@oai404iao/pi-codex-minimal-tools 4.0.0 → 4.1.0

package/LICENSE

@@ -1,8 +1,4 @@
-Licensing for pi-codex-minimal-tools
-
-Except for the third-party materials identified in THIRD_PARTY_NOTICES.md,
-the project-authored portions of this package are available under the
-following MIT License:
+MIT License
 
 Copyright (c) 2026 oai404iao
 
@@ -23,6 +19,3 @@ AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
 LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
 OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
 SOFTWARE.
-
-Third-party materials remain under their respective terms. See
-THIRD_PARTY_NOTICES.md and the license texts under LICENSES/.

package/README.md

@@ -1,482 +1,87 @@
 # @oai404iao/pi-codex-minimal-tools
 
-Codex-specific Responses support for Pi, driven by an exact per-model JSON
-catalog instead of model-name heuristics.
-
-Peer floor: Pi 0.87.0; development target: 0.87.1.
-
-> npm identity stays `@oai404iao/pi-codex-minimal-tools`. This checkout composes
-> core, web-search and imagegen through a shared runtime. The four new packages
-> are alpha bootstrap candidates. Guarded bundle publication remains blocked
-> until those dependencies are bootstrapped and separately activated.
-> Previously published tarballs are unchanged. See
-> [package composition](../../docs/codex-packages.md) for local installation tests
-> and version/lifecycle restrictions; this is not a new npm release.
-
-> **Compatibility boundary:** Responses Lite uses an internal Codex request
-> shape. This package pins and tests a compatibility serialization for exact
-> configured model profiles, but it is not an OpenAI-supported public API
-> contract and may stop working if that internal protocol changes.
-
-The extension adds:
-
-- Responses SSE, WebSocket, cached continuation, WebSocket retry, upgrade-only
-  SSE fallback, and WebSocket prewarm.
-- Standard Responses and the internal Codex Responses Lite envelope.
-- Codex freeform `apply_patch`, including streaming preview and exact replay.
-- Hosted Responses web search or standalone Codex `web.run`.
-- Hosted Responses image generation or standalone Codex `image_gen.imagegen`.
-- Codex remote compaction v2 and legacy `/responses/compact`.
-- Per-model Fast service tiers.
-- User overrides and user-added provider/model profiles.
-
-Unknown models are not guessed. They keep Pi's native provider implementation
-and do not receive package tools.
-
-## Codex Identity
-
-This package owns Codex wire identity independently from Pi's session-file id:
-
-- a root has `SessionId === ThreadId`;
-- subagents share the root `SessionId` and receive their own `ThreadId`;
-- `prompt_cache_key` is the shared `SessionId`;
-- `x-client-request-id` is the current `ThreadId`;
-- TurnId and `x-codex-turn-state` are scoped to one logical agent turn;
-- WindowId is thread-scoped and advances after successful compaction;
-- the installation UUIDv4 is stored at
-  `<PI_CODING_AGENT_DIR>/pi-codex-minimal-tools/installation_id`.
-
-SSE, Responses Lite, WebSocket, native compaction, and standalone `web.run`
-are projections of the same request identity snapshot.
-
-The package also exports
-`@oai404iao/pi-codex-minimal-tools/subagent-inline`. Its named inline
-extension installs only identity lifecycle hooks for SDK-created child
-sessions; it does not register providers, tools, commands, or renderers.
-
-## Install
-
-After the reviewed split-package alpha and its dependencies are available on npm:
-
-```bash
-pi install npm:@oai404iao/pi-codex-minimal-tools@1.4.1-alpha.0
-```
-
-Until then, test a complete repository checkout after running the root
-`npm ci --ignore-scripts`; an isolated bundle source directory cannot supply
-its unpublished workspace dependencies:
-
-```bash
-pi install /absolute/path/to/pi-codex-minimal-tools
-```
-
-Restart Pi or run `/reload`.
-
-## Commands
-
-| Command | Action |
-| --- | --- |
-| `/codex-minimal-tools` | Show the active model profile and effective capabilities. |
-| `/codex-minimal-tools:doctor` | Show config/catalog diagnostics. |
-| `/fast [on\|off\|status]` | Toggle the active profile's Fast service tier. |
-| `/image-gen <prompt> [@reference.png]` | Run background image generation or editing. |
-
-## Two Configuration Files
-
-There are two separate model-related files:
-
-1. Pi's agent `models.json` defines providers and actual models: API type,
-   base URL, authentication, headers, modalities, context window, and cost.
