Inspect a nak thread

A thread's behaviour is the product of its message history AND the cached priming payloads on its threads row (intuition_payload, context_recall_payload). When a turn answers the wrong thing, the cause is usually in one of those two places. This skill pulls both in a fixed shape so the diagnosis is one step, not four ad-hoc queries.

Background on the priming staleness failure mode it surfaces: docs/dev/intuition.md and docs/dev/context-recall.md.

Inputs

• A thread id (uuid). If you only have a user, list recent threads first (query C below).
• A target: cloud or local. Pick the runner accordingly.

Runners

• Cloud (the linked project): mcp__claude_ai_Supabase__execute_sql with project_id = the nak project (resolve via mcp__claude_ai_Supabase__list_projects if unknown). The MCP runs with management access, so the auth schema is queryable.
• Local (a running mise run dev-start stack): mise run dev-sql "<sql>". The MCP/cloud tools cannot see the local stack; this is the only way in. Loopback-guarded.

Run the same SQL either way. Substitute :tid with the thread id.

A. Messages in order

select row_number() over (order by created_at) as turn,
       role, status, model, created_at,
       left(regexp_replace(coalesce(content,''), E'[\n\r]+', ' ', 'g'), 140) as snippet,
       (tool_calls is not null) as has_tool_calls, name
from public.messages
where thread_id = ':tid'
order by created_at asc;

Read it as the transcript the model accumulates. Watch for: the last role='user' row (what the turn was actually answering), tool rows (web_search/update_title/etc.), and any status of error / aborted / suspended_for_ask_user.

B. Priming payloads with age + staleness

select
  t.title,
  (select count(*) from public.messages m
     where m.thread_id = t.id and m.role = 'user') as user_rounds,
  -- intuition
  t.intuition_payload->>'trigger'           as intu_trigger,
  (t.intuition_payload->>'computed_at_round')::int as intu_round,
  round(extract(epoch from now()
    - to_timestamp((t.intuition_payload->>'computed_at_at')::bigint / 1000)) / 60.0)::int
                                            as intu_age_min,
  -- context-recall
  t.context_recall_payload->>'trigger'      as recall_trigger,
  (t.context_recall_payload->>'computed_at_round')::int as recall_round,
  round(extract(epoch from now()
    - to_timestamp((t.context_recall_payload->>'computed_at_at')::bigint / 1000)) / 60.0)::int
                                            as recall_age_min
from public.threads t
where t.id = ':tid';

Interpreting staleness

A cached payload is injected into the wire as a <think> block every round. Two fuses force a refresh (see STALE_FUSE_ROUNDS / STALE_FUSE_MS in src/lib/intuition):

• Wall-clock: age_min >= 60. At/over an hour the chat-loop's injection guard SUPPRESSES the payload (won't steer on it) and the next turn recomputes. An age in the hundreds/thousands on a payload that still drove a turn is the classic "answered a stale situation" bug - but note the guard now prevents injection, so on current code a stale payload should be suppressed, not injected.
• Rounds: user_rounds - <payload>_round >= 8. Drift within a session; forces a refresh on the next turn regardless of age.

Flag a payload as suspect when intu_age_min / recall_age_min is large (tens of minutes climbing toward 60+) or when the round gap is near/over 8 - especially if the thread shows a topic shift the mood trigger wouldn't have caught.

To read the actual injected text, add t.intuition_payload->>'synthesis' and left(t.context_recall_payload->>'note', 400) to the select.

C. Find a thread when you only have a user

Local dev user is dev@nak.local. Cloud: use the real address.

select id, title, updated_at,
       intuition_payload is not null  as has_intuition,
       context_recall_payload is not null as has_recall
from public.threads
where user_id = (select id from auth.users where email = ':email')
order by updated_at desc
limit 10;

Output

Summarise: the last user turn vs what the response addressed; any error/aborted/suspended rows; and any priming payload whose age or round gap makes it a staleness suspect. Quote the offending synthesis / note when one looks like the culprit. Don't dump raw rows back wholesale - relay the diagnosis.