fix(agent-session): fail closed on recovered identity

This commit is contained in:
Brennan Benson
2026-08-07 14:58:41 -07:00
parent b82f73a260
commit d21382999f
10 changed files with 152 additions and 26 deletions
@@ -88,6 +88,20 @@ describe('owner identity probe', () => {
).resolves.toEqual({ outcome: 'identity-mismatch', field: 'process-start-time' })
})
it('fails closed when an exact token and the reconstructed start time disagree', async () => {
await expect(
probeAgentSessionProcessIdentity({
identity: IDENTITY,
deps: deps({
readProcessStartTimeMs: async () => START_TIME + PROCESS_START_TIME_TOLERANCE_MS + 1
})
})
).resolves.toEqual({
outcome: 'indeterminate',
reason: 'process identity evidence contradicted'
})
})
it('tolerates start-time jitter inside the tolerance', async () => {
await expect(
probeAgentSessionProcessIdentity({
@@ -45,19 +45,18 @@ function defaultIsPidPresent(pid: number): boolean {
async function readLinuxProcessStartTimeMs(pid: number): Promise<number | null> {
try {
const [stat, uptime] = await Promise.all([
const [stat, systemStat] = await Promise.all([
readFile(`/proc/${pid}/stat`, 'utf-8'),
readFile('/proc/uptime', 'utf-8')
readFile('/proc/stat', 'utf-8')
])
// Field 22 is starttime in clock ticks; the comm field can contain spaces, so cut past ") ".
const fields = stat.slice(stat.lastIndexOf(') ') + 2).split(' ')
const ticks = Number(fields[19])
const uptimeSeconds = Number(uptime.split(' ')[0])
if (!Number.isFinite(ticks) || !Number.isFinite(uptimeSeconds)) {
const bootTimeSeconds = Number(/^btime\s+(\d+)$/m.exec(systemStat)?.[1])
if (!Number.isFinite(ticks) || !Number.isFinite(bootTimeSeconds)) {
return null
}
const bootTimeMs = Date.now() - uptimeSeconds * 1000
return Math.round(bootTimeMs + (ticks / 100) * 1000)
return Math.round(bootTimeSeconds * 1000 + (ticks / 100) * 1000)
} catch {
return null
}
@@ -125,6 +124,10 @@ export async function probeAgentSessionProcessIdentity(args: {
)
if (observed !== null) {
if (Math.abs(observed - identity.processStartTimeMs) > PROCESS_START_TIME_TOLERANCE_MS) {
if (matchedOn.includes('spawn-token')) {
// Why: contradictory evidence cannot prove that a token-authenticated child is dead.
return { outcome: 'indeterminate', reason: 'process identity evidence contradicted' }
}
return { outcome: 'identity-mismatch', field: 'process-start-time' }
}
matchedOn.push('process-start-time')
@@ -36,6 +36,7 @@ export type AgentSessionStoreState = {
export type LoadedAgentSessionStore = {
state: AgentSessionStoreState
storeFound: boolean
/** True when the file was written by a newer schema; this host reads but never writes it. */
readOnly: boolean
/** True when the primary file was unusable and the previous committed copy was used. */
@@ -184,6 +185,7 @@ export async function loadAgentSessionStore(
}
return {
state: parsed.state,
storeFound: true,
readOnly: parsed.state.schemaVersion > AGENT_SESSION_STORE_SCHEMA_VERSION,
recoveredFromBackup
}
@@ -191,7 +193,12 @@ export async function loadAgentSessionStore(
if (unusableStoreFound) {
throw new Error('agent_session_store_corrupt')
}
return { state: emptyState(hostId), readOnly: false, recoveredFromBackup: false }
return {
state: emptyState(hostId),
storeFound: false,
readOnly: false,
recoveredFromBackup: false
}
}
function serializeState(state: AgentSessionStoreState): string {
@@ -356,9 +356,8 @@ describe('concurrent claims', () => {
const store = await open()
await establishOwner(store)
const before = store.getRecord('session-alpha')
// The next commit cannot land: the store directory is no longer a directory.