-2. This extension's `models.json` defines Codex behavior for an exact
-   `provider/model` ID: Responses mode, reasoning summary, transport, tools,
-   compaction, and Fast.
-
-Typical paths:
-
-```text
-<PI_CODING_AGENT_DIR>/models.json
-<PI_CODING_AGENT_DIR>/extensions/pi-codex-minimal-tools/models.json
-<PI_CODING_AGENT_DIR>/extensions/pi-codex-minimal-tools/config.json
-```
-
-Without `PI_CODING_AGENT_DIR`, Pi normally uses `~/.config/pi/agent` or
-`~/.pi/agent`, depending on the installation.
-
-### Extension Global Config
-
-`config.json` now contains only package-wide preferences:
-
-```json
-{
-  "$schema": "https://unpkg.com/@oai404iao/pi-codex-minimal-tools@1/config.schema.json",
-  "enabled": true,
-  "glyphStyle": "unicode",
-  "autoEnable": true,
-  "webSocketEnabled": true,
-  "fastMode": false,
-  "imageGeneration": true,
-  "imageOutputDir": ".pi/openai-codex-images",
-  "imageModel": "gpt-image-2",
-  "directImageApiFallback": false,
-  "viewImageWorkspaceOnly": false,
-  "deferApplyPatchRendering": false
-}
-```
-
-| Setting | Meaning |
-| --- | --- |
-| `enabled` | Enable all package behavior. |
-| `glyphStyle` | Use `unicode` or `ascii` UI glyphs. |
-| `autoEnable` | Add supported package tools automatically. |
-| `webSocketEnabled` | Global Responses WebSocket master switch. Set `false` to force SSE and disable WebSocket prewarm for every model profile. |
-| `fastMode` | Global user toggle; only profiles with `fast` are affected. |
-| `imageGeneration` | Global master switch. Set `false` to omit image tools, `/image-gen`, presentation, hosted injection, standalone requests, and direct fallback without changing other model behavior. |
-| `imageOutputDir` | Generated-image output directory. Relative paths resolve from the workspace root. |
-| `imageModel` | Image model used by standalone/hosted image requests and direct fallback. |
-| `directImageApiFallback` | Permit the separate `OPENAI_API_KEY` Images API fallback. |
-| `viewImageWorkspaceOnly` | Restrict `view_image` to the workspace. |
-| `deferApplyPatchRendering` | Use Pi's fallback renderer instead of the streaming patch preview. |
-
-The older model-level keys remain readable for one migration version, but are
-deprecated and no longer appear in `config.schema.json`. `imageGeneration` is
-the exception: it remains the supported global master switch. See
-[Legacy migration](#legacy-migration).
-
-### Per-Model Catalog
-
-Create:
-
-```text
-<PI_CODING_AGENT_DIR>/extensions/pi-codex-minimal-tools/models.json
-```
-
-Example overriding a bundled model and adding a custom provider/model:
-
-```json
-{
-  "$schema": "https://unpkg.com/@oai404iao/pi-codex-minimal-tools@1/models.schema.json",
-  "version": 1,
-  "models": [
-    {
-      "id": "openai/gpt-5.5",
-      "responses": {
-        "reasoningSummary": "none",
-        "transport": "sse",
-        "websocketPrewarm": false
-      },
-      "tools": {
-        "webSearch": false
-      }
-    },
-    {
-      "id": "my-provider/my-codex-model",
-      "extends": "openai/gpt-5.5",
-      "responses": {
-        "endpoint": "openai",
-        "transport": "websocket-cached"
-      },
-      "tools": {
-        "imageGeneration": false
-      }
-    }
-  ]
-}
-```
-
-Resolution rules:
-
-- IDs are exact, case-insensitive `provider/model` matches.
-- A user entry with the same ID deep-overrides the bundled entry.
-- `extends` can inherit from a bundled or user profile.
-- Objects merge recursively; arrays replace the inherited array.
-- Duplicate IDs, malformed JSON, missing parents, cycles, and unknown fields
-  are reported by `/codex-minimal-tools:doctor`.
-- Each effective profile has a stable hash. WebSocket continuation, sticky
-  fallback, and native-compaction replay are isolated by that hash.
-
-Full field documentation is in
-[`reference/model-catalog.md`](https://github.com/oai404iao/omp/blob/main/pi-extensions/pi-codex-minimal-tools/reference/model-catalog.md), and editor
-validation is provided by [`models.schema.json`](models.schema.json).
