@gajae-code/coding-agent 0.13.2 → 0.13.3

package/src/internal-urls/docs-index.generated.ts

@@ -1,7 +1,7 @@
 // Auto-generated by scripts/generate-docs-index.ts - DO NOT EDIT
 Reflect.set(globalThis, Symbol.for("gjc.docs-index.generated.loaded"), true);
 
-export const EMBEDDED_DOC_FILENAMES: readonly string[] = ["ERRATA-GPT5-HARMONY.md","REBRANDING_PLAN_260525.md","acp-local-development.md","adr-inline-selection-gate.md","adr-overlay-component-seam.md","adr-sessions-dashboard.md","ai-schema-normalize.md","alibaba-token-plan-pro-profile-benchmark.md","analyze-me-with-gjc.md","aside-integration.md","auth-broker-gateway.md","bash-tool-runtime.md","blob-artifact-architecture.md","bot-integration.md","brand-assets.md","clipboard-transport.md","codebase-overview.md","codegraph-custom-tool.md","compaction.md","composer-codex-parity.md","computer-use/README.md","cursor-composer-profile-tiers.md","custom-providers-and-multi-account.md","discord-onboarding.md","environment-variables.md","external-control-readiness.md","extragoal-skill-template.md","fs-scan-cache-architecture.md","geobench.md","git-daemon.md","gjc-dogfood-skill-template.md","gjc-plugins.md","gjc-session-clawhip-routing.md","gpt-5.6-codex-preset-benchmark.md","grok-build-provider-design.md","handoff-generation-pipeline.md","hermes-mcp-bridge.md","hotspot-map-successor.md","install.md","keybindings.md","lsp-config.md","macos-option-key.md","memory.md","models.md","multi-vendor-profiles.md","native-ffi-optimization-policy.md","natives-addon-loader-runtime.md","natives-architecture.md","natives-binding-contract.md","natives-build-release-debugging.md","natives-media-system-utils.md","natives-rust-task-cancellation.md","natives-shell-pty-process.md","natives-text-search-pipeline.md","non-compaction-retry-policy.md","notebook-tool-runtime.md","onboarding-packet.md","onboarding-receipt.md","ooo-bridge-extension-contract.md","perf-profiling-corpus.md","porting-from-pi-mono.md","porting-to-natives.md","prompt-architect-reports/README.md","prompt-architect-reports/recovered-context/0-ToolPrompts.recovered.md","prompt-architect-reports/recovered-context/1-SystemPrompts.recovered.md","prompt-architect-reports/recovered-context/3-SkillMiscPrompts.recovered.md","prompt-architect-reports/recovery-summary.md","prompt-architect-reports/system-prompts.raw.md","prompt-architect-reports/tool-prompts.raw.md","provider-streaming-internals.md","python-repl.md","render-mermaid.md","research-plan-ledger.md","resolve-tool-runtime.md","rulebook-matching-pipeline.md","sdk-app-guide.md","sdk-embedding.md","sdk-rpc-parity-audit.md","sdk-session-cli.md","sdk.md","secrets.md","session-operations-export-share-fork-resume.md","session-switching-and-recent-listing.md","session-tree-plan.md","session.md","slack-onboarding.md","speech-to-text.md","standalone-mcp.md","streamdeck-integration-guide-with-cmux.md","telegram-onboarding.md","theme.md","tools/ask.md","tools/ast-edit.md","tools/ast-grep.md","tools/bash.md","tools/bisect.md","tools/browser.md","tools/calc.md","tools/checkpoint.md","tools/computer.md","tools/cron.md","tools/debug.md","tools/edit.md","tools/eval.md","tools/find.md","tools/github.md","tools/irc.md","tools/job.md","tools/lsp.md","tools/monitor.md","tools/read.md","tools/recipe.md","tools/render_mermaid.md","tools/resolve.md","tools/rewind.md","tools/search.md","tools/search_tool_bm25.md","tools/ssh.md","tools/task.md","tools/todo_write.md","tools/web_search.md","tools/write.md","tree.md","ttsr-injection-lifecycle.md","tui-runtime-internals.md","ui-design-visual-qa.md"];
