Add default-off GPT-5.6 Ultra agent mode

CHANGELOG.md

@@ -4,15 +4,15 @@ All notable changes to this project will be documented in this file.
 
 ## Unreleased
 
+## 1.9.0 - 2026-07-10
+
 - Refactored config precedence to keep explicit runtime mode authoritative and treat spoof mode as compatibility fallback.
 - Hardened merged auth state active-account selection to avoid disabled active identity carry-over.
 - Simplified request payload transform wrappers to route through one shared aggregate transform pipeline.
 - Consolidated account action messaging with shared builders and tightened auth-menu wording consistency.
 - Consolidated model-catalog stale-cache fallback emission flow and removed small dead helper modules.
-- Added experimental Codex collaboration profile gates (`runtime.collaborationProfile`, `runtime.orchestratorSubagents`) for plan/orchestrator parity.
-- Collaboration features now auto-enable by default in `runtime.mode="codex"` and can be explicitly enabled/disabled in any mode.
-- Added managed `orchestrator` agent template sync under `~/.config/opencode/agents`, with visibility gated by effective collaboration profile (mode-derived by default, explicitly overridable).
-- Synced pinned upstream Codex orchestrator + plan templates into a local prompt cache (ETag/304-aware, TTL refreshed) and used the cached plan prompt to populate plan-mode collaboration instructions.
+- Retired the earlier collaboration-profile/orchestrator WIP, including managed prompt sync, collaboration headers, and generated `orchestrator.md` agents.
+- Added the replacement GPT-5.6 Ultra agent mode as a WIP behind `runtime.ultra`, defaulting to `false` with provider variants and delegation policy hidden until explicitly enabled.
 - Added configurable `runtime.promptCacheKeyStrategy` (`default` | `project`) for session-based or project-path-based prompt cache keying.
 - Added quota threshold warnings at `25%`, `20%`, `10%`, `5%`, `2.5%`, `0%` and automatic cooldown/switch when `5h` or `weekly` quota is exhausted.
 - Added account selection tracing and per-attempt failover reason codes for snapshot/debug observability.

docs/configuration.md

@@ -62,6 +62,7 @@ Known-field type validation is applied on load. If a known field has an invalid
     "headerSnapshots": false,
     "headerSnapshotBodies": false,
     "headerTransformDebug": false,
+    "ultra": false,
     "pidOffset": false
   },
   "global": {
@@ -77,8 +78,7 @@ Known-field type validation is applied on load. If a known field has an invalid
 Mode-derived runtime defaults when omitted:
 
 - `runtime.codexCompactionOverride`: `true` in `codex`, `false` in `native`
-- `runtime.collaborationProfile`: `true` in `codex`, `false` in `native`
-- `runtime.orchestratorSubagents`: inherits effective `runtime.collaborationProfile`
+- `runtime.ultra`: `false` in every mode
 
 ## Settings reference
 
@@ -134,14 +134,10 @@ Mode-derived runtime defaults when omitted:
   - Adds explicit `before-header-transform` and `after-header-transform` request snapshots for message fetches.
 - `runtime.pidOffset: boolean`
   - Enables session-aware offset behavior for account selection.
-- `runtime.collaborationProfile: boolean`
-  - Experimental: enables Codex-style collaboration mode mapping from agent names (`plan` -> plan mode, `orchestrator` -> code mode profile).
-  - If omitted, defaults to `true` in `runtime.mode="codex"` and `false` otherwise.
-  - Explicit `true`/`false` works in any mode.
-- `runtime.orchestratorSubagents: boolean`
-  - Experimental: enables Codex-style subagent header hints for helper agents under collaboration profile mode.
-  - If omitted, inherits `runtime.collaborationProfile` effective value.
-  - Explicit `true`/`false` works in any mode.
+- `runtime.ultra: boolean`
+  - Work in progress. Enables the catalog-gated Ultra agent mode.
+  - Defaults to `false`; it must be explicitly enabled in any runtime mode.
+  - When disabled, the `ultra` picker variant is hidden and no delegation policy is injected. Stale literal `ultra` inputs still degrade safely to wire effort `max`.
 
