You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
This commit was created on GitHub.com and signed with GitHub’s verified signature.
[1.14.2] — 2026-11-08
Fixed
Loop: the Tool grains are counted, not read — a pass's memory no longer
grows with them (#433). 1.14.1 bounded the watermark reads, but two
readers never used the watermark. tool_failure reads a rolling window_days window (30 days by default), which is the whole memory while
the memory is younger than that, and held it as one set of records with
every body decoded and cloned. The gate's count_new built a record for
every grain since the watermark just to count it. On a 20,145-grain
Postgres memory (11,286 Tool grains, 237 MB of bodies) a pass with nothing
new peaked at 2.7 GiB RSS and aborted under a 2.5 GiB cap; the gate alone
cost 1 GiB. The store now keeps tool_idx, one narrow row per Tool grain
(tool name, error flag, Definition flag, derived_from parent), and the
Loop asks it questions instead. tool_failure takes its per-tool counts
from the index and decodes only the error grains it clusters. The gate
counts grains per type, Tool errors and distinct Event sessions from
indexes, so a skipped pass decodes only the Loop's own state. tool_error_recurrence
re-measurement reads the errors since the apply. Liveness is unchanged:
a call superseded inside the window is still not counted, now answered by
the index's parent column. Measured with 20 KB Tool grains on a file
memory, a pass with nothing new went from 8,148 decoded blobs and 218 MB
peak RSS (8,000 grains; 93 MB at 2,000) to the window's failures plus
bounded evidence, at a flat ~33 MB. New store reads: tool_counts, tool_errors, count_of_type_since, sessions_of_type_since, rebuild_tool_index. Where the index cannot
answer (a read-only open of a pre-1.14.2 memory, which never builds it, or
a live anonymization policy) the same reads fold the grains a page at a
time (SubstrateRead::visit_grains_of_type): the same answer, holding one
page and the running counts. Pinned by the extended bounded_reads.rs (a full pass's decode count independent of the window,
and the index path and the fold proposing identical drafts).
Loop: the evalset pin and cold_grains read Tool Definitions, not every
Tool grain (#432).cold_grains' in-use set and tool_evalset (Rule
E1's pin) read the whole Tool type to find the handful of Definitions.
They now read Areev::tool_definitions, served by the same index, so the
blobs decoded are the Definitions' alone. A handle without the index falls
back to a paged scan.
Loop: the remaining whole-type Tool reads are folded. The LLM
verifier's evidence seeding and the opt-in command analyzer's snapshot keep
only what fits their caps instead of reading the type and truncating it.
areev reindex also rebuilds the Tool index. The first read-write
open after upgrading builds it with one paged pass over the Tool grains
(its own tool_index stamp, deliberately not a link_index bump, whose
rebuild reads every blob at once).
Postgres schema version 4: tool_idx and an index on thread_idx(seq).
Rolling-deploy safe from version 1. An older writer's Tool grains written
during the mixed-version window are unindexed until areev reindex, so
drain old writers or reindex once afterwards for exact Loop counts
(docs/deployment-profile.md). Conformance: tool_index_answers_counts_errors_and_definitions_without_reading_calls,
both backends.
Changed
Internal crate requirements raised to 1.14.2. The adapter and the
CLI call store and Loop API added in this release, so ^1.14.0 would let a
lockfile pair them with a 1.14.1 dependency that lacks it.