// Losing both committed copies must not reset the live store to empty authority.
await rm(directory, { recursive: true, force: true })
await writeFile(directory, 'not a directory')
await expect(
store.setJournalCheckpoint({
sessionId: 'session-alpha',
@@ -829,7 +828,7 @@ describe('orphans, claim keys, checkpoints, and unreadable rows', () => {
expect(reopened.recoveredFromBackup).toBe(true)
expect(reopened.getRecord('session-alpha')?.lease.runtimeFence).toBe(1)
await reopened.retireClaimKey('key-2', NOW)
await expect(reopened.retireClaimKey('key-2', NOW)).rejects.toThrow()
const backup = JSON.parse(await readFile(`${agentSessionStorePath(directory)}.bak`, 'utf-8'))
expect(backup.records['session-alpha'].lease.runtimeFence).toBe(1)
})
@@ -80,6 +80,7 @@ export class AgentSessionRecordStore {
args.hostId,
loaded.readOnly,
loaded.recoveredFromBackup,
loaded.storeFound,
loaded.state,
diskRevision
)
@@ -53,6 +53,7 @@ export class AgentSessionStoreTransactionQueue {
readonly hostId: string,
readonly readOnly: boolean,
readonly recoveredFromBackup: boolean,
private diskStoreFound: boolean,
public state: AgentSessionStoreState,
private diskRevision: string
) {
@@ -66,6 +67,10 @@ export class AgentSessionStoreTransactionQueue {
throw new Error('agent_session_legacy_required')
}
await this.refreshExternallyChangedState()
if (this.diskRecoveredFromBackup) {
// Why: the missing commit may hold a newer fence, so rollback cannot mint another writer.
throw new Error('execution_owner_reconciling')
}
const records = new Map(this.state.records)
const operations = new Map(this.state.operations)
const retiredClaimKeys = [...this.state.retiredClaimKeys]
@@ -80,6 +85,7 @@ export class AgentSessionStoreTransactionQueue {
this.state.schemaVersion = AGENT_SESSION_STORE_SCHEMA_VERSION
this.diskRevision = agentSessionStoreRevision(this.state)
this.diskRecoveredFromBackup = false
this.diskStoreFound = true
return result
} catch (error) {
this.state.records = records
@@ -95,6 +101,10 @@ export class AgentSessionStoreTransactionQueue {
private async refreshExternallyChangedState(): Promise<void> {
const loaded = await loadAgentSessionStore(this.filePath, this.hostId)
if (this.diskStoreFound && !loaded.storeFound) {
throw new Error('agent_session_store_corrupt')
}
this.diskStoreFound ||= loaded.storeFound
const diskRevision = agentSessionStoreRevision(loaded.state)
this.diskRecoveredFromBackup = loaded.recoveredFromBackup
if (diskRevision === this.diskRevision) {
@@ -168,6 +168,7 @@ describe('persisted row validation', () => {
expect(isAgentSessionOperationRow({ ...row, callerKey: '' })).toBe(false)
expect(isAgentSessionOperationRow({ ...row, expiresAt: 1.5 })).toBe(false)
expect(isAgentSessionOperationRow({ ...row, outcome: { status: 'succeeded' } })).toBe(false)
expect(isAgentSessionOperationRow({ ...row, outcome: null })).toBe(false)
expect(isAgentSessionOperationRow(null)).toBe(false)
})
})
+2 -1
View File
@@ -151,7 +151,8 @@ export function isAgentSessionOperationRow(value: unknown): value is AgentSessio
const row = value as Partial<AgentSessionOperationRow>
const outcome = row.outcome as AgentSessionOperationOutcome | undefined
const outcomeValid =
outcome !== undefined &&
typeof outcome === 'object' &&
outcome !== null &&
((outcome.status === 'pending' && true) ||
(outcome.status === 'succeeded' && typeof outcome.sessionId === 'string') ||
(outcome.status === 'failed' && typeof outcome.code === 'string') ||
@@ -43,8 +43,8 @@ describe('handle identity', () => {
it('keys a Codex handle by thread id alone', () => {
const codex: AgentSessionProviderHandle = { provider: 'codex', threadId: 'thread-1' }