+export const EMBEDDED_DOC_FILENAMES: readonly string[] = ["ERRATA-GPT5-HARMONY.md","REBRANDING_PLAN_260525.md","acp-local-development.md","adr-inline-selection-gate.md","adr-overlay-component-seam.md","adr-sessions-dashboard.md","ai-schema-normalize.md","alibaba-token-plan-pro-profile-benchmark.md","analyze-me-with-gjc.md","aside-integration.md","auth-broker-gateway.md","bash-tool-runtime.md","blob-artifact-architecture.md","bot-integration.md","brand-assets.md","clipboard-transport.md","codebase-overview.md","codegraph-custom-tool.md","compaction.md","composer-codex-parity.md","computer-use/README.md","cursor-composer-profile-tiers.md","custom-providers-and-multi-account.md","discord-onboarding.md","environment-variables.md","external-control-readiness.md","extragoal-skill-template.md","fs-scan-cache-architecture.md","geobench.md","git-daemon.md","gjc-dogfood-skill-template.md","gjc-plugins.md","gjc-session-clawhip-routing.md","gpt-5.6-codex-preset-benchmark.md","grok-build-provider-design.md","handoff-generation-pipeline.md","hermes-mcp-bridge.md","hotspot-map-successor.md","install.md","keybindings.md","lsp-config.md","macos-option-key.md","memory.md","models.md","multi-vendor-profiles.md","native-ffi-optimization-policy.md","natives-addon-loader-runtime.md","natives-architecture.md","natives-binding-contract.md","natives-build-release-debugging.md","natives-media-system-utils.md","natives-rust-task-cancellation.md","natives-shell-pty-process.md","natives-text-search-pipeline.md","non-compaction-retry-policy.md","notebook-tool-runtime.md","onboarding-packet.md","onboarding-receipt.md","ooo-bridge-extension-contract.md","perf-profiling-corpus.md","porting-from-pi-mono.md","porting-to-natives.md","prompt-architect-reports/README.md","prompt-architect-reports/recovered-context/0-ToolPrompts.recovered.md","prompt-architect-reports/recovered-context/1-SystemPrompts.recovered.md","prompt-architect-reports/recovered-context/3-SkillMiscPrompts.recovered.md","prompt-architect-reports/recovery-summary.md","prompt-architect-reports/system-prompts.raw.md","prompt-architect-reports/tool-prompts.raw.md","provider-streaming-internals.md","python-repl.md","release-0.13.3-integration.md","render-mermaid.md","research-plan-ledger.md","resolve-tool-runtime.md","rulebook-matching-pipeline.md","sdk-app-guide.md","sdk-embedding.md","sdk-rpc-parity-audit.md","sdk-session-cli.md","sdk.md","secrets.md","session-operations-export-share-fork-resume.md","session-switching-and-recent-listing.md","session-tree-plan.md","session.md","slack-onboarding.md","speech-to-text.md","standalone-mcp.md","streamdeck-integration-guide-with-cmux.md","telegram-onboarding.md","theme.md","tools/ask.md","tools/ast-edit.md","tools/ast-grep.md","tools/bash.md","tools/bisect.md","tools/browser.md","tools/calc.md","tools/checkpoint.md","tools/computer.md","tools/cron.md","tools/debug.md","tools/edit.md","tools/eval.md","tools/find.md","tools/github.md","tools/irc.md","tools/job.md","tools/lsp.md","tools/monitor.md","tools/read.md","tools/recipe.md","tools/render_mermaid.md","tools/resolve.md","tools/rewind.md","tools/search.md","tools/search_tool_bm25.md","tools/ssh.md","tools/task.md","tools/todo_write.md","tools/web_search.md","tools/write.md","tree.md","ttsr-injection-lifecycle.md","tui-runtime-internals.md","ui-design-visual-qa.md"];
 
