TB2 Kimi fixes: empty-turn synth, pre-submission prompt guidance, reasoning_content NUL sanitize#41

Open

n-WN wants to merge 25 commits into

masterfrom

n-WN/kimi-tb2-empty-turn-nul-sanitize

Commits on Apr 21, 2026

Kimi OAuth login + reflection completion-claim trigger + UTF-8 fix
n-WN
committed
reflection: actually fire completion-claim trigger (v1 patch was inert)
n-WN
committed
openai: call reflection on terminal turn so completion-claim trigger fires
n-WN
committed
reflection: drop keyword gate, use structural terminal-turn signal
n-WN
committed
openai: preserve reasoning_content on terminal-turn reflection hook
n-WN
committed
bash: flock dpkg frontend when the command invokes apt/dpkg
n-WN
committed
openai: restore Chat Completions reasoning_content round-trip for Kimi
n-WN
committed
reflection: adopt cc-src verificationAgent's adversarial nudge
n-WN
committed
reflection: stderr diagnostics on terminal-turn path
n-WN
committed
openai: proactive OAuth refresh before HTTP send
n-WN
committed
oauth+reflection: adopt disk creds + drop terminal-turn batch cooldown
n-WN
committed
reflection: show completion claim verbatim + skeptic criteria
n-WN
committed
oauth: temporarily disable disk-adopt path
n-WN
committed
openai: raise HTTP request timeout 300s → 900s for reasoning models
n-WN
committed
reflection: snap byte offsets to UTF-8 char boundaries
n-WN
committed
benchmark-run: add pre-submission verification guidance to system prompt
n-WN
committed
reflection: tell the LLM judge to think briefly
n-WN
committed
conversation: wire Read-tool image output into a vision content part
n-WN
committed
conversation: tests for Read-image → vision content part
n-WN
committed
reflection+benchmark: remove TB2 eval-path hints, cap reflection injections
n-WN
committed
reflection+benchmark: strip remaining TB2-specific path hints
n-WN
committed
reflection: stop the nudge from inviting filesystem probing
n-WN
committed

Commits on Apr 22, 2026

reflection: synth observation on empty terminal turn
n-WN
committed
benchmark: cleanup + python-interp guidance in pre-submission prompt
n-WN
committed
openai: sanitize NUL/C0 control bytes from reasoning_content
n-WN
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

TB2 Kimi fixes: empty-turn synth, pre-submission prompt guidance, reasoning_content NUL sanitize#41

TB2 Kimi fixes: empty-turn synth, pre-submission prompt guidance, reasoning_content NUL sanitize#41
n-WN wants to merge 25 commits into
masterfrom
n-WN/kimi-tb2-empty-turn-nul-sanitize

Commits on Apr 21, 2026

Kimi OAuth login + reflection completion-claim trigger + UTF-8 fix

reflection: actually fire completion-claim trigger (v1 patch was inert)

openai: call reflection on terminal turn so completion-claim trigger fires

reflection: drop keyword gate, use structural terminal-turn signal

openai: preserve reasoning_content on terminal-turn reflection hook

bash: flock dpkg frontend when the command invokes apt/dpkg

openai: restore Chat Completions reasoning_content round-trip for Kimi

reflection: adopt cc-src verificationAgent's adversarial nudge

reflection: stderr diagnostics on terminal-turn path

openai: proactive OAuth refresh before HTTP send

oauth+reflection: adopt disk creds + drop terminal-turn batch cooldown

reflection: show completion claim verbatim + skeptic criteria

oauth: temporarily disable disk-adopt path

openai: raise HTTP request timeout 300s → 900s for reasoning models

reflection: snap byte offsets to UTF-8 char boundaries

benchmark-run: add pre-submission verification guidance to system prompt

reflection: tell the LLM judge to think briefly

conversation: wire Read-tool image output into a vision content part

conversation: tests for Read-image → vision content part

reflection+benchmark: remove TB2 eval-path hints, cap reflection injections

reflection+benchmark: strip remaining TB2-specific path hints

reflection: stop the nudge from inviting filesystem probing

Commits on Apr 22, 2026

reflection: synth observation on empty terminal turn

benchmark: cleanup + python-interp guidance in pre-submission prompt

openai: sanitize NUL/C0 control bytes from reasoning_content