Repository navigation
Releases
Releases · Lisa-Harness/lisa
v0.6.0
v0.5.4
v0.5.3
Lisa Harness v0.5.3 — patch release.
Highlights
- One shared distillation ladder. The two memory distillers' hand-rolled
retry loops are one kernel escalation
(kernel.operations.structured_output.run_distillation_ladder) behind a
DistillationLadderpolicy object: structured output → JSON-only re-ask →
validation-feedback re-ask naming the failing fields and the per-fact cap →
assistant-bootstrap prefill. The payload type is generic, so the distillers'
typed returns (DistilledLearning,MinedFacts) type-check without casts. - Mining distiller crash-loop fixed end to end (2026-11-08..09 incident,
21 consecutive failures, zero segments mined): budget escalation on retries
(16384-token module default, authored 300s timeout), a per-segment attempt
budget that advances the marker past a poisoned segment instead of starving
every later one, and a third bootstrap-prefill rung so a reasoning-first
model continues inside the JSON envelope. - Paused-turn in-place resumes work again (2026-10-22 23:31 incident):
TurnIdentity.same_turncomparedturn_index, a reserved forensic join
key that a resumed turn re-reserves, so every in-place resume failed with
"Continuation belongs to a different turn identity or session plan". The
index is excluded; the fields that say whose turn it is stay decisive. - Rejected and off-turn dispatches are forensic evidence: pre-execution
rejections publishtool.telemetryrows before raising, slice-owned
operations stamp the session telemetry bus, forensic reads merge
errors.jsonl(ERROR+) withruntime.jsonl(WARNING+), and behavioral
signals detect repeated-failure crash-loops a per-turn view never sees. docs/SPEC/PROVIDER_ARGUMENT_NORMALIZATION.md— spec (no implementation)
for provider argument normalization at the gateway boundary.
Migration
No breaking changes; no migration steps required. Full details in
CHANGELOG.md
under [0.5.3].
Evidence
- Offline test + coverage gate: 4,628 passed, 3 skipped, 1 xfailed; all
nine package coverage floors met. Full OS (Ubuntu/macOS/Windows) ×
Python (3.11/3.12) matrix green on the tagged commit. - Ruff, mypy, direct-I/O policy, Vulture, import boundaries, documentation
gates clean. - All 20 pack validations from source passed; offline harness-smoke and
memory dedup-recall e2e scenarios passed. - Ten wheels + ten sdists built and Twine-checked; clean-environment import
and entry-point smoke tests passed. - No live provider or external-host compatibility testing was performed.
Install: pip install lisa-harness-runtime lisa-harness-sdk lisa-pack-memory — all ten distributions are on PyPI at 0.5.3.
v0.5.2
Added
- Relocatable profiles root:
LISA_PROFILES_DIRmoves the profiles root
(<profiles-dir>/<name>layout unchanged), so profiles can live in a
synced folder while the packs venv stays machine-local. A relocated
profile shares the base data dir's packs venv. Kernel fail-closed
hardening: missing lifecycle targets and tool errors under an abort
pre_turn_failure_policyfail the turn instead of proceeding. - Memory pack evals (
tests/evals/test_pack_validity.py,evalmarker) and
an offline scripted e2e scenario (tests/e2e/dedup_recall.scenario.yml,
registered ine2etest/packs.yml) that proves the store-then-recall flow
end to end with surface-scoped records.
Changed
- Memory capture is now a write-time upsert (breaking). The memory pack's
synthesize_and_store_learningandmine_thread_segmentoperations search
the scoped store for similar existing records before distilling and persist
a per-fact create/update/skip decision: updates write a superseding record
(supersedes_id) and retire the scoped active original in one transaction;
duplicates are skipped. Captures are stored asmemory_kind="semantic"—
distilled facts, not raw
conversation events. This replaces the batch
consolidate-episodic-pressure pipeline (the removed
consolidate_semanticoperation), aligning the write path with the
extract-then-dedup design current agent-memory systems converge on. context.memory.consolidation(MemoryConsolidationTuning) is removed
(breaking) and replaced bycontext.memory.mining
(MemoryMiningTuning):enabled,trigger,preferred_model_id,
mine_operation,max_segments_per_slice. Thread mining keeps running on
the slice provider under the new tuning.dispute_memoryperforms a real supersession: the correction carries the
supersedes_idcolumn link, inherits the disputed record's kind and scope,
and the original record is retired so it drops out of the active recall
pool (still queryable withstatus=Nonefor audit). Previously the
original stayed active at full salience forever.- Memory recall uses one shared tokenizer (
sdk.memory.lexical) across the
store's recall pools and the pack's evidence ranker, and the operation asks
the backend for candidate headroom above the final evidence limit so the
relevance-first re-ranker sees the full pool before truncating. - The memory-mining slice is fixed for production wiring: the slice-provider
view now carries the durable data root, active user, and model seams, and
the mining slice builds a realOperationExecutionContextinstead of a
bare params dict. Previously the factory leftmemory_data_dir=Noneand
the user slug defaulted, so mining and consolidation silently never ran. - The
mine.markersidecar is positional: a marker id missing from the
segment index (rebuilt store) re-mines from the start instead of silently
stopping forever — non-destructive because write-time dedup skips records
that already exist.
