agent-stream: start with what every session is doing now

[asubbarao]

Adds a first query to the agent-stream skill for "what is every session doing now" (other terminals, Claude Desktop): one row per Claude session active in the last hour, with its titles, branch, the PRs it linked, minutes idle, and the other live sessions that share its branch or one of its PRs, which is where two sessions are about to collide.

• Reads only top-level fields of agent.lines (sessionId, timestamp, gitBranch, aiTitle/customTitle, prRepository/prNumber): no message text, and nothing from agent.messages and agent.docs: the stream and the memories as plain rows #58 or One indexed transcript table: stream.blocks, agent.messages and agent.sessions on top #68.
• staging, the shared checkout's branch, doesn't count as an overlap.
• Limits stated in the skill: titles have no timestamp, so a renamed session lists them all; only PRs Claude Code recorded as linked appear; Codex sessions aren't covered.

Tested by running the block from the file on dev: 4 live sessions, 2 of them on one inf-1656 branch. An earlier run showed 14 sessions, 3 sharing that branch.

🤖 Generated with Claude Code

[asubbarao]

duckstack/rules: success

• warn sql_json_extract in skills/agent-stream/SKILL.md#2.sql: Read JSON with read_json and its typed columns (dot access on structs); no json_extract on nested fields.
• warn sql_json_extract in skills/agent-stream/SKILL.md#4.sql: Read JSON with read_json and its typed columns (dot access on structs); no json_extract on nested fields.
• warn sql_json_extract in skills/agent-stream/SKILL.md#5.sql: Read JSON with read_json and its typed columns (dot access on structs); no json_extract on nested fields.
• warn sql_json_extract in skills/agent-stream/SKILL.md#6.sql: Read JSON with read_json and its typed columns (dot access on structs); no json_extract on nested fields.

duckstack review queue, rule gate (hooks/hook_checks.sql)

[asubbarao]

duckstack/review: error

The reviewer gave no verdict. Its answer, cut at 3000 characters:

Opus 5.5, duckstack review queue (duckstack-review skill)

[asubbarao]

duckstack/rules: success

• warn sql_json_extract in skills/agent-stream/SKILL.md#2.sql: Read JSON with read_json and its typed columns (dot access on structs); no json_extract on nested fields.
• warn sql_json_extract in skills/agent-stream/SKILL.md#4.sql: Read JSON with read_json and its typed columns (dot access on structs); no json_extract on nested fields.
• warn sql_json_extract in skills/agent-stream/SKILL.md#5.sql: Read JSON with read_json and its typed columns (dot access on structs); no json_extract on nested fields.
• warn sql_json_extract in skills/agent-stream/SKILL.md#6.sql: Read JSON with read_json and its typed columns (dot access on structs); no json_extract on nested fields.

duckstack review queue, rule gate (hooks/hook_checks.sql)

[asubbarao]

duckstack/review: failure

• skills/agent-stream/SKILL.md:84 Check 1 / RULES.md 3 (a column passed into a reader); the skill's own rule at line 10 (do not re-parse the raw JSONL): read_ndjson(file_path) puts a row's column into a table function, which hits the binder error that self-dispatch exists to avoid, and it re-parses raw JSONL that agent.lines already stores with its message column.
  Rewrite: head is cut. For a message's full text, join the store instead of re-reading the file:

SELECT d.session_id, d.ts, l.message -> 'content' AS content
FROM day d
JOIN agent.lines l ON l.transcript = d.file_path AND (l.timestamp ->> '$')::TIMESTAMPTZ = d.ts;

• skills/agent-stream/SKILL.md:78 RULES.md 10 (lossless: keep columns, SELECT * EXCLUDE): The query keeps only six columns and drops file_path, tail, block_type and is_prompt, even though line 84 tells the reader to use file_path at that row's timestamp to get the full text, so the follow-up step has no key to start from.
  Rewrite: WITH day AS (
  SELECT *, timestamp AS ts,
  ts > now() - INTERVAL 1 HOUR AS live,
  first_value(timestamp) OVER w AS last_ts,
  row_number() OVER w AS back
  FROM agent.messages
  WHERE timestamp > now() - INTERVAL 24 HOUR AND block_type = 'text' AND NOT is_agent
  AND CASE message_role WHEN 'user' THEN is_prompt ELSE true END
  WINDOW w AS (PARTITION BY session_id ORDER BY timestamp DESC)
  )
  SELECT live, back, * EXCLUDE (live, back)
  FROM day
  ORDER BY live DESC, CASE WHEN live THEN back END, last_ts DESC, ts DESC
  LIMIT 20;