-
-## Custom Provider Example
-
-The custom model must exist in Pi's agent `models.json`. For provider-shim
-features, its Pi `api` must be `openai-responses` or
-`openai-codex-responses`:
-
-```json
-{
-  "providers": {
-    "my-provider": {
-      "baseUrl": "https://api.example.com/v1",
-      "apiKey": "$MY_PROVIDER_API_KEY",
-      "api": "openai-responses",
-      "models": [
-        {
-          "id": "my-codex-model",
-          "name": "My Codex Model",
-          "reasoning": true,
-          "input": ["text", "image"],
-          "contextWindow": 272000,
-          "maxTokens": 16384,
-          "cost": {
-            "input": 0,
-            "output": 0,
-            "cacheRead": 0,
-            "cacheWrite": 0
-          }
-        }
-      ]
-    }
-  }
-}
-```
-
-The extension dynamically attaches only its stream handler to a selected
-custom provider. It does not replace the provider's URL, authentication,
-headers, or model definitions.
-
-If a profile requests the provider shim but the Pi model uses another API
-type, wire-only features are disabled:
-
-- hosted web/image tools;
-- custom/freeform `apply_patch`;
-- native Responses compaction;
-- Fast service tiers.
-
-Standalone web/image tools and function `apply_patch` can still be used when
-the profile enables them.
-
-## Model Profile Fields
-
-```json
-{
-  "id": "provider/model",
-  "extends": "provider/parent-model",
-  "enabled": true,
-  "responses": {
-    "providerShim": true,
-    "endpoint": "auto",
-    "mode": "standard",
-    "reasoningSummary": "auto",
-    "systemPromptPlacement": "instructions",
-    "transport": "auto",
-    "websocketPrewarm": true
-  },
-  "tools": {
-    "parallelCalls": true,
-    "applyPatch": "custom",
-    "webSearch": {
-      "implementation": "hosted",
-      "contentTypes": ["text", "image"]
-    },
-    "imageGeneration": "standalone",
-    "viewImage": false
-  },
-  "compaction": "responses",
-  "fast": {
-    "serviceTier": "priority",
-    "costMultiplier": 2
-  }
-}
-```
-
-Important constraints:
-
-- `responses.reasoningSummary` accepts `"auto"`, `"concise"`, `"detailed"`,
-  or `"none"`. `"none"` omits `reasoning.summary`; when absent, Standard
-  defaults to `"auto"` and Lite defaults to `"none"`.
-- `responses.mode:"lite"` always uses a developer message, namespace tools in
-  `additional_tools`, `parallel_tool_calls:false`,
-  `reasoning.context:"all_turns"`, and strips input-image `detail`.
-- Lite cannot use hosted `web_search` or hosted `image_generation`; choose
-  `standalone` instead.
-- `tools.applyPatch:"custom"` uses the Codex freeform grammar and requires the
-  provider shim. `"function"` works as a normal Pi tool.
-- `responses.endpoint:"openai"` uses API-key endpoint/auth semantics.
-  `"codex"` uses ChatGPT/Codex endpoint/auth semantics. `"auto"` infers from
-  the provider.
-- `compaction:"responses"` uses `compaction_trigger` through the selected
-  SSE/WebSocket transport. `"responses-compact"` uses the legacy unary
-  endpoint. `"pi"` keeps Pi summaries.
-
-## Built-In Profiles
-
-The bundled catalog is based on local Codex commit
-`eb9dceba1a2e658142a456c5898836774835616b` from August 12, 2026, with the
-Astra profile updated from `ddea03ad049142943bdbf13e937b1d67e8c1ba0c`.
-
-| Models | Responses | Web | Image | Patch | Compaction |
-| --- | --- | --- | --- | --- | --- |
-| `openai-codex/gpt-6-astra` | Lite, auto WS/SSE | standalone text+image | standalone | custom | responses |
-| `gpt-5.6-sol`, `gpt-5.6-terra`, `gpt-5.6-luna` | Lite, auto WS/SSE | standalone text+image | standalone | custom | responses |
-| `gpt-5.5`, `gpt-5.4` | Standard, auto WS/SSE | hosted text+image | standalone | custom | responses |
-| `gpt-5.4-mini`, `codex-auto-review` | Standard, auto WS/SSE | hosted text+image | standalone | custom | responses |
-| `gpt-5.2` | Standard, auto WS/SSE | hosted text | standalone | custom | responses |
-| legacy GPT-5/Codex entries | inherited Standard profile | profile-specific | standalone | custom | responses |
-| `gpt-4.1` | Standard SSE | off | hosted | off | Pi |
-| `o4-mini` | Standard SSE | off | off | off | Pi |
-
-The pre-Astra entries include equivalent `openai/...` and
-`openai-codex/...` IDs; the latter switch only the endpoint/auth shape to
-`codex`. Astra is bundled only for `openai-codex`, matching the analyzed
-ChatGPT subscription route. A public-API or proxy Astra deployment requires an
-explicit user profile for that endpoint.
