fix: LangChain/LangGraph parity audit — 40+ correctness fixes across session, cache, tool calling, reasoning, agent loop and local models - #95

Merged

senamakel merged 178 commits into

mainfrom

tinyagents-langchain-parity

Aug 8, 2026

Commits on Aug 8, 2026

feat(session): versioned schema migrations, busy_timeout, missing indexes

senamakel
and
medullabot
committed
62cff37
fix(retry): additive jitter, backoff-on-by-default, Retry-After, retry_on predicate

senamakel
and
medullabot
committed
db37294
feat(events): failure-path variants and a panic-safe dispatch guard

senamakel
and
medullabot
committed
70f0c80
feat(limits): fail-closed cap reconciliation and stop-with-partial exhaustion

senamakel
and
medullabot
committed
90fc612
fix(harness): pairing-safe trimming, tool-aware token estimates, tool-history summaries, and tool-result artifacts

senamakel
and
medullabot
committed
cc3a177
fix(steering): validate batches atomically and latch pauses across checkpoints

senamakel
and
medullabot
committed
b5b6c3c
fix(session): guard the task-claim CAS on status, keep upsert read-backs transactional

senamakel
and
medullabot
committed
69339e4
feat(session): atomic FTS indexing, retention/pruning entry points, FTS reindex

senamakel
and
medullabot
committed
f1231a0
feat(checkpoint): put_writes/get_writes protocol, durable file writes, safe lineage walks

senamakel
and
medullabot
committed
47762fa
feat(harness/tool): unique synthetic call ids, callable schema projection, injected args, per-tool error policy

senamakel
and
medullabot
committed
7ec1b94
feat(stream,no-progress): event projection and a drivable no-progress API

senamakel
and
medullabot
committed
f071be0
chore: files changed src/harness/embeddings/cloud.rs,src/harness/embeddings/cohere.rs,src/harness/em

senamakel
and
medullabot
committed
8328c9d
test(context): add tests for context and schema tool surface

senamakel
and
medullabot
committed
b7c0118
test(runtime): add resilience tests for runtime primitives

senamakel
and
medullabot
committed
63457f1
test(persistence): add conformance test for persistence module

senamakel
and
medullabot
committed
4bfb2c6
test(persistence): update import path for conformance test helpers

senamakel
and
medullabot
committed
48b3ef4
fix(checkpoint): disable cycle detection and skip thread existence check

senamakel
and
medullabot
committed
01a3bd8
test: public-API regression suite for the runtime primitives

senamakel
and
medullabot
committed
3a5ec45
test(context_and_schema_tool_surface): reformat artifact assertion

senamakel
and
medullabot
committed
f3fb9ce
fix(steering): clamp steering angle to valid range

senamakel
and
medullabot
committed
63a75a5
fix(steering): pass current checkpoint to pause commands

senamakel
and
medullabot
committed
7f9ba12
test(persistence): add session persistence tests

senamakel
and
medullabot
committed
d2c8776
test(persistence): extract test helpers for team and task creation

senamakel
and
medullabot
committed
6fef42d
test(feature_infra_resilience): add tests for infrastructure resilience

senamakel
and
medullabot
committed
6442424
test(e2e): add public API contract tests

senamakel
and
medullabot
committed
84558b1
test(persistence_store): add tests for persistence store

senamakel
and
medullabot
committed
b65e891
chore(store): add namespaced types module

senamakel
and
medullabot
committed
d4f859b
chore(tests): reformat long lines to comply with style guide

senamakel
and
medullabot
committed
7c62890
fix(checkpoint): correct cycle detection and thread copy logic

senamakel
and
medullabot
committed
f87dcc8
fix(store): handle missing namespace in namespaced store

senamakel
and
medullabot
committed
809bb47
docs(store): document the new namespaced store module

senamakel
and
medullabot
committed
7284cb2
fix(openai): add CacheTokenAccounting to stream accumulator and parse_chat_response calls

senamakel
and
medullabot
committed
ff537b6
fix(openai): handle empty response body in provider

senamakel
and
medullabot
committed
c1f5ff4
refactor(openai): remove unused usage conversion and probe result alias

senamakel
and
medullabot
committed
3d06bf4
chore: files changed src/harness/providers/openai/test.rs

senamakel
and
medullabot
committed
ef72af1
fix(openai): handle empty tool call arguments

senamakel
and
medullabot
committed
e4a3e1c
refactor(openai): unify schema preparation for response format and tools

senamakel
and
medullabot
committed
27d0fa6
fix(openai): handle empty assistant message content

senamakel
and
medullabot
committed
aa06395
chore(tests): update doc references to renamed RunPolicy field

