v0.8.5

Directive delivery and daemon stability. Two of the fixes in this release correct regressions introduced earlier in the same release cycle and caught by the pre-release QA waterfall before any tag existed — the notes below describe net behaviour against 0.8.4.

Fixed

• Tier-0 directives were silently truncated at 800 chars. MAX_CORE_MEMORY_CHARS applied a flat per-item guillotine with no signal to anyone. Because importance correlates with length and truncation eats the tail — where a rule's concrete specifics live — this preferentially destroyed the most important directives. On a real 25-entry install, a majority were being cut mid-clause every turn while ~4k chars of the tier-0 budget sat unused. perItemCapFor() now scales the cap with priority, and applyCoreBudgetVerbose() fits entries in two passes: admission at the flat floor first, then leftover slack spent raising caps highest-priority-first (partial upgrades allowed). Nothing is dropped that would have fit before, nothing gets less room than before, and the budget is actually used. Measured on the same install: 25/25 entries injected, 0 dropped, full budget consumed. (#19)
• TIER0_MAX_TOTAL capped tier 0 by entry COUNT, which is the wrong unit — entries vary by an order of magnitude in size, so a count cap refuses writes while under budget and permits sets that overflow. Admission now tests the candidate against the same character budget the renderer enforces, and refuses with something actionable (budget_full names the weakest entries; would_evict names only what the candidate would actually displace, diffed against a baseline so pre-existing overflow is reported separately rather than blamed on the new entry). The update path is checked too, not just add. (#19)
• The injection envelope was deleting its own section tags. wrapMemoryContext ran stripStructuralTags over the fully assembled context, and recalled_memory / active_directives / session_directives / reflection_context are all on that list — so the labels laqrumcode had just written were stripped on the way out and tier-0 and tier-1 directives reached the model as one unlabelled run of bullets. Breakout protection moved to the content (retrieved nodes, core-memory text and category, skill and reflection fields, resurfacing memories, wakeup briefing, compaction summary, soul summary), where the wrapper tags survive; stripReminderWrapper now removes only <system-reminder>. (#21)
• stripStructuralTags was single-pass and defeatable by nesting — deleting an inner tag splices the outer halves into a live one, so depth-N nesting needed N+1 passes. Both strippers now run to a fixpoint (guaranteed to converge, since every effective pass strictly shortens the string) and neutralize any residue if the safety bound is ever reached.
• Tier-1 "session-pinned" directives were pinned to nothing. getAllCoreMemory had no session filter and nothing retires the rows, so every session directive ever written loaded into every later session — including entries whose own text says "this session". Root cause was deeper than the read path: the MCP relay invented mcp-client-<pid> as its session id while the hook path keyed on Claude Code's session UUID, so the two id spaces never intersected. The relay now sends CLAUDE_CODE_SESSION_ID (resolveSessionId(), with an explicit LAQRUMCODE_SESSION_ID pin still winning for auto-drain), unifying the spaces; tier-1 reads are scoped to the asking session; and a one-shot migration retires legacy mcp-client-* rows explicitly rather than letting them go quietly invisible. This also fixes injectedSections invalidation (it was clearing the cache on the session that does not build the prompt) and makes the per-session tier-0 write cap survive a relay restart. (#21)
• archiveOldTurns could stop draining silently. The anti-join ran in-DB with LIMIT applied after the membership filter, so the whole backlog paid an O(stale × referenced) per-row test and crossed the 8s TIMEOUT once it grew. Rewritten as indexed scans plus a pure O(stale + referenced) Set anti-join (selectUnreferencedTurns), now paged rather than capped at a single window — referenced turns stay candidates forever and pile up at the head of the scan, so a fixed window eventually yields nothing on every run. The job also records a status='error' row on failure so the 30-minute backoff engages instead of re-firing the full scan every boot, and warns when a non-empty candidate set yields nothing archivable.
• Bootstrap maintenance no longer re-hammers the database during a daemon flap — deferred and gated on the daemon having retained an IPC client, so transient cold-boot daemons skip it.
• Daemon spawn-lock race. The holder pid is now stamped at O_EXCL acquisition, so a racer can no longer read an empty lock (Number("") === 0), judge it dead, and spawn a second daemon concurrently.
• Reconnect thundering herd — jittered backoff before an mcp-client reconnect so a single daemon exit doesn't make every relay re-handshake in the same instant.

Changed

• The tool-call planning gate keys on toolCallsSinceLastText rather than total tool calls — the signal that separates a loop from productive work. Documented as inert: nothing in the production hook path produces that signal today, so it fires exactly where the old predicate did. See #20.
• The softInterrupt block message no longer claims the user pressed Ctrl+C; nothing sets that flag from a real interrupt.

Known follow-ups

• #20 — wire an assistant-text producer so the loop guard can actually distinguish looping from work.
• Age-based archival of orphaned tier-1 rows (hygiene; read-scoping is the correctness mechanism).
• retrieval_outcome id prefetch in archiveOldTurns is unbounded and re-Set per page.
• Integration tests gated on a live SurrealDB return early rather than skipping, so they report green in CI, which provides no database.