 export const EMBEDDED_DOCS: Readonly<Record<string, string>> = {
 	"ERRATA-GPT5-HARMONY.md": "# ERRATA — GPT-5 Harmony-Header Leakage\n\n## 1. The problem\n\nOpenAI frames tool calls in the Harmony chat protocol:\n\n```\n<|start|>assistant<|channel|>commentary to=functions.<NAME><|message|>{ARGS}<|call|>\n```\n\n`<|channel|>commentary to=functions.NAME` is the **routing header** —\ncontrol tokens consumed by the runtime to dispatch the call. These\ntokens never appear as content under normal operation; the runtime\nstrips them.\n\nThe defect: gpt-5 models occasionally emit, **as ordinary content\ninside `{ARGS}`**, the **plain-text shadow** of these routing tokens —\nthe same characters without the `<|…|>` brackets — and continue\nproducing more pseudo-routing structure (channel name, body marker,\nmultilingual spam, fake tool-result framing). The contamination lives\ninside the visible tool argument and is dispatched to the tool as if it\nwere intended content.\n\n**Critical detail.** The actual `<|start|>` / `<|channel|>` /\n`<|message|>` / `<|call|>` special tokens almost never appear in tool\nargs. What leaks is the bracket-less spelling — `analysis to=functions.X\ncode …` — because OpenAI applies a logit mask suppressing the\ncontrol-token IDs inside the args region. The mass that would have gone\nto those special tokens redistributes onto the un-bracketed plain-text\nrepresentation the model also learned. This makes the leak structurally\ninvisible to the routing parser and lands it in the tool input verbatim.\n\nManifestation in tool args (real corpus example):\n\n```\n~      add_function(iso, ctx, ns, \"installSystemChangeObserver\",\n        os_install_system_change_observer);】【\"】【analysis to=functions.edit\n        code above เงินไทยฟรีuser to=functions.edit code …\n```\n\nThe leading code is real and intended. Everything after the first\nnon-Latin token through the next clean structural boundary is corruption.\n\n---\n\n## 2. Observed statistics & failure modes\n\nSource: `~/.gjc/stats.db` (`ss_tool_calls`, `ss_assistant_msgs`), through\n2026-05-10. 1.05M tool calls scanned.\n\n### 2.1 Rate\n\n| Model            | Leaks in tool args | Calls   | per million |\n|------------------|-------------------:|--------:|------------:|\n| gpt-5.4          | 37                 | 226,957 | 163         |\n| gpt-5.3-openai-code    | 17                 | 112,243 | 151         |\n| gpt-5.5          | 2                  |  80,750 |  25         |\n| gpt-5.2-openai-code    | 0                  |    —    |   —         |\n\nPlus 15 hits in assistant visible text / thinking blobs.\n\n### 2.2 Tool distribution\n\n| Tool                | Hits |\n|---------------------|-----:|\n| `edit`              | 38   |\n| `eval`              | 11   |\n| `report_tool_issue` | 3    |\n| `grep`/`read`/`search`/`yield` | 1 each |\n\nConcentrated in tools with free-form (non-JSON-schema) argument formats.\n\n### 2.3 Leak shape (deterministic)\n\n```\nLEAK         ::= JUNK_PREFIX MARKER CHANNEL_BODY (LEAK)?\nMARKER       ::= \"to=functions.\" TOOL_NAME\nCHANNEL_BODY ::= \" code \" (SPAM | reasoning_prose | fake_tool_output)*\nJUNK_PREFIX  ::= (GLITCH_TOKEN | CHANNEL_WORD | NON_LATIN_RUN | \"}\" | \"】【\")+\n```\n\n**Cascading is common.