-
-Pi supplies the Astra model descriptor. The extension composes its stream shim over that provider and does not
-replace authentication, streams, or the model catalog.
-The descriptor keeps the production default at 272,000 context tokens and
-128,000 output tokens. A backend-authorized larger context must be selected
-explicitly through Pi's provider `modelOverrides`; the extension does not infer
-that entitlement from the Astra slug.
-
-These entries are defaults, not claims that every proxy with the same model
-slug supports the protocol. Override or disable a profile for the endpoint
-actually in use.
-
-`openai-codex/gpt-6-sol` and `openai-codex/gpt-6-luna` use separately verified
-Lite profiles: developer placement, custom patch, standalone web/image,
-parallel calls disabled on the wire, and no reasoning summary. Pi 0.87.1 supplies
-their descriptors and `off -> none` mapping; no `ultra`, larger context,
-pricing multiplier or public-API cache TTL is inferred. Fast mode is disabled.
-See `provenance/openai-codex-40eac3ce-sol-luna.json`. Source/fixture verification
-does not establish that a particular account can access these endpoints.
-
-## Web Search
-
-`tools.webSearch` supports two implementations:
-
-- `hosted`: rewrites the Pi placeholder to Responses
-  `{"type":"web_search"}` and preserves progress, sources, results, and
-  citation replay. Standard Responses only.
-- `standalone`: exposes `web.run` as a client-executed namespace tool and calls
-  the provider's `alpha/search` endpoint. It supports search/image queries,
-  open/click/find, PDF screenshots, finance, weather, sports, and time.
-
-Standalone search sends a bounded recent visible conversation tail, resolved
-Pi authentication, direct-caller/live-web settings, the active turn metadata,
-and the model's 10,000-token truncation budget. Search rows stay compact by
-default and show deduplicated source hosts; expand the tool row to inspect the
-raw result text.
-
-## Image Generation
-
-Set global `config.json.imageGeneration` to `false` to disable the entire
-generation capability. Per-model `tools.imageGeneration:false` disables only
-that profile. Neither setting disables image input or historical image replay.
-
-`tools.imageGeneration` supports:
-
-- `hosted`: Responses `image_generation`.
-- `standalone`: `image_gen.imagegen`, backed by the provider's
-  `images/generations` and `images/edits` endpoints.
-
-Standalone input follows current Codex:
-
-- required `prompt`;
-- up to five `referenced_image_paths`; or
-- `num_last_images_to_include` from 1 through 5.
-
-Generated PNGs are saved under `imageOutputDir`, mirrored to `latest.png`, and
-returned to the model before the saved-path text. `/image-gen` selects any
-loaded image-capable model with an enabled catalog profile.
-
-## WebSocket And Compaction
-
-Global `config.json.webSocketEnabled:false` forces `sse` and disables
-WebSocket prewarm without changing the selected per-model profile. When it is
-`true` (the default), each profile's `responses.transport` and
-`responses.websocketPrewarm` values apply.
-
-`transport` values:
-
-- `sse`: HTTP streaming only.
-- `websocket`: reusable WebSocket; exact logical prefixes opportunistically use
-  `previous_response_id` input deltas.
-- `websocket-cached`: compatibility alias with the same safe continuation
-  behavior.
-- `auto`: retry transient WebSocket failures, but use sticky per-session SSE
-  fallback only when the WebSocket upgrade is rejected with HTTP 426. Model,
-  request, output-limit, context-limit, and post-upgrade connection errors do
-  not change transports.
-
-Prewarm is scheduled once per session startup (including resume/fork), in the
-background without delaying session initialization. If the first WebSocket
-request starts while prewarm is pending, it consumes that one startup handle
-before sending. Prewarm sends a best-effort `response.create` with
-`generate:false` containing only the startup system/tool/request
-envelope—never conversation history or the pending user message. The first real
-request appends its full conversation input, and later turns continue from real
-response IDs. A failed or timed-out prewarm is not retried on every user
-message. Continuation reuse always requires the new request to extend the
-previous logical request exactly.