senamakel
and
medullabot
committed
523339e
fix(tests): update provider local wire test assertions

senamakel
and
medullabot
committed
e68d323
chore: format test files with rustfmt

senamakel
and
medullabot
committed
60037db
fix(harness): replace manual index with enumerate and use const assertion

senamakel
and
medullabot
committed
3837b55
feat(openai): add responses provider harness

senamakel
and
medullabot
committed
d6a15ee
fix(test): record request before writing reply in mock server

senamakel
and
medullabot
committed
537d1b6
refactor(session): remove unnecessary Ok wrapping in prune functions

senamakel
and
medullabot
committed
0ec000d
docs(checkpoint): document performance characteristics of scoped lookups and state_history

senamakel
and
medullabot
committed
548d130
chore(session): make busy-timeout guarantee explicit

senamakel
and
medullabot
committed
32e039e
docs(session): correct the busy-timeout test rationale to match the verified rusqlite default

senamakel
and
medullabot
committed
99d7f18
chore(cache): add types module for harness cache

senamakel
and
medullabot
committed
ae5c01d
chore(cache): use stable hash for cache keys

senamakel
and
medullabot
committed
cabab11
chore(harness): add context type definitions

senamakel
and
medullabot
committed
0012f71
chore(harness): add context types module

senamakel
and
medullabot
committed
4d31c13
chore(harness): remove unused context module

senamakel
and
medullabot
committed
c1fb635
chore(harness): remove unused context module

senamakel
and
medullabot
committed
ffb8eba
chore(cache): derive Hash for cache key struct

senamakel
and
medullabot
committed
a0a9c8b
chore(harness): remove unused context module

senamakel
and
medullabot
committed
488a0ba
test(harness): update run config tests for optional call caps

senamakel
and
medullabot
committed
fd4f8b2
fix(agent_loop): handle empty agent output gracefully

senamakel
and
medullabot
committed
440e03e
fix(harness): correct agent loop cache layout handling

senamakel
and
medullabot
committed
665eda9
perf(cache): reduce memory cache lock contention

senamakel
and
medullabot
committed
fa33922
fix(agent_loop): handle tool call errors gracefully

senamakel
and
medullabot
committed
8c5929b
fix(runtime): make runtime types Send and Sync

senamakel
and
medullabot
committed
23ee1a7
fix(runtime): make runtime types Send and Sync

senamakel
and
medullabot
committed
970afa6
fix(harness): use sqlite cache for test results

senamakel
and
medullabot
committed
b6cdf79
fix(agent_loop): handle tool call errors gracefully

senamakel
and
medullabot
committed
14da7e7
chore(cache): add singleflight deduplication for cache loads

senamakel
and
medullabot
committed
3850612
chore(cache): add cache module

senamakel
and
medullabot
committed
7c79e60
chore(model): derive common traits for harness types

senamakel
and
medullabot
committed
5a2ec88
fix(agent_loop): preserve partial transcript on run failure

senamakel
and
medullabot
committed
6e2269a
fix(agent_loop): handle empty run loop gracefully

senamakel
and
medullabot
committed
d521b5b
fix(harness): use agent loop for model execution

senamakel
and
medullabot
committed
28ff4de
chore(model): derive common traits for model types

senamakel
and
medullabot
committed
81e39f9
fix(agent_loop): handle empty agent output gracefully

senamakel
and
medullabot
committed
efcaf45
fix(agent_loop): handle empty agent output gracefully

senamakel
and
medullabot
committed
3dc608d
fix(agent_loop): handle tool call errors gracefully

senamakel
and
medullabot
committed
7b563bd
fix(agent_loop): handle tool call errors gracefully

senamakel
and
medullabot
committed
6b0f384
fix(agent_loop): handle empty agent output gracefully

senamakel
and
medullabot
committed
9eef9d2
fix(agent_loop): handle tool call errors gracefully

senamakel
and
medullabot
committed
874cd43
fix(agent_loop): handle empty agent output gracefully

senamakel
and
medullabot
committed
003a5c6
fix(agent_loop): handle empty agent output gracefully

senamakel
and
medullabot
committed
48b6938
fix(agent_loop): handle tool call errors gracefully

senamakel
and
medullabot
committed
d3596f8
fix(agent_loop): handle empty model responses

senamakel
and
medullabot
committed
79b3f9c
feat(agent_loop): add LoopExit enum for deliberate loop stops

senamakel
and
medullabot
committed
8580233
refactor(agent_loop): simplify middleware type handling

senamakel
and
medullabot
committed
4c67a2a
fix(agent_loop): handle empty model response

senamakel
and
medullabot
committed
8c90fda
fix(agent_loop): handle empty entry list gracefully