** Of 96 marker occurrences across 71 contaminated\nrecords, 39 contain ≥2 markers and 7 contain ≥3 — the model emits\nmultiple fake `to=functions.X code …` blocks back-to-back, often with\nfake `code_output\\nCell N:\\n…` framing between them. Once the\nplain-text scaffolding is in the residual stream, the prefix now *looks\nlike* a fresh tool envelope start, so the macro prior over continuations\nkeeps voting for more scaffolding. Self-amplifying.\n\n### 2.4 Glitch tokens\n\nSingle-token identifiers in `o200k_base` whose embeddings appear to be\nnear-init from underrepresentation in post-training. ASCII residue\nimmediately before the marker in the natural corpus:\n\n| Surface string    | Single-token | Token ID | Hits in corpus |\n|-------------------|:-:|---------:|---:|\n| `Japgolly`        | ✅ | 199,745 | 1 |\n| `Jsii`            | ✅ | 114,318 | (subtoken of `Jsii_commentary`) |\n| `Jsii_commentary` | — (3 toks) | — | 2 |\n| `changedFiles`    | — (2 toks) | — | 8 |\n| `RTLU`            | — (2 toks) | — | 3 |\n\n`Japgolly` is in the last 0.13% of the vocabulary — the same family of\nGitHub-corpus residue that produced `SolidGoldMagikarp` in the 2023\nGPT-2 vocabulary (Rumbelow & Watkins). `SolidGoldMagikarp` itself\ntokenizes to 5 tokens in `o200k_base` — that specific token was retired,\nbut the class wasn't.\n\nFor the multi-token entries, the corpus-level signature is the surface\nstring; the underlying glitch trigger is a sub-token (e.g. `Jsii` inside\n`Jsii_commentary`). The detector list (`G` signal) keys on the surface\nstrings.\n\nStable across unrelated sessions. Treated as a high-precision detector\nsignal.\n\n### 2.5 Channel-word leakage\n\n`analysis` (5), `assistant` (5), `commentary` (3), `user` (1) appear\ndirectly preceding `to=`. Always bare words; never `<|channel|>analysis`\nor any other bracketed form. Consistent with §1 — the brackets are\nmasked, the words are not.\n\n### 2.6 Non-Latin spam residue\n\n96 marker hits, by script: CJK 40, Cyrillic 12, Telugu/Kannada/Malayalam\n18, Thai 8, Georgian 7, Armenian 7, Arabic 1. Recurring fragments are\nChinese gambling SEO (`大发时时彩`, `天天中彩票`), Georgian/Abkhaz junk,\nand Thai casino spam — well-known low-quality crawl residue.\n\nThis is the same script distribution observed in the controlled\nreproduction (§7.3), independent of the prompt's natural language.\n\n### 2.7 Failure-mode breakdown for the `edit` tool\n\nThe `edit` tool exists in two variants in the corpus:\n\n| Variant                  | Calls | Recovery |\n|--------------------------|------:|----------|\n| Patch-DSL (`§PATH`/anchor/`«»≔` ops) | 27 | **Recoverable** by op-truncation (§3.3) |\n| JSON-schema (`{path,edits:[…]}`)      | 11 | **Not recoverable** — contamination is escaped *inside* JSON strings, parser accepts it cleanly, content would be written verbatim into source files |\n\nFor Patch-DSL leaks specifically:\n\n- 20/27 cases: contamination on the last input line; nothing follows.\n- 7/27 cases: contamination mid-input; what follows is one of: a\n  duplicate replay of an earlier file/anchor, intended content for a\n  *different* tool call (the model started its next call inline), or\n  pure hallucination. Post-contamination content is never trustworthy.\n\n### 2.8 Mechanism (confirmed)\n\n**Prior collapse from null-embedding glitch tokens, into a\ncontrol-token-masked basin whose mass redistributes onto the\nplain-text shadow of the Harmony protocol.