-
-Native compaction stores opaque encrypted state in the Pi session. Version 4
-keeps no preceding messages in model context; the raw session/UI history remains.
-It preserves other context handlers' changes and lets Pi restore current
-system/tool state. Replay requires the same provider, model, API and effective
-profile hash. If these change, restore the original configuration or navigate
-before compaction: an opaque checkpoint's placeholder is not a usable text summary.
-
-Legacy checkpoints still replay unchanged canonical contexts. If an earlier
-context handler transforms a legacy checkpoint, continuation stops rather than
-guessing its boundary or restoring filtered content. Run `/compact` on the
-original model to migrate it. Failed recompression preserves the old checkpoint.
-Treat sessions containing native compaction as sensitive data; older extension
-versions cannot replay v4 checkpoints.
-
-## Apply Patch
-
-When `apply_patch` activates, the extension temporarily hides Pi's `edit` and
-`write` tools and restores their prior positions after switching away.
-
-The executor supports Codex `@@ class/function` contexts, ordered update
-chunks, `*** Move to:`, `*** End of File`, fuzzy matching, CRLF preservation,
-and atomic verification before writes. See
-[`reference/apply-patch-behavior.md`](https://github.com/oai404iao/omp/blob/main/pi-extensions/pi-codex-minimal-tools/reference/apply-patch-behavior.md).
-
-## Legacy Migration
-
-These `config.json` keys are deprecated but still projected into an in-memory
-model override for compatibility:
-
-| Old key | New model-profile field |
-| --- | --- |
-| `nativeProviderTools` | `responses.providerShim` plus hosted/standalone tool selection |
-| `openaiTransport` | `responses.transport` |
-| `openaiWebSocketPrewarm` | `responses.websocketPrewarm` |
-| `compactionMode` | `compaction` |
-| `requestProfile.responsesMode` | `responses.mode` |
-| `requestProfile.reasoningSummary` | `responses.reasoningSummary` |
-| `requestProfile.systemPromptPlacement` | `responses.systemPromptPlacement` |
-| `requestProfile.patchTransport` | `tools.applyPatch` |
-| `apiKeyMode` | `responses.endpoint` |
-| `webSearchEnabled` | `tools.webSearch` |
-| `viewImage` | `tools.viewImage` |
-| `applyPatchEnabled` | `tools.applyPatch` |
-| `additionalModelIds` | one exact user catalog entry per model |
-
-Move these values to extension `models.json`; legacy projection is intended
-only as a transition path. `imageGeneration` is no longer a legacy projection:
-it is the supported global gate, while per-model selection stays in
-`tools.imageGeneration`.
-
-## Protocol Reference
-
-[`reference/`](https://github.com/oai404iao/omp/tree/main/pi-extensions/pi-codex-minimal-tools/reference) documents the Codex snapshot, Responses
-Standard/Lite envelopes, namespace tools, WebSocket continuation, custom-tool
-streaming/replay, standalone web/image endpoints, compaction, and apply-patch
-protocol.
-
-## License and publication status
-
-Project-authored portions are MIT-licensed, copyright 2026 oai404iao.
-Third-party material retains its own terms; see
-[LICENSE](LICENSE) and [THIRD_PARTY_NOTICES.md](THIRD_PARTY_NOTICES.md).
-
-The Codex namespace-tool source attribution is recorded in
-`provenance/openai-codex-eb9dceba-reserved-tools.json`. The existing bundle has
-configured trusted publishing, but the split-package alpha still requires
-dependency-first bootstrap and activation of the four new packages before
-guarded publication can proceed. Responses Lite remains an internal Codex compatibility path,
-not an OpenAI-supported public API contract.
+Composition-only entry for Pi's Codex extensions. It loads:
+
+- [`pi-codex-core`](../pi-codex-core/README.md): Responses transport, compaction,
+  `apply_patch`, `view_image` and diagnostics.
+- [`pi-codex-web-search`](../pi-codex-web-search/README.md): hosted/standalone search.
+- [`pi-codex-imagegen`](../pi-codex-imagegen/README.md): image generation and display.
+
+All three use [`pi-codex-runtime`](../pi-codex-runtime/README.md). Compatible
+duplicate installations are deduplicated by the shared broker.
+Peer floor and development target: Pi 0.99.1.