expect(agentSessionProviderHandleKey(codex)).toBe('codex:thread-1')
expect(agentSessionProviderHandleRoot(codex)).toBe('codex:thread-1')
expect(agentSessionProviderHandleKey(codex)).toBe('codex:"thread-1"')
expect(agentSessionProviderHandleRoot(codex)).toBe('codex:"thread-1"')
expect(
agentSessionProviderHandlesEqual(codex, { provider: 'codex', threadId: 'thread-2' })
).toBe(false)
@@ -57,6 +57,21 @@ describe('handle identity', () => {
expect(isAgentSessionProviderHandle({ provider: 'gemini', sessionId: 'x' })).toBe(false)
expect(isAgentSessionProviderHandle({ ...CLAUDE, sessionId: ' sess-1 ' })).toBe(false)
})
it('uses collision-free keys when Claude ids contain delimiters', () => {
const left: AgentSessionProviderHandle = {
provider: 'claude',
sessionId: 'a#b',
leafUuid: 'c'
}
const right: AgentSessionProviderHandle = {
provider: 'claude',
sessionId: 'a',
leafUuid: 'b#c'
}
expect(agentSessionProviderHandleKey(left)).not.toBe(agentSessionProviderHandleKey(right))
expect(agentSessionProviderHandlesEqual(left, right)).toBe(false)
})
})
describe('chain append', () => {
@@ -153,6 +168,19 @@ describe('chain append', () => {
).toHaveLength(2)
})
it('rejects reuse of a stable link id for a different proof', () => {
expect(() =>
appendAgentSessionProviderHandleLink(
[link()],
link({
origin: 'resumed',
handle: { ...CLAUDE, leafUuid: 'leaf-2' },
mintedAtFence: 2
})
)
).toThrow('agent_session_provider_handle_invalid')
})
it('refuses to grow past the cap rather than dropping fork provenance', () => {
const chain: AgentSessionProviderHandleLink[] = [link()]
for (let index = 1; index < MAX_AGENT_SESSION_PROVIDER_HANDLE_LINKS; index += 1) {
@@ -218,6 +246,9 @@ describe('chain lookup and validation', () => {
it('rejects a persisted chain that is over the cap or holds a malformed link', () => {
expect(isAgentSessionProviderHandleChain([{ ...link(), mintedAtFence: -1 }])).toBe(false)
expect(isAgentSessionProviderHandleChain([{ ...link(), linkId: 'not a link id!' }])).toBe(false)
expect(isAgentSessionProviderHandleChain([{ ...link(), forkedFromKey: 'claude:seed' }])).toBe(
false
)
expect(
isAgentSessionProviderHandleChain(
Array.from({ length: MAX_AGENT_SESSION_PROVIDER_HANDLE_LINKS + 1 }, (_value, index) =>
@@ -226,4 +257,38 @@ describe('chain lookup and validation', () => {
)
).toBe(false)
})
it('rejects persisted chains that bypass append invariants', () => {
expect(
isAgentSessionProviderHandleChain([
link(),
link({
linkId: 'link-2',
origin: 'created',
mintedAtFence: 2
})
])
).toBe(false)
expect(
isAgentSessionProviderHandleChain([
link(),
link({
origin: 'resumed',
handle: { ...CLAUDE, leafUuid: 'leaf-2' },
mintedAtFence: 2
})
])
).toBe(false)
expect(
isAgentSessionProviderHandleChain([
link(),
link({
linkId: 'link-2',
origin: 'resumed',
handle: { provider: 'claude', sessionId: 'sess-2', leafUuid: 'leaf-2' },
mintedAtFence: 2
})
])
).toBe(false)
})
})
+38 -13
View File
@@ -64,8 +64,8 @@ export function isAgentSessionProviderHandle(value: unknown): value is AgentSess
/** Stable string identity for one handle. Two handles with the same key name the same writer target. */
export function agentSessionProviderHandleKey(handle: AgentSessionProviderHandle): string {
return handle.provider === 'claude'
? `claude:${handle.sessionId}#${handle.leafUuid ?? ''}`
: `codex:${handle.threadId}`
? `claude:${JSON.stringify([handle.sessionId, handle.leafUuid])}`
: `codex:${JSON.stringify(handle.threadId)}`
}
/**
@@ -73,7 +73,9 @@ export function agentSessionProviderHandleKey(handle: AgentSessionProviderHandle
* whatever the provider called it.