Opus 5.5, duckstack review queue (duckstack-review skill)

[asubbarao]

Closed: the search moves to #68's BM25 index instead of a long skill section. — Opus 5.5

[asubbarao]

duckstack/rules: success

• warn sql_json_extract in skills/agent-stream/SKILL.md#2.sql: Read JSON with read_json and its typed columns (dot access on structs); no json_extract on nested fields.
• warn sql_json_extract in skills/agent-stream/SKILL.md#4.sql: Read JSON with read_json and its typed columns (dot access on structs); no json_extract on nested fields.
• warn sql_json_extract in skills/agent-stream/SKILL.md#5.sql: Read JSON with read_json and its typed columns (dot access on structs); no json_extract on nested fields.
• warn sql_json_extract in skills/agent-stream/SKILL.md#6.sql: Read JSON with read_json and its typed columns (dot access on structs); no json_extract on nested fields.

duckstack review queue, rule gate (hooks/hook_checks.sql)

[asubbarao]

duckstack/review: failure

• skills/agent-stream/SKILL.md:61 RULES.md 25 (delete wrong advice): The new prose says never to use ORDER BY session, ts and that origin.kind = 'human' replaces the hand-written prefix filters. But the Turns query in the same file still ends with ORDER BY session, ts and still filters prompts by the <, '# AGENTS.md' and 'This session is being continued' prefixes, so the file now contradicts itself.
  Rewrite: In the Turns prompts CTE: WHERE role = 'user' AND NOT is_subagent AND origin_kind = 'human' AND NOT starts_with(ltrim(text), '') -- with origin ->> 'kind' AS origin_kind projected in recent; and end it with ORDER BY ts DESC
• skills/agent-stream/SKILL.md:81 RULES.md 10 (lossless): LIMIT 20 silently drops sessions in a committed query that is meant to show what every session is doing. The description says 37 sessions over 24 hours, so the cap already cuts real sessions. The 24-hour window already bounds the result.
  Rewrite: FROM texts GROUP BY session_id ORDER BY last_ts DESC;
• skills/agent-stream/SKILL.md:76 RULES.md 10 / RULES.md 13 (lossless; OR in disguise): CASE role WHEN 'user' THEN … ELSE true END in a WHERE is an OR chain written as a CASE, and it drops rows silently instead of deriving a flag and filtering in the next step.
  Rewrite: texts AS (
  SELECT session_id, ts, role, coalesce(block ->> 'text', block ->> '$') AS text,
  CASE WHEN role = 'assistant' THEN 'reply'
  WHEN origin_kind = 'human' AND NOT starts_with(ltrim(text), '') THEN 'prompt' END AS kind
  FROM blocks WHERE coalesce(block ->> 'type', 'text') = 'text'
  )
  SELECT session_id, max(ts) AS last_ts, arg_max(kind, ts) AS last_kind, arg_max(text, ts) AS last_text, strlen(last_text) AS last_text_bytes
  FROM texts WHERE kind IS NOT NULL GROUP BY session_id ORDER BY last_ts DESC;
• skills/agent-stream/SKILL.md:68 RULES.md 13 / check 7 (selector-taking extractors): CASE json_type(message -> 'content') … message -> '$.content[*]', together with type ->> '$', timestamp ->> '$' and origin ->> 'kind', is the arrow ladder the rubric shows as its BAD example, and the rule gate flagged it as json_extract.
  Rewrite: Read the typed columns of the shared messages relation (session, ts, role, text, is_prompt) rather than digging into agent.lines' JSON columns in each query: FROM agent.messages WHERE ts > now() - INTERVAL 24 HOUR AND NOT is_agent
• skills/agent-stream/SKILL.md:65 check 5 / check 9 (missing shared relation, bloat): The new query is the fifth copy of the blocks/messages unnest in this file, and it duplicates the existing 'One row per session' query a few lines below it. The author's reason, not depending on agent.messages and agent.docs: the stream and the memories as plain rows #58, doesn't hold: the description says the result was checked against agent.messages on dev, so that shared relation already exists and is what this query should read.
  Rewrite: SELECT session, max(ts) AS last_ts, arg_max(role, ts) AS last_role, arg_max(text, ts) AS last_text,
  strlen(last_text) AS last_text_bytes
  FROM agent.messages
  WHERE ts > now() - INTERVAL 24 HOUR AND NOT is_agent
  GROUP BY session ORDER BY last_ts DESC;
  -- and delete the duplicate 'One row per session' query, or make it this one