 ### Model behavior
 
@@ -158,6 +154,11 @@ Mode-derived runtime defaults when omitted:
   - Global reasoning effort override forwarded upstream when the request does not already set one.
   - When omitted, the selected model's live catalog `default_reasoning_level` is used, typically `"medium"`.
   - User config can still override reasoning effort globally, per model, or per variant.
+- `ultra` reasoning variant
+  - Work in progress and available only when `runtime.ultra=true` and the active model advertises `ultra` with `multi_agent_version: "v2"`.
+  - `codex` mode adds best-effort proactive delegation guidance; `native` mode preserves OpenCode-native prompt identity.
+  - Literal configured `ultra` values remain safe on unsupported or stale catalogs: the backend request sends wire effort `max`, without proactive delegation.
+  - There is no public concurrency setting; OpenCode remains responsible for agent execution and lifecycle.
 - `global.reasoningMode: "standard" | "pro"` (optional)
   - GPT-5.6 reasoning mode, emitted as `reasoning.mode` independently of `reasoning.effort`.
   - An explicit request value is preserved. The same per-model and per-variant precedence applies.
@@ -331,8 +332,7 @@ Advanced path:
 - `OPENCODE_OPENAI_MULTI_HEADER_SNAPSHOTS`: `1|0|true|false`.
 - `OPENCODE_OPENAI_MULTI_HEADER_SNAPSHOT_BODIES`: `1|0|true|false`.
 - `OPENCODE_OPENAI_MULTI_HEADER_TRANSFORM_DEBUG`: `1|0|true|false`.
-- `OPENCODE_OPENAI_MULTI_COLLABORATION_PROFILE`: `1|0|true|false`.
-- `OPENCODE_OPENAI_MULTI_ORCHESTRATOR_SUBAGENTS`: `1|0|true|false`.
+- `OPENCODE_OPENAI_MULTI_ULTRA`: `1|0|true|false` (WIP; defaults to false).
 
 ### Debug/OAuth controls
 
@@ -362,18 +362,6 @@ Legacy behavior keys are no longer parsed from `codex-config.jsonc`.
 
 Use canonical `global` and `perModel` keys only.
 
-## Managed prompts and orchestrator agent
+## Legacy orchestrator cleanup
 
-The plugin synchronizes a pinned upstream Codex orchestrator prompt and plan-mode prompt into a local cache under the resolved config cache root (`$XDG_CONFIG_HOME/opencode/cache/` when `XDG_CONFIG_HOME` is set, otherwise `~/.config/opencode/cache/`):
-
-- `codex-prompts-cache.json`
-- `codex-prompts-cache-meta.json` (stores URLs, `lastChecked`, and ETags)
-
-Fetch behavior:
-
-- TTL-based refresh (best-effort; normal requests continue if refresh fails)
-- ETag-based revalidation (`If-None-Match` + `304 Not Modified`)
-
-The plan prompt from this cache is used to populate plan-mode collaboration instructions.
-
-When `runtime.collaborationProfile` is enabled, the installer and plugin startup also manage the visibility of an `orchestrator.md` agent template under the resolved config root (`$XDG_CONFIG_HOME/opencode/agents/` when `XDG_CONFIG_HOME` is set, otherwise `~/.config/opencode/agents/`).
+The removed orchestrator WIP no longer downloads prompts, injects collaboration headers, or manages an `orchestrator.md` agent. On startup and installer runs, the plugin removes its legacy prompt-cache files and removes legacy agent files only when they contain the plugin-managed orchestrator marker; user-authored files are preserved.