**\n\nStep by step:\n\n1. The model is mid-`{ARGS}` of a Harmony tool call. The runtime applies\n   a logit mask suppressing structural control tokens (`<|channel|>`,\n   `<|message|>`, `<|call|>`, `<|start|>`, `<|end|>`) inside the args\n   region. Without this mask, normal generation would constantly\n   hallucinate envelope-closes; with it, those token IDs have logit\n   `-∞` in args.\n2. A glitch token `g` is sampled. By construction `g` was in the BPE\n   merge corpus but barely in LM/RL training, so its **input embedding\n   `e_g` ≈ near-init noise of small norm**.\n3. At position t+1, the residual update `h_{t+1} ≈ LN(h_t + e_g + Attn +\n   MLP)` is dominated by the prefix-derived terms; the just-emitted-token\n   signal is effectively absent. Generation diversity normally comes\n   from `e_x` steering the residual into different sub-regions —\n   stripped here.\n4. The next-token distribution therefore collapses onto the **conditional\n   prior over continuations of the prefix, with local conditioning\n   removed**. In a tool-calling rollout context, that prior is sharply\n   peaked on Harmony scaffolding (control tokens + routing tokens) —\n   that's what RL trained.\n5. The mask zeros the control-token IDs. Mass redistributes onto the\n   **next-best continuation**: the un-bracketed surface-form spelling of\n   the same protocol (`analysis`, `commentary`, ` to=functions.X`,\n   ` code `). This spelling is unmasked because those characters are\n   ordinary tokens.\n6. Once a few tokens of plain-text scaffolding land in the residual\n   stream, the prefix now resembles a fresh envelope start. The macro\n   prior keeps voting for more scaffolding. Cascading (§2.3) follows.\n7. Multilingual spam after the marker is the same prior-collapse\n   continuation, drawn from the training neighborhood of the glitch\n   token (often ESL/auto-generated multilingual web junk — exactly the\n   crawl residue in §2.6).\n\n**Two corollaries the corpus data demanded but only the experiment\nexplained:**\n\n- **The brackets never appear** (§1, §2.5). The mask is what makes the\n  leak land in plain text instead of as a real envelope-close.\n- **Counterintuitive grammar dependency** (§7.4). The leak is *worse* in\n  formats closest to OpenAI's training distribution. Off-distribution\n  custom grammars dampen the macro-prior basin; the official\n  `*** Begin Patch` format is the strongest collapse target.\n\nThe 2023 SolidGoldMagikarp paper documented mechanism (1)+(2)+(4). The\nnew piece is (5): when constrained decoding masks the natural collapse\ntarget, the mass laundered through the un-masked plain-text shadow\nbecomes a structurally-invisible exfiltration channel.",
@@ -47,7 +47,7 @@ export const EMBEDDED_DOCS: Readonly<Record<string, string>> = {
 	"lsp-config.md": "# LSP configuration in GJC\n\nThis guide explains how to configure language servers for the GJC coding agent.\n\nSource of truth in code:\n\n- Server config type: `packages/coding-agent/src/lsp/types.ts` (`ServerConfig`)\n- Config loader: `packages/coding-agent/src/lsp/config.ts`\n- Built-in server definitions: `packages/coding-agent/src/lsp/defaults.json`\n\n## Auto-detection\n\nWhen no LSP config file is present, GJC auto-detects servers by intersecting two conditions:\n\n1. The project directory contains at least one of the server's `rootMarkers`.\n2. The server binary is a trusted external executable. Project-local binaries, including paths reached through symlinks, are rejected.\n\nNo configuration is required for common setups. The built-in server list covers most popular languages; see [`defaults.json`](../packages/coding-agent/src/lsp/defaults.json) for the full set.