Opus 5.5, duckstack review queue (duckstack-review skill)

[asubbarao]

duckstack/rules: success

• warn sql_json_extract in skills/agent-stream/SKILL.md#2.sql: Read JSON with read_json and its typed columns (dot access on structs); no json_extract on nested fields.
• warn sql_json_extract in skills/agent-stream/SKILL.md#4.sql: Read JSON with read_json and its typed columns (dot access on structs); no json_extract on nested fields.
• warn sql_json_extract in skills/agent-stream/SKILL.md#5.sql: Read JSON with read_json and its typed columns (dot access on structs); no json_extract on nested fields.
• warn sql_json_extract in skills/agent-stream/SKILL.md#6.sql: Read JSON with read_json and its typed columns (dot access on structs); no json_extract on nested fields.

duckstack review queue, rule gate (hooks/hook_checks.sql)

[asubbarao]

duckstack/review: failure

• skills/agent-stream/SKILL.md:77 RULES.md 12 / 10 / 5 (results come back as columns): The kind of overlap is encoded by prefixing the string 'branch ', and line 81 then builds shares_with by concatenating session_id, ' ' and that string. Structure turns into text that the reader has to parse by position.
  Rewrite: ), touches AS (
  SELECT session_id, 'branch' AS kind, branch AS value FROM per_session WHERE branch <> 'staging'
  UNION ALL BY NAME
  SELECT session_id, 'pr' AS kind, unnest(prs)::VARCHAR AS value FROM per_session
  ), shared AS (
  SELECT a.session_id, list_distinct(list({'session_id': b.session_id, 'kind': a.kind, 'value': a.value})) AS shares_with
  FROM touches a JOIN touches b ON a.kind = b.kind AND a.value = b.value AND a.session_id <> b.session_id
  GROUP BY a.session_id
  )
• skills/agent-stream/SKILL.md:71 skill: shorter version / RULES.md 9: live_sessions computes max(ts) per session, and per_session then joins it back and regroups on the same key. A single GROUP BY with HAVING does both, which removes a CTE and a join (6 lines → 4).
  Rewrite: ), per_session AS (
  SELECT session_id, max(ts) AS last_ts, list_distinct(list(title)) AS titles, arg_max(branch, ts) AS branch,
  list_distinct(list(CASE WHEN pr_number IS NOT NULL THEN {'repo': pr_repo, 'number': pr_number} END)) AS prs
  FROM session_lines GROUP BY session_id HAVING max(ts) > now() - INTERVAL 1 HOUR
• skills/agent-stream/SKILL.md:67 RULES.md 12 (no || chains) / 10 (lossless): The PR key is glued together with a || chain mixed with ->>. Both operators share one precedence level in DuckDB's parser, so this can parse as ((… || prNumber) ->> '$'). It also folds repo and number into one string that a later reader would have to split apart.
  Rewrite: -- session_lines: keep the parts as columns
  prRepository ->> '$' AS pr_repo, prNumber ->> '$' AS pr_number
  -- per_session: one struct per linked PR; list_distinct drops the NULLs
  list_distinct(list(CASE WHEN s.pr_number IS NOT NULL THEN {'repo': s.pr_repo, 'number': s.pr_number} END)) AS prs
• skills/agent-stream/SKILL.md:68 skill: keep the read bounded / manifesto §9 / skill rule 'One statement covers every session. Filter by a time window, transcript IN (…)': session_lines parses every top-level JSON field of every line of every Claude session in the store, and only drops old sessions later in a HAVING. That is the full transcript scan the rubric says once took dev down. The PR's note (title lines carry no timestamp) explains why there is no filter on line time, but it does not justify skipping a file-level filter: filtering on the transcript file's modified time keeps every title line of a live session.
  Rewrite: WITH live_files AS (
  SELECT path FROM agent.entries
  WHERE modified > now() - INTERVAL 1 HOUR
  AND NOT contains(path, '/subagents/') AND NOT contains(path, '/.codex/')
  ), session_lines AS (
  SELECT sessionId ->> '$' AS session_id, (timestamp ->> '$')::TIMESTAMPTZ AS ts,
  gitBranch ->> '$' AS branch, coalesce(customTitle ->> '$', aiTitle ->> '$') AS title,
  prRepository ->> '$' AS pr_repo, prNumber ->> '$' AS pr_number
  FROM agent.lines l SEMI JOIN live_files f ON f.path = l.transcript
  )

Opus 5.5, duckstack review queue (duckstack-review skill)