+
+Use Pi's `openai` provider for API keys or ChatGPT OAuth (`/login openai`).
+`openai-codex` remains deprecated compatibility only. No credentials, model
+selection or session history are migrated automatically. New requests follow
+Pi's API/base URL rather than plugin `apiKeyMode` or `responses.endpoint`
+overrides outside the legacy provider.
+
+`openai/gpt-6.1-sol` defaults to the same profile as `openai/gpt-5.6-sol`:
+Lite, auto WS/SSE, prewarm, standalone web/image, custom patch, native compaction
+and Fast availability (priority, 2x cost multiplier). `view_image` is off.
+These defaults do not establish real-account endpoint access or Fast pricing.
+Exact profiles for `openai/gpt-6-astra`, `gpt-6-sol` and
+`gpt-6-luna` retain Standard SSE with local custom patch and image viewing.
+Remote search/image generation, native compaction, prewarm and Fast default
+off. Enable only the implementations you intend to use in the model profile,
+and declare endpoint support separately; allowlists do not enable tools.
+
+Native codemode may call active `apply_patch` and standalone web/image tools.
+Hosted web/image tools and `view_image` remain model-only. The bundle does not
+activate codemode or load MCP servers.
+Script results are structured: patch `{ summary, files }`, search
+`{ output, results }`, and image generation `{ path, latestPath?, image }`.
+With native codemode/tool-search active, core skips speculative prewarm and
+uses Pi text compaction only without an existing opaque native checkpoint;
+otherwise it preserves the checkpoint and refuses recompaction.
+
+The package name, default extension factory and
+`@oai404iao/pi-codex-minimal-tools/subagent-inline` remain supported.
+The latter re-exports runtime's identity-only SDK extension.
+
+## Configuration and installation
+
+Existing configuration stays at
+`<agentDir>/extensions/pi-codex-minimal-tools/{config,models}.json`.
+The file locations are unchanged. Move nonlegacy endpoint overrides to Pi's
+own `models.json`; `directImageApiFallback` now uses the selected provider's
+Pi credentials and endpoint, not a separate `OPENAI_API_KEY` account.
+Schemas belong only to runtime:
+
+- [config.schema.json](https://unpkg.com/@oai404iao/pi-codex-runtime/config.schema.json)
+- [models.schema.json](https://unpkg.com/@oai404iao/pi-codex-runtime/models.schema.json)
+
+Global `codexRequestExtensions` defaults to `true`. Set it to `false` to stop
+plugin-generated Codex wire metadata and enhancements while keeping Standard
+tools and hosted result handling; Lite requests and incompatible opaque
+checkpoint transitions are blocked rather than silently downgraded.
+
+`endpoint_config` accepts exact `{ provider, baseUrl, webSearch?, imageGeneration?,
+compaction? }` entries. Tool lists contain `"hosted"`/`"standalone"`; compaction
+lists contain `"responses"`/`"responses-compact"`. Omitted lists inherit the
+model profile, `[]` denies the capability, and lists never select a fallback.
+Matching uses Pi's authenticated endpoint. Explicit unsupported responses warn
+and disable that mode only for the current session/endpoint, without changing
+credentials or writing config. See the configuration guide for an example and
+error-classification limits; `/reload` clears transient rejections.
+
+Old version-pinned `pi-codex-minimal-tools@1` schema URLs still refer to their
+unchanged published artifacts. This checkout no longer duplicates schemas or
+private `src/*` forwards in the bundle.
+
+See the [configuration guide](https://github.com/oai404iao/omp/blob/main/pi-extensions/pi-codex-runtime/reference/configuration.md)
+and [installation/composition guide](https://github.com/oai404iao/omp/blob/main/docs/codex-packages.md).
+For local testing, run root `npm ci --ignore-scripts`, then
+`pi install /absolute/path/to/omp/pi-extensions/pi-codex-minimal-tools`.
+These source changes do not themselves publish a new npm release.
+
+Tests live with their implementation owners; cross-package tests are in root
+`tests/codex/`. Protocol references and provenance are indexed in runtime's
+[source map](https://github.com/oai404iao/omp/blob/main/pi-extensions/pi-codex-runtime/reference/source-map.md).
+Responses Lite is an internal compatibility path, not a supported public API.
+
+## License
+
+The composition entry is [MIT](LICENSE). Dependencies retain their own licenses
+and source notices; see [THIRD_PARTY_NOTICES.md](THIRD_PARTY_NOTICES.md).