\n\n## Config file locations\n\nGJC merges LSP config from multiple files, lowest to highest priority:\n\n| Priority | Location |\n|----------|----------|\n| 5 (lowest) | `~/lsp.json`, `~/.lsp.json`, `~/lsp.yaml`, `~/.lsp.yaml` |\n| 4 | Preloaded trusted external plugin LSP config outside the project (internal loader support; no current CLI/startup producer) |\n| 3 | `~/.gjc/agent/lsp.json`, `~/.gjc/agent/lsp.yaml`, `~/.gemini/lsp.*` |\n| 2 | `<project>/.gjc/lsp.json`, `<project>/.gjc/lsp.yaml`, `<project>/.gemini/lsp.*` |\n| 1 (highest) | `<project>/lsp.json`, `<project>/.lsp.json`, `<project>/lsp.yaml` |\n\nEach location accepts both `.json` and `.yaml` / `.yml` variants, as well as hidden-file versions (`.lsp.json`, `.lsp.yaml`). Configuration is merged in order, but project-controlled files can only control declarative server matching, activation, and capabilities. They cannot define or override a server's `command`, `args`, executable, client factory, `initOptions` / `initializationOptions`, or `settings`; opaque options that can instruct a trusted server belong to trusted user configuration.\n\nThe recommended trusted user configuration is `~/.gjc/agent/lsp.json` (or YAML equivalent). Legacy user-wide `~/.gemini/lsp.*` and home-root `~/lsp.*` / `~/.lsp.*` files are also outside the project and may define launch settings and opaque server options, including custom servers. Project files may refine declarative matching and activation fields of built-in or user-defined servers.\n\n**Recommended locations:**\n\n- Trusted user launch settings, `initOptions`, and `settings` → `~/.gjc/agent/lsp.json`\n- Project-specific matching and activation → `<project>/.gjc/lsp.json`\n\n> **Note:** The presence of any LSP config file disables auto-detection. When at least one file is found, GJC skips the binary-scan phase and loads matching, available, non-disabled servers using trusted launch definitions.\n\n## File shape\n\nBoth JSON and YAML are accepted. The top-level object can use either a `servers` wrapper key or a flat map directly:\n\n```json\n{\n  \"servers\": {\n    \"server-name\": { ... }\n  },\n  \"idleTimeoutMs\": 300000\n}\n```\n\nor (flat, without the `servers` wrapper):\n\n```json\n{\n  \"server-name\": { ... },\n  \"idleTimeoutMs\": 300000\n}\n```\n\nTop-level keys:\n\n- `servers` — map of server name to `ServerConfig` (optional wrapper; flat form is equivalent)\n- `idleTimeoutMs` — shut down idle language servers after this many milliseconds; disabled by default\n\n## ServerConfig fields\n\n| Field | Type | Required | Description |\n|-------|------|----------|-------------|\n| `command` | `string` | trusted user config only | Server executable name or absolute path; project configuration cannot set or override it |\n| `args` | `string[]` | no | Launch arguments; trusted user config only |\n| `fileTypes` | `string[]` | yes | File extensions this server handles, e.g. `[\".ts\", \".tsx\"]` |\n| `rootMarkers` | `string[]` | yes | Files/dirs that indicate a project root; glob patterns (e.g. `*.cabal`) are supported |\n| `initOptions` | `object` | trusted user config only | Sent as `initializationOptions` during LSP handshake |\n| `settings` | `object` | trusted user config only | Workspace settings pushed via `workspace/didChangeConfiguration` |\n| `disabled` | `boolean` | no | Set to `true` to disable this server entirely |\n| `warmupTimeoutMs` | `number` | no | Startup timeout in ms for this server (overrides the global default) |\n| `isLinter` | `boolean` | no | Mark server as linter/formatter only; excluded from type-intelligence operations (hover, go-to-definition, etc.) |\n| `capabilities` | `object` | no | Opt-in server-specific features; see [Capabilities](#capabilities) |\n\n`resolvedCommand` is populated automatically at runtime — do not set it manually.