senamakel
and
medullabot
committed
e88cb9b
fix(agent_loop): handle empty model response

senamakel
and
medullabot
committed
e4aea6c
fix(agent_loop): handle empty entry list gracefully

senamakel
and
medullabot
committed
1b79fbc
refactor(agent_loop): simplify model call and tool handling

senamakel
and
medullabot
committed
986a94a
chore(agent_loop): rename types module for clarity

senamakel
and
medullabot
committed
a074996
fix(agent_loop): handle tool call errors gracefully

senamakel
and
medullabot
committed
163a01a
fix(agent_loop): handle empty model response

senamakel
and
medullabot
committed
630fc42
fix(agent_loop): handle empty model responses

senamakel
and
medullabot
committed
f81a397
fix(agent_loop): handle empty agent output gracefully

senamakel
and
medullabot
committed
e8f3fe2
fix(agent_loop): track emitted deltas across streaming retries

senamakel
and
medullabot
committed
e6a809b
fix(agent_loop): handle empty model responses

senamakel
and
medullabot
committed
1b62e82
fix(agent_loop): handle empty model responses

senamakel
and
medullabot
committed
826779d
chore(context): remove unused context middleware

senamakel
and
medullabot
committed
41c8a4d
fix(agent_loop): handle empty model response

senamakel
and
medullabot
committed
269c472
test(wave2): add execution tests for wave2 tools

senamakel
and
medullabot
committed
a33af39
fix(harness): enforce budget limits in middleware library

senamakel
and
medullabot
committed
9a02bf5
test(wave2_tools_execution): use push_middleware and trim unused imports

senamakel
and
medullabot
committed
596c7c1
fix(openai): handle empty tool call arguments

senamakel
and
medullabot
committed
1b9b466
chore(harness): remove unused context middleware

senamakel
and
medullabot
committed
9deb1af
fix(openai): handle empty tool call arguments

senamakel
and
medullabot
committed
130e311
fix(openai): handle empty tool call arguments

senamakel
and
medullabot
committed
5f19989
test(wave2): update tool execution tests for new context API

senamakel
and
medullabot
committed
47c44b2
fix(harness): retry transient failures with backoff

senamakel
and
medullabot
committed
d71a43a
test(wave2_loop_limits): add tests for loop limit edge cases

senamakel
and
medullabot
committed
3c48e8d
fix(tests): use default run limits in concurrent admission test

senamakel
and
medullabot
committed
bef0ca3
feat(harness): add cache metadata to model responses

senamakel
and
medullabot
committed
11f6bc9
test(wave2_loop_structured): add structured loop test

senamakel
and
medullabot
committed
7c13985
fix(error): include source error in display output

senamakel
and
medullabot
committed
0db70a4
test: update test fixtures for new model response fields

senamakel
and
medullabot
committed
c53f8a2
test(agent_loop): complete cache policy in test request

senamakel
and
medullabot
committed
8c0f164
test(wave2): add loop recovery test

senamakel
and
medullabot
committed
03e3062
test(e2e_observability): add served_from_cache field to test responses

senamakel
and
medullabot
committed
464154a
refactor(harness): rename agent loop tests and update provider types

senamakel
and
medullabot
committed
7b756d3
fix(model): remove stray field from ModelResponse

senamakel
and
medullabot
committed
937a21b
fix(repair): handle missing parent nodes during repair

senamakel
and
medullabot
committed
c7c1462
fix(harness): remove stray retry_after_ms fields

senamakel
and
medullabot
committed
3fe9a7c
fix(test): remove invalid field from test model

senamakel
and
medullabot
committed
f428a36
fix(validate): allow empty structured values

senamakel
and
medullabot
committed
25b0f84
test: rework wave2 loop estimator tests around shared token counting

senamakel
and
medullabot
committed
47f89d0
chore(types): add structured type definitions

senamakel
and
medullabot
committed
92fbc38
test: remove unused retry_after_ms field from test fixtures

senamakel
and
medullabot
committed
e7a5975
chore(types): add missing Debug derive to structured types

senamakel
and
medullabot
committed
8f4d6ba
chore(types): add structured type definitions

senamakel
and
medullabot
committed
a6d30bb
chore(harness): remove unused structured module

senamakel
and
medullabot
committed
df35356
chore(harness): remove unused structured module

senamakel
and
medullabot
committed
a6c3031
fix(harness): correct loop control and limit test expectations

senamakel
and
medullabot
committed
6e36bc5
chore(harness): remove unused structured module

senamakel
and
medullabot
committed
97c9a81
test(wave2): add loop control tests

senamakel
and
medullabot
committed
a3d7c94
feat(harness): repair invalid tool-call arguments during extraction

