[johnnyelwailer]

Two changes that share one idea: the things an agent needs to remember must live outside its context window.

Both are grounded in one real 31-hour nexplore orchestration run (thread fbdb583b): 128 turns, 22 compactions, 12 hard mid-conversation truncations, 10 session restarts.

1. Search the whole transcript, newest first (6e51a22)

t3team.thread.search scanned only messages. A thread detail also carries activities — every command and its output, every file read, every tool call. In that run: 1319 searchable messages against 1250 invisible activities, and the invisible half is where the evidence lives.

• messages and activities interleaved chronologically, so a match ranks by when it happened, not which stream it came from
• order now defaults to recent. Matches are capped, so transcript order returned the oldest — one real call reported totalMatches: 70, returnedMatches: 5, the five earliest of seventy
• offset paging, hasMore, and a hint naming the exact next offset
• scope narrows to one stream; role filters a message role or an activity kind
• a multi-word query matching nothing verbatim retries requiring every term (matchMode); terms under 3 chars are excluded so "a b" cannot match inside "bash"
• activity matches report activity_id, not message_id — an activity id can never be resolved by thread.read_message, and offering it under that key invites a lookup that always fails

The messages-only helper and search_source's divergent index result shape are removed rather than kept alongside. Both tools return the same shape now.

2. Durable per-thread task journal (b77b219)

That run made zero task records — no such tool existed — and 75 t3team_children polls, because the orchestrator had no durable plan and re-derived its own state by interrogating its children.

TaskRecord contract, migration 056 (thread_task_records, indexed on (thread_id, position)), a persistence service, and two broker tools. t3team.task.write replaces the whole list rather than patching by id — deliberate, because the primary consumer is a weak local model that handles re-sending a list far more reliably than addressing individual records. note is part of the record so a failure keeps its reason.

Review

Ran a refute-first adversarial pass. It found 4 real defects in the first draft of the search work, all fixed here: stream-ordered instead of time-ordered ranking, unreachable arguments (schemas never declared offset/scope/order), activity ids masquerading as message ids, and 1–2 char terms driving the fallback.

One finding was not fixable here and is worth knowing: the nexplore pack records only the first 500 characters of a tool result, so activity search finds which command ran and its first 500 chars, not full output. Both tool descriptions now say so. A distro-side change raises that budget separately.

Verification

t3team-toolBrokerBindingSearch{Thread,Source}   25 passed
t3team-toolBrokerBindingTaskJournal             11 passed
t3team-ThreadTaskRecords                         5 passed
t3team-toolkit (MCP parity)                      8 passed
t3teamToolCatalog                                5 passed
                                    total 49 + 5 passed, 0 failed

Additive guard: no errors from these files; one warning (t3team-ThreadTaskRecords.ts, 161 lines, under the 200 cap). Remaining guard/lint/typecheck failures reproduce at clean fork/main.

Also fixes three catalog tests that were already red on main: t3team.thread.ask_user is status: "implemented" but was missing from both expected lists and the catalog doc.

Not included: web UI, and the pack-side capability policy that would auto-enable the journal per provider.

🤖 Generated with Claude Code