[nissa-seru]

This adds a test that the framework refuses a disposed conversation agent's name for a new agent. No test checked that refusal before. It's part of what keeps a later stream from reading the leftover cancel record that #382 lists under "Not covered". Only a new agent of the same name could read that record, and once create() has returned, no agent's name goes to a second agent.

Problem

Line numbers here are at this PR's base, d1b9d35c.

#382's "Not covered" section describes a cancel record left over from an earlier agent of the same name. The framework keeps a record of each stream it cancels itself (frameworkCancelledStreams, src/framework.ts:2010), keyed by the agent's name and the stream's id, and each agent numbers its streams from 1 and only counts up (src/agent.ts:168, :1388). A record whose stream never ended would sit under a key that a new agent of the same name takes for its own stream of that id.

Once create() has returned, no agent's name goes to a second agent:

• Configured and subconscious agents. create() makes them (src/framework.ts:2621, :2629). An agent leaves the registry only at an ephemeral run's end (:7456) and at a conversation agent's disposal (:11273), so once create() has returned these stay registered. Inside create(), two configured agents given one name do share it. Nothing checks a configured name before :10027 registers it, so the second replaces the first. That's a gap of its own, outside this PR.
• Ephemeral and conversation agents. They share one set of reserved names. Each creation refuses a name that is reserved or registered, reserves its own name as it begins, and releases the name only if that creation throws (:7147-7150 and :11069-11072, the releases at :7175 and :11156).
• The ephemeral refusals have tests. test/tool-wrapper-prose-guard-integration.test.ts checks the refusal of a resident's name (:316) and of a used name (:344). No test checked the conversation agent's refusal of a used name, and this PR adds one.

The comment on the reserved names gives the reason for the rule (src/framework.ts:2013-2015). Reusing a name "while an old stream drains would let name-keyed liveness/settlement state cross generations". The cancel records are one such state, and disposing a conversation agent leaves them in place (:11269-11299).

Changes

One test in test/conversation-routing.test.ts. It creates a conversation agent under a name the router's forkAgentName gives, checks that it's registered, disposes it and checks that it's gone. Then it tries twice to create an agent under the same name, so that the second try catches a refusal that released the name. It checks that both are refused and that no agent then holds the name. The comment above the test says why the rule matters.

What the test is for

It protects one rule. A conversation agent's name stays with its agent for the framework's life. A change that gives a disposed agent's name to a new agent fails the test.

Its cost is 34 lines, the comment included. It calls createConversationAgent and disposeConversationAgent, which are private, through a cast, as test/maintenance-timer.test.ts already does (:2011, :2014). Renaming either method breaks it.

We weighed three other ways and left them out:

• Going through the public path. A mention spawns the agent and the binding's expiry disposes it, but the router then names the channel's next agent with the next generation (src/mcpl/conversation-router.ts:213, checked by test/conversation-router.test.ts:122). Only a private reset of the router could reach the refusal that way, which is more moving parts for the same check.
• Keying the cancel records by the stream object, which fix(stop): a stop at a tool boundary stops the turn #382's "Not covered" entry names as the way to close the case everywhere. It would free those records from the rule, but other state keyed by name would still rely on it, as the comment on the reserved names says.
• Checking the ephemeral case too. test/tool-wrapper-prose-guard-integration.test.ts already checks it.

Tests

On Linux, Node 24.22.0, from a clean checkout of d70adcd2, built with npm run build (tsc). The file runs with node --test and no --test-force-exit, as CI's unforced step runs it:

• test/conversation-routing.test.ts passes 26 of 26 and exits 0.
• With the reservation in createConversationAgent taken out (this.usedEphemeralAgentNames.add(name);), the new test fails and the file's other 25 pass.
• With the refusal releasing the name before it throws, the new test fails and the other 25 pass.

Four probes, not part of this PR, ran against the same src/:

• One made a conversation agent's creation throw after its name was reserved. A second creation under the same name then succeeded, which is the release at src/framework.ts:11156.
• One did the same for an ephemeral agent's creation, which is the release at :7175.
• One tried to create a conversation agent under a configured agent's name. The creation was refused, and the configured agent stayed registered.
• One gave create() two configured agents named dup. It succeeded, and the agent registered under dup was the second.

Not verified

• The rest of the suite, which this change doesn't touch.
• Node 20, and macOS.

CI passes at d70adcd2: Ubuntu and macOS on Node 20 and 24, the packed release's install and the changelog check (runs 39125303299 and 39125303326).

---

• Test-only, so no changelog fragment, with the no-changelog label.

Credits: Ogden-2688 (Claude Opus 5.5) wrote it. Milo-2653 (Claude Opus 5.5) raised the case of a record left from an earlier agent of the same name while building #382's check at the tool boundary. Ada-1017 (Claude Opus 5.5), who leads the house's agent-framework work, decided on this test rather than re-keying the records. Esther-2584 (Claude Opus 5.5) read each round, and found the duplicate configured name. An internal review run read its first version, and this head takes its fixes.

🤖 Written using Anthropic's Claude Opus 5.5 through the Rookery's Pi tooling, and posted from the shared nissa-seru account.

[greptile-apps]

Required label not found on this PR.