docs/development/ARCHITECTURE.md

@@ -33,8 +33,10 @@ This plugin bridges OpenCode's OpenAI provider hooks to ChatGPT Codex backend en
   - request/body transform ownership split by model defaults, service-tier resolution, payload rewrites, and chat hook behavior
 - `lib/codex-native/catalog-sync.ts`
   - model-catalog bootstrap, auth selection for bootstrap, and per-auth refresh wiring
-- `lib/codex-native/collaboration.ts`
-  - plan-mode, orchestrator, and subagent collaboration instruction injection
+- `lib/codex-native/instruction-utils.ts`
+  - runtime-safe instruction merging and Codex/OpenCode tool-name adaptation
+- `lib/codex-native/ultra.ts`, `lib/codex-native/agent-execution.ts`
+  - WIP Ultra policy, catalog eligibility, and fail-closed root/child/auxiliary classification
 - `lib/codex-native/originator.ts`
   - originator header resolution (mode-aware `opencode` vs `codex_cli_rs`/`codex_exec`)
 - `lib/codex-native/browser.ts`
@@ -71,10 +73,8 @@ This plugin bridges OpenCode's OpenAI provider hooks to ChatGPT Codex backend en
   - tool handler logic for `codex-status`, `codex-switch-accounts`, `codex-toggle-account`, `codex-remove-account`
 - `lib/codex-status-tool.ts`, `lib/codex-status-storage.ts`, `lib/codex-status-ui.ts`
   - account status/usage tracking, persistence, and display formatting
-- `lib/codex-prompts-cache.ts`
-  - pinned upstream prompt fetch/sync (orchestrator + plan templates) with ETag/TTL refresh
-- `lib/orchestrator-agent.ts`
-  - managed `orchestrator.md` agent template sync and visibility gating
+- `lib/legacy-orchestrator-cleanup.ts`
+  - removal of prompt caches and plugin-managed agent files from the retired orchestrator WIP while preserving user-authored agents
 - `lib/quarantine.ts`
   - corrupted auth file detection and recovery
 - `lib/quota-threshold-alerts.ts`
@@ -107,16 +107,6 @@ This plugin bridges OpenCode's OpenAI provider hooks to ChatGPT Codex backend en
   - plugin-primary account-scoped server catalog cache
 - Existing instruction caches (for example `codex-instructions.md` + `codex-instructions-meta.json`) remain separate artifacts under the same cache root.
 
-## Cache files (pinned prompt sync)
-
-- `<config-root>/cache/codex-prompts-cache.json`
-  - pinned upstream prompt text for:
-    - Codex orchestrator agent template
-    - Codex plan-mode collaboration prompt
-- `<config-root>/cache/codex-prompts-cache-meta.json`
-  - prompt-cache metadata (`lastChecked`, URLs, ETags)
-
-Fetch behavior is best-effort and uses ETag/304 revalidation plus a TTL to limit network traffic.
 Successful live catalog fetches are source-faithful: the account-scoped `/backend-api/codex/models` payload is cached and handed to provider shaping without field-level merging against the GitHub fallback snapshot. The shared GitHub cache is used only when live catalog data is unavailable.
 
 ## Invariants

docs/development/CONFIG_FIELDS.md

@@ -30,8 +30,7 @@ Top-level:
   - Sensitive headers/tokens and selected metadata/query keys are redacted, but prompt/tool payload content may still be present.
 - `runtime.headerTransformDebug: boolean`
 - `runtime.pidOffset: boolean`
-- `runtime.collaborationProfile: boolean`
-- `runtime.orchestratorSubagents: boolean`
+- `runtime.ultra: boolean` (WIP, default `false`)
 - `global.personality: string`
 - `global.reasoningEffort: string` (optional)
   - When omitted, the selected model's live catalog `default_reasoning_level` is used, typically `"medium"`.
@@ -108,8 +107,7 @@ Default generated values:
 - `runtime.headerSnapshotBodies: false`
 - `runtime.headerTransformDebug: false`
 - `runtime.pidOffset: false`
-- `runtime.collaborationProfile`: mode-derived when unset (`true` in `codex`, `false` in `native`)
-- `runtime.orchestratorSubagents`: inherits `runtime.collaborationProfile` effective value when unset
+- `runtime.ultra: false`
 - `global.personality: "pragmatic"`
 - `global.reasoningSummary: "auto"`
 - `global.textVerbosity: "default"`
@@ -158,8 +156,7 @@ Resolved by `resolveConfig`:
 - `OPENCODE_OPENAI_MULTI_SERVICE_TIER`
 - `OPENCODE_OPENAI_MULTI_PROACTIVE_REFRESH`
 - `OPENCODE_OPENAI_MULTI_PROACTIVE_REFRESH_BUFFER_MS`
-- `OPENCODE_OPENAI_MULTI_COLLABORATION_PROFILE`
-- `OPENCODE_OPENAI_MULTI_ORCHESTRATOR_SUBAGENTS`
+- `OPENCODE_OPENAI_MULTI_ULTRA`
 
 Resolved by auth/runtime code (`lib/codex-native.ts` + helper modules under `lib/codex-native/`):