*/
export function agentSessionProviderHandleRoot(handle: AgentSessionProviderHandle): string {
return handle.provider === 'claude' ? `claude:${handle.sessionId}` : `codex:${handle.threadId}`
return handle.provider === 'claude'
? `claude:${JSON.stringify(handle.sessionId)}`
: `codex:${JSON.stringify(handle.threadId)}`
}
export function agentSessionProviderHandlesEqual(
@@ -103,29 +105,48 @@ export function isAgentSessionProviderHandleLink(
return false
}
const link = value as Partial<AgentSessionProviderHandleLink>
const originValid =
link.origin === 'created' ||
link.origin === 'adopted' ||
link.origin === 'resumed' ||
link.origin === 'forked'
return (
typeof link.linkId === 'string' &&
LINK_ID_PATTERN.test(link.linkId) &&
isAgentSessionProviderHandle(link.handle) &&
(link.origin === 'created' ||
link.origin === 'adopted' ||
link.origin === 'resumed' ||
link.origin === 'forked') &&
originValid &&
Number.isSafeInteger(link.mintedAtFence) &&
(link.mintedAtFence as number) >= 0 &&
Number.isSafeInteger(link.observedAt) &&
(link.forkedFromKey === undefined || isHandleField(link.forkedFromKey))
(link.origin === 'forked'
? isHandleField(link.forkedFromKey)
: link.forkedFromKey === undefined)
)
}
export function isAgentSessionProviderHandleChain(
value: unknown
): value is AgentSessionProviderHandleLink[] {
return (
Array.isArray(value) &&
value.length <= MAX_AGENT_SESSION_PROVIDER_HANDLE_LINKS &&
value.every((link) => isAgentSessionProviderHandleLink(link))
)
if (!Array.isArray(value) || value.length > MAX_AGENT_SESSION_PROVIDER_HANDLE_LINKS) {
return false
}
let validated: AgentSessionProviderHandleLink[] = []
try {
for (const link of value) {
if (!isAgentSessionProviderHandleLink(link)) {
return false
}
const next = appendAgentSessionProviderHandleLink(validated, link)
// A persisted chain must name every link exactly once; retry elision belongs at append time.
if (next.length !== validated.length + 1) {
return false
}
validated = next
}
return true
} catch {
return false
}
}
/**
@@ -178,6 +199,10 @@ export function appendAgentSessionProviderHandleLink(
// Why: re-proving the same handle at the same fence is a retry, not a new identity.
return [...chain]
}
if (findAgentSessionProviderHandleLink(chain, link.linkId)) {
// Why: the lease names its exact proof by link id; reuse would make that reference ambiguous.
throw new Error('agent_session_provider_handle_invalid')
}
if (chain.length >= MAX_AGENT_SESSION_PROVIDER_HANDLE_LINKS) {
// Why: dropping older links would erase fork provenance, so refuse and let the caller roll
// the journal epoch instead of silently losing where this conversation came from.