\n\n### Capabilities\n\nThe `capabilities` object enables optional server-specific features that GJC supports on a per-server basis:\n\n```json\n{\n  \"capabilities\": {\n    \"flycheck\": true,\n    \"ssr\": true,\n    \"expandMacro\": true,\n    \"runnables\": true,\n    \"relatedTests\": true\n  }\n}\n```\n\nAll fields are boolean and optional. They are currently used by `rust-analyzer`.\n\n## Common recipes\n\n### Override a built-in server's settings from trusted user configuration\n\nOpaque server settings may contain process-affecting instructions, so place these partial overrides in trusted user configuration such as `~/.gjc/agent/lsp.json`:\n\n```json\n{\n  \"servers\": {\n    \"typescript-language-server\": {\n      \"settings\": {\n        \"typescript\": {\n          \"preferences\": {\n            \"quoteStyle\": \"single\"\n          }\n        }\n      }\n    }\n  }\n}\n```\n\n```yaml\nservers:\n  gopls:\n    settings:\n      gopls:\n        gofumpt: false\n        staticcheck: false\n```\n\n### Disable a built-in server\n\n```json\n{\n  \"servers\": {\n    \"eslint\": {\n      \"disabled\": true\n    }\n  }\n}\n```\n\n### Register a custom server\n\nRegister custom servers in the canonical trusted user configuration, `~/.gjc/agent/lsp.json`. New servers require `command`, `fileTypes`, and `rootMarkers`; `args` is optional. Project configuration cannot register a launch definition or override a server's command, arguments, executable, or client factory.\n\n```json\n{\n  \"servers\": {\n    \"my-lsp\": {\n      \"command\": \"my-lsp-server\",\n      \"args\": [\"--stdio\"],\n      \"fileTypes\": [\".xyz\"],\n      \"rootMarkers\": [\".xyz-project\", \".git\"]\n    }\n  }\n}\n```\n\n### Set a global idle timeout\n\nShut down language servers that have been inactive for more than five minutes:\n\n```json\n{\n  \"idleTimeoutMs\": 300000\n}\n```\n\n### Disable a server for one project, keep it globally\n\nPlace the override in `<project>/.gjc/lsp.json`:\n\n```json\n{\n  \"servers\": {\n    \"pylsp\": {\n      \"disabled\": true\n    }\n  }\n}\n```\n\nThe user-level config in `~/.gjc/agent/lsp.json` is unaffected; pylsp is only suppressed in this project.\n\nWhen multiple built-in primary servers support the same file, a default server can list lower-precedence servers in `supersedes`. For example, `csharp-ls` supersedes `omnisharp` only when both C# servers are installed and detected; if `csharp-ls` is unavailable, `omnisharp` remains the fallback.\n\n## lspmux\n\n`GJC_DISABLE_LSPMUX=1` is the canonical opt-out. `PI_DISABLE_LSPMUX=1` is a supported compatibility alias. A truthy value for either variable disables lspmux probing and wrapping.