agent.sessions: one view per session, every column carried

[asubbarao]

Folds agent.prompts into agent.sessions, so there is one per-session view, and stops dropping columns.

• prompts is every typed prompt as its whole stream.blocks row (all columns), in order; first_prompt / last_prompt are full text, not 100-character heads.
• Per session also: cwd, git_branch, entrypoint (latest), models, tools, files, has_agents, bytes, beside the existing activity and block_keys.
• tool_name no longer falls back to call_id (Codex tool results have no name; tool_use_id already holds the id). That fallback made one session show 557 'tools'.
• what-alok-said's collect.sql is one SELECT over agent.sessions; the agent-stream skill names the new columns.

Tested on dev as a query over the live stream.blocks: 1,082 sessions, 12,515 prompts after near-repeat dropping; the query runs and returns whole prompt rows (two sessions' rows were 421 KB, which is why readers select columns).

🤖 Generated with Claude Code

https://claude.ai/code/session_01NCnrCZ6kWmagfJjoK8jVGQ

[asubbarao]

duckstack/review: failure

• server/messages.sql:23 RULES.md 10 (lossless: derive a flag with CASE and filter in the next CTE; never silently drop rows): The view drops near-repeat prompts inside the WHERE clause, so a reader of agent.sessions can't get those rows back or tell how many were dropped, while the PR says it 'stops dropping columns'.
  Rewrite: ), prompt AS (
  SELECT *, coalesce(rapidfuzz_ratio(left(message_content, 100), lag(left(message_content, 100)) OVER (PARTITION BY session_id ORDER BY timestamp)) >= 90, false) AS is_near_repeat
  FROM session_blocks WHERE is_prompt
  ), prompts AS (
  SELECT session_id, array_agg(prompt ORDER BY timestamp) AS prompts
  FROM prompt GROUP BY session_id
  )
  -- and readers filter with [p FOR p IN prompts IF NOT p.is_near_repeat]; prompt_count/first_prompt/last_prompt use that filtered list

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

[asubbarao]

duckstack/review: failure

• server/messages.sql:19 Check 9 / RULES.md 9 (short): The prompt_openings and prompt CTEs are two steps for one derived column. previous_opening exists only to feed the next CTE, and a window function can sit directly inside the comparison.
  Rewrite: ), prompt AS (
  SELECT *, coalesce(rapidfuzz_ratio(left(message_content, 100),
  lag(left(message_content, 100)) OVER (PARTITION BY session_id ORDER BY timestamp)) >= 90, false) AS is_near_repeat
  FROM session_blocks WHERE is_prompt
  ), prompts AS ( -- 6 lines become 4; prompt_openings is deleted

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

[asubbarao]

duckstack/review: failure

• skills/what-alok-said/SKILL.md:16 RULES.md 26 (tested before offered) / RULES.md 23: collect.sql ends with ORDER BY started;, so appending AND session_id IN (<its group>) to it, as this line says, gives invalid SQL. The advice cannot have been run.
  Rewrite: Nothing is stored: each reader in step 2 runs SELECT * EXCLUDE (block_keys) FROM agent.sessions WHERE session_id IN (<its group>) ORDER BY started;. Alternatively, collect.sql can put the window in a WHERE that readers replace, e.g. WHERE last_active > now() - INTERVAL 14 DAY with ORDER BY on its own line, and this step says to replace that line rather than append to the file.
• server/blocks.sql:14 Check 9 (bloat) / shorter version; RULES.md 9: Removing 'call_id' leaves name_keys with a single key, so the list, its comprehension at the tool_name line and its EXCLUDE entry now do the work of one ->> and can be deleted.
  Rewrite: Line 14: SELECT ['text', 'thinking', 'output', 'content'] AS content_keys,
  parsed CTE: SELECT * EXCLUDE (b, content_keys, id_keys), ... , b ->> 'name' AS tool_name, ...
  (3 expressions become 1; the line count stays the same)

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

[asubbarao]

duckstack/review: success

No findings.

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