senamakel
and
medullabot
committed
25fe0ea
test(wave2): add cache key scope tests

senamakel
and
medullabot
committed
5681b8e
test(wave2_tools_structured): add structured output tests

senamakel
and
medullabot
committed
7336f44
test(wave2_tools_structured): update provider error construction in tests

senamakel
and
medullabot
committed
8aaced4
test(wave2_cache_loop): add test for cache loop behavior

senamakel
and
medullabot
committed
8c881bd
fix(error): include source error in display output

senamakel
and
medullabot
committed
b27e773
test: update cache loop tests for new RunPolicy API

senamakel
and
medullabot
committed
893f92e
test(wave2_loop_structured): enable tool calling in test model profile

senamakel
and
medullabot
committed
573093c
fix(agent_loop): handle empty agent output gracefully

senamakel
and
medullabot
committed
eaab3a5
test(wave2_cache_store): add tests for cache store behavior

senamakel
and
medullabot
committed
9f7d720
test(wave2_loop_structured): clarify default profile comment

senamakel
and
medullabot
committed
e00c834
test(wave2_cache_layout): add cache layout tests

senamakel
and
medullabot
committed
fb69ab5
test(wave2): add structured feature harness and cache retry-after tests

senamakel
and
medullabot
committed
ef458d2
test(agent_loop): pin fail-closed behavior in tool error tests

senamakel
and
medullabot
committed
247f538
test(openai): add retry-after header parsing tests

senamakel
and
medullabot
committed
0c873c0
test: derive expected token counts from shared estimator

senamakel
and
medullabot
committed
515e016
chore(cache): add singleflight deduplication for cache misses

senamakel
and
medullabot
committed
257017d
chore(cache): add singleflight deduplication for cache misses

senamakel
and
medullabot
committed
1b8d65f
fix(harness): correct tool-call identity, started/terminal pairing, error policy and admission ordering in the agent loop's tool path

senamakel
and
medullabot
committed
f16cf75
style: apply rustfmt formatting to structured harness code

senamakel
and
medullabot
committed
d3c6468
test(wave2): clarify retry-after fallback test messages

senamakel
and
medullabot
committed
d66b75f
chore: apply rustfmt formatting across cache and loop modules

senamakel
and
medullabot
committed
e6bdfd9
test(feature_infra_resilience): add retry_after_ms field to test fixtures

senamakel
and
medullabot
committed
1938bbc
test(e2e): opt into fail-closed policies in middleware and unknown-tool tests

senamakel
and
medullabot
committed
3501af9
test(e2e_middleware): import InvalidArgsPolicy in test

senamakel
and
medullabot
committed
88a3d34
chore(harness): update cache docs and context middleware

senamakel
and
medullabot
committed
cc42044
docs(harness): document cache module

senamakel
and
medullabot
committed
057dd4f
fix(agent_loop): handle empty agent response gracefully

senamakel
and
medullabot
committed
2cfe7af
fix(harness): enforce budget limits in middleware library

senamakel
and
medullabot
committed
81b30d9
fix(agent_loop): handle empty agent output gracefully

senamakel
and
medullabot
committed
5c138e2
test(wave2): add loop cache accounting tests

senamakel
and
medullabot
committed
2b65e28
fix(tests): update snapshot usage field access

senamakel
and
medullabot
committed
d870ed1
test(wave2_loop_cache_accounting): reformat model registration

senamakel
and
medullabot
committed
03ee52e
test(wave2_cache_layout): add cache layout tests

senamakel
and
medullabot
committed
594a295
chore(harness): add middleware types module

senamakel
and
medullabot
committed
a86a27e
chore(context): remove unused context middleware

senamakel
and
medullabot
committed
fc26013
chore(context): remove unused context middleware

senamakel
and
medullabot
committed
4332785
chore(harness): add middleware types module

senamakel
and
medullabot
committed
95d3447
test(live_local_models): add tests for local model loading

senamakel
and
medullabot
committed
0ebf35f
test(live_local_models): add tests for local model loading

senamakel
and
medullabot
committed
d3acf7f
test(cache): pin run-scoped prompt-cache guard baseline

senamakel
and
medullabot
committed
b7d5c76
test(wave2_cache_store): add tests for cache store behavior

senamakel
and
medullabot
committed
283af0e
test(wave2_cache_store): add tests for cache store edge cases

senamakel
and
medullabot
committed
c9e2330
Merge remote-tracking branch 'refs/remotes/upstream/main' into pr/95
senamakel
committed
b1582f5
Fix PR review findings
senamakel
committed
5877372

Uh oh!

Commits

Commits on Aug 8, 2026