Fixed
- Thread mining leaves a segment pending after an operation exception or
memory-write failure, stopping the pass before a later segment can advance
the marker past it. A partial write is retried using write-time dedup. - Slice-provider content reads report an uninitialized content registry with
setup guidance instead of an attribute error. - Memory capture and mining log malformed dedup-search responses before
proceeding without candidates. - A failed supersession transaction rolls back its successor and retirement;
failed dispute writes report an error status instead of a successful result. - Supersession rejects a missing target ID that would make the successor
reference and retire itself. - Dispute and mining report the retirement performed by each superseding
write rather than assuming a predecessor was still active. Providers that
do not report the outcome leave it explicitly unknown. - File-backed approval and recovery stores fail closed on corruption. An
unreadable, malformed, or non-object store file, or an invalid record,
raisesApprovalStoreCorruptErrororRecoveryStoreCorruptErrorinstead of
loading as empty, so the next mutation can no longer overwrite pending
approvals or open recovery records.
Removed
MemoryRecallContext(unused) and the vestigial
used_semantic_search/auto_deep_used/uncertainty_note/
semantic_score/recency_scorerecall fields; the dead
_store_mined_factshelper; thedeeprecall parameter (the deep-synthesis
flow it described never existed).
Migration
- Replace
context.memory.consolidationwithcontext.memory.miningin
session configuration. Remove calls toconsolidate_semanticand the
deeprecall parameter. See
memory mining configuration. - External
memory.addproviders must atomically retire the scoped active
predecessor whensupersedes_idis supplied; capture, mining, and dispute
no longer issue a separatememory.retirecall. Superseding writes now
return a JSON metadata line containingmemory_idand boolean
supersession_retiredbefore the built-in server's existing text receipt.
External providers should return the same metadata so retirement can be
confirmed; see memory architecture. dispute_memory'sretiredresult can benullwhen the provider does not
confirm the outcome. Mining'sretired_idscontains only confirmed
retirements;not_retired_by_write_idsandretirement_unknown_idsreport
the other outcomes, andsupersession_outcomesassociates each outcome
with its predecessor and successor.- Capture results no longer expose
retired_ids; inspect the stored
successors'supersedes_idlinks when the prior IDs are needed. - A corrupt approval or recovery store file is no longer silently reset.
Repair or remove the file explicitly when
ApprovalStoreCorruptErrororRecoveryStoreCorruptErroris reported;
a missing file is still treated as empty.
Full Changelog: v0.5.1...v0.5.2
v0.5.1
What's Changed
- Test MCP capability and binary edge cases by @amigus in #12
- chore(release): prepare v0.5.1 by @amigus in #13
Full Changelog: v0.5.0...v0.5.1
v0.5.0
v0.4.0
v0.3.0
v0.2.0
[0.2.0] - 2026-10-24
Added
- Selectable MCP capability providers, including version-pinned binary servers
with checksum-verified downloads and caching. - Hosts can pass
surface_typeandsurface_idthroughbuild_runtimeand
supply extra context item builders tohandle_message. Pack-provided items
take precedence when item IDs collide. - End-to-end scenarios can declare surface identity and verify it through
the harness.
Fixed
- The stuck-in-stage guard counts only iterations in the current turn, not
earlier conversation turns in the same workflow mode. New turns and operator
resumes get a fresh stage budget; within-turn nudges and caps still apply. - Git subdirectory pack installs use the pack key for package-name policy
checks, including URLs with multiple fragment parameters. - In-process memory server initialization uses the default backup retention
when the shared argument namespace lacksbackup_keep.
Changed
- Catalog and authored-content adapters reuse long-lived stdio MCP sessions
instead of starting a Python server for every call. Framework setup prewarms
catalog and content concurrently; catalog invalidation also clears the
persistent server cache. Sessions close at process exit and can be closed
explicitly withclose_shared_sessions(). - A local startup profile measured catalog/content adapter time falling from
approximately 24.8 seconds to 4.4 seconds. These are diagnostic observations,
not an end-to-end latency guarantee; the model calls differed between runs. - Tracked follow-up: same-configuration framework rebuilds can retain
cached authored text after files change. The
prompt startup spec prioritizes refresh and
runtime-owned teardown before further in-process startup optimization. - Memory provider bindings now live in pack MCP manifests instead of runtime
aliases. Providers publish the stablememory.add,memory.search, and
memory.listcontract; additional memory tools are provider-dependent.
Upgrade notes
- Upgrade the framework distributions and bundled packs together to 0.2.0,
then restart the host. Hosts release separately; reinstalling an older
framework version does not include the turn-budget fix. - Alternate memory providers must declare the
memorycapability and its
required tools in their MCP manifest. Select the active provider with
providers.capabilities.memory; runtime-specific memory aliases are gone.
See the memory architecture. - New host context-builder and surface-identity parameters are optional.
Existing callers can omit them.