Preserve reasoning and replay tool history across turns #2470
Preserve reasoning and replay tool history for multi-turn tool calling
4e9a80ef
Measure streaming TTFT and throughput in the replay harness
f8865aa3
Harden tool history replay: provider-safe ids, honest interrupts, siz…
ce6265ad
Extend reasoning_content replay to the plain completion flow
56382cfe
gary149
changed the title Preserve reasoning and replay tool history across turns in the MCP flow Preserve reasoning and replay tool history across turns 21 days ago
Clarify why replayed reasoning attaches to the final message
c04f2e9c
Make replay budget degradation monotonic
2f773dd3
Attribute replayed reasoning to its tool round; honor user reasoning …
d7fcbfb4
Accept reasoning_text as a third incoming reasoning field
24ec8592
Flag reasoning support for DeepSeek-V4-Flash and Qwen3.6 models
2ba08f81
Teach sync-models that supportsReasoning also gates reasoning replay
e1c5f1dc
Fix reasoning byte fidelity, preamble attribution, and budget-fallbac…
20407036
Fix preamble duplication on replay and whitespace loss in reasoning s…
120ffdb1
Strip Gemma-flow think leak, gate cross-producer reasoning, persist r…
2530e4f1
Fix leading-whitespace reasoning loss, stale-producer reasoning after…
64425f07
Omit empty replayed final messages; fail the harness when the nonce r…
d8a97004
Scope round-reasoning dedup to a positional match; omit phantom empty…
05aa6162
Prefix-only preamble dedup, empty-flat guard, per-rep coherence gate
7e0c7a24
pngwn
commented
on 2026-08-03
Merge remote-tracking branch 'origin/main' into preserve-reasoning-to…
bc8e231e
test: end-to-end harness for reasoning + tool-history replay
831d0f1b
test: re-scope replay findings as characterisation tests
24e87797
Apply the whitespace-reasoning rule to the tool-less flow too
de3a0399
Backfill producer metadata without replacing the message array
3d798dce
Lock in that recording a producer never moves the pinned model
9ffae6d2
Budget the whole outgoing history, not just the replayed part
36e75a3e
Fail the harness on a successful response with no answer
5eead9e0
Bound the history budget by the model's context window
14d9ce81
Gate the in-loop reasoning echo on the same capability as replay
1354c4ea
Follow the production gate in the in-loop harness scenario
1c5bcea5
Merge branch 'main' into preserve-reasoning-tool-replay
f6cfcda8
Share the model-switch backfill with the v2 API endpoint
1c6e3e7c
Reserve the model's configured reply allowance, not a constant
8879bd0c
Default reasoning replay on, with a blocklist for the exceptions
2fa0ef1b
abidlabs
approved these changes
on 2026-08-11
pngwn
merged
108b97bc
into main 9 days ago
pngwn
deleted the preserve-reasoning-tool-replay branch 9 days ago
Assignees
No one assigned
Login to write a write a comment.
Login via GitHub