Small, individually confirmed, none blocking on their own:
session_commands.go:563 -- the compression summary is persisted as a USER turn. CompressHistory emits it with Role 'system' (compression.go:118) but the write-back only maps 'assistant' -> Identity, so system falls through to 'user'. The model replays '[Earlier conversation has been summarised...]' as if the human said it. (Retained bot turns are handled correctly.)
session_commands.go:488 -- compressPreview reads RecentMessages(200) while reporting count from MessageCount. With 100 large + 250 tiny messages the preview says 'Compression not needed (350 messages, ~0 tokens)' while apply on the same session reports '350 -> 24 (~100000 -> ~3063 tokens)'. The two numbers in one sentence describe different histories, and the preview advises against a compression that is needed.
session_commands.go:329 -- handleUndo discards the deleted count from DeleteRecentMessages, so /undo 10 failing after 6 removals reports failure while six messages are gone.
session_store.go:626 -- nextTimestamp has no upper bound and lastTimestamp re-reads from the store, so one future-dated write (host clock 24h fast) pins the session's clock for its lifetime; every later At handed to callers is wrong by a day.
session_store.go:268 -- Close() calls s.store.Close() on the SHARED nl.Store (in production the daemon's tasks/events log) and, unlike every other method, does not take s.mu. Nothing calls it today.
session_store.go:429 -- newMessageID's UUIDv5 input sessionID+"\x00"+sourceID is not injectively encoded: newMessageID('a\x00b','c') == newMessageID('a','b\x00c'). No reachable input today; hardening, not a live bug.
turns.go:79-85 -- Turns.lanes grows without bound; a lane is inserted per session key and never deleted. No goroutine leak.
Imported from beads archie-core-1786637499459-369-0770c1e8 · GitHub #390 · type bug · priority P2 · created 2026-08-08 · closed 2026-08-09
Small, individually confirmed, none blocking on their own:
1. session_commands.go:563 -- the compression summary is persisted as a USER turn. CompressHistory emits it with Role 'system' (compression.go:118) but the write-back only maps 'assistant' -> Identity, so system falls through to 'user'. The model replays '[Earlier conversation has been summarised...]' as if the human said it. (Retained bot turns are handled correctly.)
2. session_commands.go:488 -- compressPreview reads RecentMessages(200) while reporting count from MessageCount. With 100 large + 250 tiny messages the preview says 'Compression not needed (350 messages, ~0 tokens)' while apply on the same session reports '350 -> 24 (~100000 -> ~3063 tokens)'. The two numbers in one sentence describe different histories, and the preview advises against a compression that is needed.
3. session_commands.go:329 -- handleUndo discards the deleted count from DeleteRecentMessages, so /undo 10 failing after 6 removals reports failure while six messages are gone.
4. session_store.go:626 -- nextTimestamp has no upper bound and lastTimestamp re-reads from the store, so one future-dated write (host clock 24h fast) pins the session's clock for its lifetime; every later At handed to callers is wrong by a day.
5. session_store.go:268 -- Close() calls s.store.Close() on the SHARED nl.Store (in production the daemon's tasks/events log) and, unlike every other method, does not take s.mu. Nothing calls it today.
6. session_store.go:429 -- newMessageID's UUIDv5 input sessionID+"\x00"+sourceID is not injectively encoded: newMessageID('a\x00b','c') == newMessageID('a','b\x00c'). No reachable input today; hardening, not a live bug.
7. turns.go:79-85 -- Turns.lanes grows without bound; a lane is inserted per session key and never deleted. No goroutine leak.
---
_Imported from beads `archie-core-1786637499459-369-0770c1e8` · GitHub #390 · type bug · priority P2 · created 2026-08-08 · closed 2026-08-09_
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
Small, individually confirmed, none blocking on their own:
session_commands.go:563 -- the compression summary is persisted as a USER turn. CompressHistory emits it with Role 'system' (compression.go:118) but the write-back only maps 'assistant' -> Identity, so system falls through to 'user'. The model replays '[Earlier conversation has been summarised...]' as if the human said it. (Retained bot turns are handled correctly.)
session_commands.go:488 -- compressPreview reads RecentMessages(200) while reporting count from MessageCount. With 100 large + 250 tiny messages the preview says 'Compression not needed (350 messages, ~0 tokens)' while apply on the same session reports '350 -> 24 (~100000 -> ~3063 tokens)'. The two numbers in one sentence describe different histories, and the preview advises against a compression that is needed.
session_commands.go:329 -- handleUndo discards the deleted count from DeleteRecentMessages, so /undo 10 failing after 6 removals reports failure while six messages are gone.
session_store.go:626 -- nextTimestamp has no upper bound and lastTimestamp re-reads from the store, so one future-dated write (host clock 24h fast) pins the session's clock for its lifetime; every later At handed to callers is wrong by a day.
session_store.go:268 -- Close() calls s.store.Close() on the SHARED nl.Store (in production the daemon's tasks/events log) and, unlike every other method, does not take s.mu. Nothing calls it today.
session_store.go:429 -- newMessageID's UUIDv5 input sessionID+"\x00"+sourceID is not injectively encoded: newMessageID('a\x00b','c') == newMessageID('a','b\x00c'). No reachable input today; hardening, not a live bug.
turns.go:79-85 -- Turns.lanes grows without bound; a lane is inserted per session key and never deleted. No goroutine leak.
Imported from beads
archie-core-1786637499459-369-0770c1e8· GitHub #390 · type bug · priority P2 · created 2026-08-08 · closed 2026-08-09