feat(mcp): hold a tool listing for an hour, and let something end that early #365

Merged

coilyco-ops merged 2 commits from feat/roster-hour-ttl-and-refresh into main

2026-08-13 11:08:59 +00:00

coilyco-ops commented

2026-08-13 11:05:53 +00:00

Member

Kai chose a 1 hour TTL plus an agent-invokable refresh, over the boot-only
resolution recorded earlier on 163. This is the TTL half and the seam the
refresh needs.

The interval moves 5 minutes to an hour. Refresh marks every rostered server
stale so the next turn re-lists.

The reason an hour is safe is worth stating, because it is the thing I got
wrong when I argued against removing the TTL entirely. Expiry is not the only
trigger. dropSession nils tools, and needsTools returns true whenever tools is
nil, so a server that failed to list retries on every turn its backoff allows
regardless of the interval. The interval governs only servers that already
answered. Removing the TTL would have broken that; lengthening it does not.

Refresh dials nothing. It sets a flag the next Open reads, so twenty calls
collapse into one listing. That is why it carries no rate limit of its own,
which corrects what I said I would build when I claimed this.

The three tests bind those three claims rather than the implementation: an
unlisted server ignores the interval, a refresh re-lists inside it, and repeated
refreshes cost one listing. Mutation checked by making Refresh mark nothing.

The model-callable half is not here. Every tool Echo offers comes from an MCP
server today, so a refresh the agent can invoke needs the first harness-native
tool, which is a new seam rather than a wiring change. Filed separately.

closes #363

Co-authored-by: Kai Siren coilysiren@gmail.com
Co-authored-by: Claude noreply@anthropic.com
Co-authored-by: Kai Siren coilysiren@gmail.com
Co-authored-by: Claude noreply@anthropic.com

Kai chose a 1 hour TTL plus an agent-invokable refresh, over the boot-only resolution recorded earlier on 163. This is the TTL half and the seam the refresh needs. The interval moves 5 minutes to an hour. Refresh marks every rostered server stale so the next turn re-lists. The reason an hour is safe is worth stating, because it is the thing I got wrong when I argued against removing the TTL entirely. Expiry is not the only trigger. dropSession nils tools, and needsTools returns true whenever tools is nil, so a server that failed to list retries on every turn its backoff allows regardless of the interval. The interval governs only servers that already answered. Removing the TTL would have broken that; lengthening it does not. Refresh dials nothing. It sets a flag the next Open reads, so twenty calls collapse into one listing. That is why it carries no rate limit of its own, which corrects what I said I would build when I claimed this. The three tests bind those three claims rather than the implementation: an unlisted server ignores the interval, a refresh re-lists inside it, and repeated refreshes cost one listing. Mutation checked by making Refresh mark nothing. The model-callable half is not here. Every tool Echo offers comes from an MCP server today, so a refresh the agent can invoke needs the first harness-native tool, which is a new seam rather than a wiring change. Filed separately. closes #363 Co-authored-by: Kai Siren <coilysiren@gmail.com> Co-authored-by: Claude <noreply@anthropic.com> Co-authored-by: Kai Siren <coilysiren@gmail.com> Co-authored-by: Claude <noreply@anthropic.com>

coilyco-ops added 1 commit

2026-08-13 11:05:54 +00:00

feat(mcp): hold a tool listing for an hour, and let something end that early

ci / test (pull_request) Successful in 30s

Details

ci / publish-echo-image (pull_request) Has been skipped

Details

ci / publish-observed (pull_request) Has been skipped

Details

ci / image-build (pull_request) Successful in 19s

Details

eb1eb1dffd

Kai chose a 1 hour TTL plus an agent-invokable refresh, over the boot-only
resolution recorded earlier on 163. This is the TTL half and the seam the
refresh needs.

The interval moves 5 minutes to an hour. Refresh marks every rostered server
stale so the next turn re-lists.

The reason an hour is safe is worth stating, because it is the thing I got
wrong when I argued against removing the TTL entirely. Expiry is not the only
trigger. dropSession nils tools, and needsTools returns true whenever tools is
nil, so a server that failed to list retries on every turn its backoff allows
regardless of the interval. The interval governs only servers that already
answered. Removing the TTL would have broken that; lengthening it does not.

Refresh dials nothing. It sets a flag the next Open reads, so twenty calls
collapse into one listing. That is why it carries no rate limit of its own,
which corrects what I said I would build when I claimed this.

The three tests bind those three claims rather than the implementation: an
unlisted server ignores the interval, a refresh re-lists inside it, and repeated
refreshes cost one listing. Mutation checked by making Refresh mark nothing.

The model-callable half is not here. Every tool Echo offers comes from an MCP
server today, so a refresh the agent can invoke needs the first harness-native
tool, which is a new seam rather than a wiring change. Filed separately.

closes #363

Co-authored-by: Kai Siren <coilysiren@gmail.com>
Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: Kai Siren <coilysiren@gmail.com>
Co-authored-by: Claude <noreply@anthropic.com>

coilyco-ops requested review from coilysiren

2026-08-13 11:05:54 +00:00

coilyco-ops referenced this pull request

2026-08-13 11:06:23 +00:00

Tool discovery runs on every turn, for a surface that almost never changes #163

coilyco-ops added 1 commit

2026-08-13 11:06:52 +00:00

Merge branch 'main' into feat/roster-hour-ttl-and-refresh

ci / test (pull_request) Successful in 30s

Details

ci / publish-echo-image (pull_request) Has been skipped

Details