mirror of
https://gh-proxy.org/https://github.com/klinsmaya/multica.git
synced 2026-11-10 10:26:46 +08:00
Bohan Jiangandmultica-agent
d54300abaa
Status leaves the compared set, so the snapshot is now just the two body fields. The comparison answers one question — must the agent re-read the issue BODY? — and status never helped answer it: its current value ships on every claim, so the agent learns it without a read. Worse, comparing it worked against the feature. The workflow has the agent set in_progress when it starts and in_review when it delivers, so the agent's OWN bookkeeping came back the next turn as "changed: status" and demanded a body re-read that nothing in the body justified. Replaying this issue's 21 follow-ups, that alone cut "unchanged" from 17 to 4. The earlier rationale for keeping it — a free signal that somebody intervened — was wrong on both counts: it is not free, and the agent can see an intervention by comparing the shipped status against the value it last set itself. issueSnapshotVersion stays 1, and this time rows in the old shape DO exist. That is still correct: removing a field is backward compatible, because every field the new code reads is present in an old row and still means the same thing, and the stale "status" key is ignored. The comment now states the actual bump rule rather than "the set changed" — bump when an old row can no longer answer the current question, which means adding a field or changing how a value is derived, not removing one. A test pins the legacy shape comparing correctly in both the unchanged and the changed direction, so that claim is checked rather than asserted. No migration: the column is JSONB and the schema is unchanged. TestClaimTaskByRuntime_AssigneeAndPriorityAreOutOfScope becomes NonBodyFieldsAreOutOfScope and now moves status alongside assignee and priority, asserting the current status still ships with its NEW value — without that, the fixture's UPDATE could silently not apply and the empty changed-list would prove nothing. Co-authored-by: multica-agent <[email protected]>
44 lines
2.6 KiB
SQL
44 lines
2.6 KiB
SQL
-- issue_snapshot records the comparable state of the issue AS THIS CLAIM SAW IT,
|
|
-- so the next run this agent takes on the same issue can be told whether the
|
|
-- issue itself moved instead of re-reading it unconditionally.
|
|
--
|
|
-- Written once per claim, next to the comment-delivery receipt and under the
|
|
-- same CAS. It is read back by a LATER claim only off the row whose provider
|
|
-- session that claim actually resumes — the row GetLastTaskSession returns, or
|
|
-- the operator-chosen source of a manual rerun — and only when that row
|
|
-- actually started. "The run that started last" is a different row whenever a
|
|
-- poisoned session is skipped or an older run is rerun, and comparing against
|
|
-- it would report "unchanged" to an agent whose resumed memory predates the
|
|
-- change.
|
|
--
|
|
-- One writer per row is why the column lives here and not on `issue`: a
|
|
-- content_revision on the issue row would have to be bumped by every one of the
|
|
-- dozen-odd writers that touch any compared field, and the first one missed
|
|
-- would report "unchanged" forever with no symptom. A snapshot taken at claim
|
|
-- time has exactly one writer, whatever the compared set happens to be.
|
|
--
|
|
-- Shape is {"v":1,...}: version plus sha256 of title and description. Hashes,
|
|
-- not bodies — this column is a comparison key, never a second copy of the
|
|
-- issue text. `v` gates the comparison: a snapshot written by a different
|
|
-- version is treated as unknown, which degrades to the pre-existing "read the
|
|
-- issue" instruction.
|
|
--
|
|
-- The set answers only "must the agent re-read the issue BODY?", so only title
|
|
-- and description qualify — they are the task itself and nothing but a read
|
|
-- delivers them. Status, assignee and priority are out: the claim ships the
|
|
-- current status and assignee on every response, so no comparison is needed to
|
|
-- learn them, and priority does not change the agent's work. Labels, parent,
|
|
-- due date, stage, project and metadata are out too.
|
|
--
|
|
-- Rows written by an earlier shape of this column also carry a "status" key;
|
|
-- it is ignored. Removing a field from the set is backward compatible, so it
|
|
-- does not bump `v` — only adding one, or changing how a value is derived,
|
|
-- makes an old row unable to answer the current question.
|
|
--
|
|
-- Nullable, and NULL is the normal state for every row written before this
|
|
-- migration. A reader must treat NULL as "not compared", never as "unchanged".
|
|
-- No index: the column is never a lookup key. It is only ever projected off a
|
|
-- row some other query already located by (agent_id, issue_id), alongside the
|
|
-- started_at that dates the comment delta.
|
|
ALTER TABLE agent_task_queue ADD COLUMN IF NOT EXISTS issue_snapshot JSONB;
|