one asserting that without the pin it can, so the difference the pin makes is visible in the suite rather than discovered later. The load bound has a test above it that counts the reads.

Not in scope¶

• Catalog hot-reload. CMCP_CATALOG_HASH has exactly the same pin, and load_catalog the same shape, so whatever is decided here should be applied there deliberately rather than by copy. It is not analysed in this document.
• Revocation of a policy signing key, which is a real gap rather than a non-goal: see "The limit: no revocation" above.