std: add caos-conversation-list tool

[malcolmredheron]

What

Adds std/caos-conversation-list, a tool that lists the conversations recorded on the server, newest first. std/caos-conversation can read one conversation but needs its tip hash; nothing could produce that hash before.

Each line has the UTC time of the last message, the conversation id, the tip commit hash, the title, and user:active / user:archived membership when there is any:

2026-10-03T00:44:51Z  cc/de6ac24a-…  b6a82041…  "Import metta-ai/caos. Then read about where the mcp server …"
2026-10-03T00:43:02Z  cc/86f1e6e6-…  5036546e…  "There's unfinished work on the actors-design branch of mett…"

Parameters: filter (substring of id or title), limit (default 50, 0 for all), titles (set 0 to skip title reads). Because the list is newest first, limit=N is the N most recent.

How

• Conversations are git refs on the server: refs/caos/v3/conversations/<hex id>/head, plus per-user refs/caos/v3/users/<hex user>/conversations/{active,archived}/<hex id> (see conversation-protocol/src/v3/refs.rs). The worker lists them with git ls-remote against $CAOS_SERVER_URL; std/go already ships git for refs work.
• The time is the committer time of the tip commit, since refs carry none. Sorting newest first therefore reads every tip commit (one caos get-hash each) before applying limit. Titles are read only for the conversations printed, unless filter is set.
• Layout follows std/caos-conversation: .caos-expr with the help as a here-string, DEPS on std/go, and worker.go on the w prelude. Every outcome, including an unreachable server, is returned as text rather than a job error. A conversation whose tip cannot be read is still listed, last, with the reason.
• std/README.md has a row for the tool.

Testing

• caos-test on this branch's tree: SUITE OK, 49/49 passed (293s), including lint, unit-fmt, unit-clippy, caos-tools, deep-deps and eval-path. The run used the cache, so tests whose inputs did not change replayed earlier results; I did not salt it to force every test to re-run.
• The suite has no test for this tool, so that result shows the change breaks nothing else, not that the tool works. No tests/ entry is added here.
• The tool itself was exercised through run_tool against a live server with 59 conversations: ordering, limit, titles and timestamps check out.
• The members tags were not exercised: none of the conversations printed in my runs had a membership ref.

Related

std/caos-conversation/recording-gaps.md (item 5) says every commit has the same timestamp. That does not hold on the server I tested: tips have distinct committer times, and this tool depends on that. I left the note unchanged here; it may be worth correcting separately.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Qmq9SYrkF8KYc2iZYjtgoh