Cooperative is the sole production Chat orchestrator. The default remains
professional; COOPERATIVE_WORK_PLANNER_ENABLED=false selects basic. A failed
decision never falls back to a different strategy. Host permissions, durable
Work identity, replay fences and execution acceptance remain authoritative.
core/chat_runtime.py now contains shared presentation only: text filtering,
sentence splitting, expressions and TTS delivery. It cannot generate a turn or
dispatch a DELEGATE. Original Chat generation, its separate adjudicator,
decomposition query, recovery/resend chain and direct Browser/AUIP entry points
are retired. Shared source/reference validation and parse_decomposition_reply
remain; professional decoding still hands off at the first DELEGATE.
COOPERATIVE_CHAT_ENABLED is a retired setting. Explicit false produces a
warning and starts Cooperative. The old Off value selected Original; it never
prohibited Work. Existing Work permission policy is unchanged.
Desktop Settings keeps a saved false through ordinary reads, unrelated saves
and backup recovery. Its migration notice explains the change; confirmation
durably deletes the old key. A failed save retains the notice. Environment or
.env values are read-only explanations: remove the obsolete assignment from
that source and restart. The backend exposes these non-sensitive facts through
retired_settings; they do not authorize execution.
The transitional reader is retained for the first release containing this convergence. Before removing it in a later release, retain a file-format migration for users skipping releases. Rollback requires the previous software version; there is no hidden Original runtime switch in this version. Existing historical authority records are retained and do not become fresh effect grants.
After acknowledging the desktop migration, rolling back does not restore the deleted false value: the older version uses its default until you explicitly set its old route option. Do not restore an old settings file over later edits or an old database snapshot over newer receipts. Stop ingress and drain/stop owned work before a version change; an unknown Provider state remains unknown. Storage compatibility checks cover only the recorded current/baseline commits on test copies, not arbitrary releases or external process recovery.
COOPERATIVE_CHAT_TEMPERATURE=0.7 is the default for Main Chat role replies,
including replies to Provider updates and Host receipts. Set it in the process
environment or repository .env, then restart the backend. Values must be
finite and between 0 and 2; the selected model may have narrower limits.
Set 0 explicitly to restore the previous sampling behavior. Planner, reference
resolution and structured control queries remain at 0.0; Hybrid local-head
sampling is independent. DeepSeek, Gemini, Bedrock and local HTTP adapters forward
this value. The current OpenAI / Hybrid3 adapter omits temperature, so this
setting does not take effect on that path. This setting does not change history retention.
COOPERATIVE_CHAT_TEMPERATURE 默认 0.7，用于主对话角色回复，包含对 Provider
更新及 Host 回执的回复。在进程环境或仓库 .env 中设置后，需重启后端。值必须为
0 到 2 之间的有限数值；所选模型可能有更窄的限制。显式设为 0 可恢复之前的采样
行为。规划、引用解析和结构化控制查询保持 0.0，Hybrid 本地首句采样独立，
DeepSeek、Gemini、Bedrock 和本地 HTTP 适配器会传递该值；当前 OpenAI / Hybrid3
适配器不发送 temperature，因此此设置在该路径不生效。历史保留策略不变。
For pure-local Chat, use LOCAL_LLM_TYPE=llama_server. The managed default
context is 16384; an explicitly saved 4096 is preserved and may need increasing.
The messages boundary measures the actual server slot and rendered input plus
output budget, and rejects insufficient capacity before generation. Larger
context consumes additional device memory.
The managed Hybrid launcher keeps its 4096 default. An explicit
LOCAL_LLM_CLI_CONTEXT overrides both launch profiles; leave it unset to use
their separate defaults. The pure-local llama_server backend requires the
server's /props, /apply-template and /tokenize APIs, in addition to chat
completions. Older llama.cpp builds or OpenAI-only proxies without those APIs
are rejected because their capacity cannot be verified before generation.
Persistent cli sessions cannot satisfy independent messages requests and are
rejected with a migration hint. Point the executable setting at llama-server
from the same llama.cpp package. LM Studio and Ollama endpoint capacity and
overflow behavior remain pending live acceptance; no endpoint was available
in this environment.
hybrid* now actually requests its configured local HTTP head plus the existing
remote role model. This is an intentional behavior change for users who had
selected Hybrid while Cooperative previously used only the remote model. No new
response-mode setting is introduced. Head configuration is independent of
LOCAL_LLM_TYPE; the existing Hybrid BAT launch profile is preserved.
The delivery lock stays inside the user turn. Streamed remote replies, and non-streamed replies with no Host action, share one history entry with the head. For a non-streamed Host action the head is committed separately and the lock is released before the action. An already-ready remote reply cancels an unused head; pending choices never inherit its lock.
RAG uses the shared process index and adds a per-turn reference to the role system prompt. Display/voice formatting cannot create execution authority.
The linked evidence records its tested revisions and retained failures. Historical routing scores remain historical, including results for the now-retired Original route.
The corrected 30-pair ordinary Chat comparison measured send-to-first-text medians of 1.039 s before and 1.047 s after the capability migration. This is a pre-retirement comparison, not a new performance claim for every later edit. The earlier 3.9 s result included input-readiness wait and is invalid for this metric. First model content, first Chat text and first audio-device write are different boundaries; device writes do not prove acoustic arrival.
External local endpoints and acoustic loopback remain unverified. The full current routing/persona corpus was observed in a new controlled context; results retain all mismatches. Local heads now retain a complete sentence across commas, but copied input language and repetitive openings remain model-quality limitations. No extra model pass or language classifier was added. The user explicitly deferred model quality tuning from this retirement after reviewing these observations. A bounded pure-local Host probe also retained basic's missed task handoff and professional's unrequested desktop target; neither is relabeled as a fully accepted local model outcome. See the quality disposition and local Host evidence. AUIP J7 uses the shipping entry with a seeded application and real action receipts. Parent review found wrong actor/winner narration after resignation; the candidate catalog now preserves declared action semantics, and B2 receives the original user wording instead of a lossy derived instruction. Both strategies' repeated journeys correctly attributed resignation and the winner through follow-up dialogue. See the bounded repair evidence. Gomoku does not cover open-payload actions or continuous controllers. Separate bounded journeys now cover both, with the open case explicitly using a declaration variant of an existing app. Local cancellation and fresh-loop checks cover transport lifecycle, not durable desktop restoration or general model quality. The old private-runtime harness is not retained as a compatibility path.
Browser is limited to regression observation and issue recording in this retirement, by the user's explicit scope decision. The live cross-domain journey retained seven failed checks; it is not marked accepted. The workspace-less OpenClaw amendment restriction predates this work; other Project/steering mismatches remain unattributed. They are deferred for the future Browser decision, while regressions proven to be introduced by this convergence remain in scope. See the retained observations and baseline evidence.
Real native Codex approval was checked in both directions: allow continued the same run and wrote the requested file; deny durably recorded refusal and left the target absent. Both closed the permission card and retained one WorkItem and Attempt. These are direct provider/Host checks, not GUI permission-button or natural-language routing checks. See permission evidence.
A real interruption/restart journey preserved failed status, workspace and focus across backend restart. Explicit retry added one Attempt to the same WorkItem and produced the exact recovery artifact; read-only queries added no Work. This is bounded headless professional/DeepSeek/Codex evidence, with TTS off.
Provider handoff keeps the selected operation as the assigned task. The earlier experiment that replaced its source evidence with the full message was withdrawn from retirement. A separately reviewed handoff change retains the selected source and additionally quotes the complete current message as context for applicable constraints. It does not expand Host task, target or permission authority; its bounded scope evidence is recorded separately.