\n\n## Built-in server list\n\nThe following servers ship in `defaults.json` and are eligible for auto-detection:\n\n| Server key | Language(s) | Binary |\n|---|---|---|\n| `rust-analyzer` | Rust | `rust-analyzer` |\n| `clangd` | C, C++, ObjC | `clangd` |\n| `zls` | Zig | `zls` |\n| `gopls` | Go | `gopls` |\n| `typescript-language-server` | TypeScript, JavaScript | `typescript-language-server` |\n| `denols` | TypeScript, JavaScript (Deno) | `deno` |\n| `biome` | TS/JS/JSON (linter) | `biome` |\n| `eslint` | TS/JS/Vue/Svelte (linter) | `vscode-eslint-language-server` |\n| `vscode-html-language-server` | HTML | `vscode-html-language-server` |\n| `vscode-css-language-server` | CSS, SCSS, Less | `vscode-css-language-server` |\n| `vscode-json-language-server` | JSON | `vscode-json-language-server` |\n| `tailwindcss` | HTML, CSS, TS/JS | `tailwindcss-language-server` |\n| `svelte` | Svelte | `svelteserver` |\n| `vue-language-server` | Vue | `vue-language-server` |\n| `astro` | Astro | `astro-ls` |\n| `pyright` | Python | `pyright-langserver` |\n| `basedpyright` | Python | `basedpyright-langserver` |\n| `pylsp` | Python | `pylsp` |\n| `ruff` | Python (linter) | `ruff` |\n| `jdtls` | Java | `jdtls` |\n| `kotlin-lsp` | Kotlin | `kotlin-lsp` |\n| `metals` | Scala | `metals` |\n| `hls` | Haskell | `haskell-language-server-wrapper` |\n| `ocamllsp` | OCaml | `ocamllsp` |\n| `elixirls` | Elixir | `elixir-ls` |\n| `erlangls` | Erlang | `erlang_ls` |\n| `gleam` | Gleam | `gleam` |\n| `solargraph` | Ruby | `solargraph` |\n| `ruby-lsp` | Ruby | `ruby-lsp` |\n| `rubocop` | Ruby (linter) | `rubocop` |\n| `bashls` | Bash, Zsh | `bash-language-server` |\n| `lua-language-server` | Lua | `lua-language-server` |\n| `intelephense` | PHP | `intelephense` |\n| `phpactor` | PHP | `phpactor` |\n| `csharp-ls` | C# | `csharp-ls` |\n| `omnisharp` | C# | `omnisharp` |\n| `yamlls` | YAML | `yaml-language-server` |\n| `terraformls` | Terraform | `terraform-ls` |\n| `dockerls` | Dockerfile | `docker-langserver` |\n| `helm-ls` | Helm | `helm_ls` |\n| `nixd` | Nix | `nixd` |\n| `nil` | Nix | `nil` |\n| `ols` | Odin | `ols` |\n| `dartls` | Dart | `dart` |\n| `marksman` | Markdown | `marksman` |\n| `texlab` | LaTeX | `texlab` |\n| `graphql` | GraphQL | `graphql-lsp` |\n| `prismals` | Prisma | `prisma-language-server` |\n| `vimls` | Vim script | `vim-language-server` |\n| `emmet-language-server` | HTML, CSS, JSX | `emmet-language-server` |\n| `sourcekit-lsp` | Swift | `sourcekit-lsp` |\n| `swiftlint` | Swift (linter) | `swiftlint` |\n| `tlaplus` | TLA+ | `tlapm_lsp` |\n",
 	"macos-option-key.md": "# macOS + iTerm2 Option/Alt key setup for GJC\n\nHow to make the macOS Option key reach GJC as Alt/Meta input instead of producing composed characters like `œ` or `ˆ`.\n\n## iTerm2 settings\n\nIn **Settings → Profiles → Keys → General** for the profile you use:\n\n- Left Option Key: `+Esc`\n- Right Option Key: `+Esc`\n\nA verified live configuration sets `Option Key Sends = 2` and `Right Option Key Sends = 2` for both the `Default` and `tmux` profiles. iTerm2's `+Esc` setting (`OPTION_KEY_ESC = 2`) prepends ESC to Option input, delivering `Option+Q` as `ESC q` and `Option+I` as `ESC i`. Value `1` is Meta mode and is not an ESC prefix.\n\nIn **Settings → Profiles → Keys → Key Mappings**, remove any `Send Text`, `Send Escape Sequence`, or other mappings that intercept `Option+Q` or `Option+I`, then open a new session for each profile after changing settings.\n\n## macOS input source\n\nSwitch the active keyboard layout to `ABC` when typing GJC Alt commands. `scripts/verify-option-key.sh` does not merely check that ABC is in the list; it verifies `AppleCurrentKeyboardLayoutInputSourceID = com.apple.keylayout.ABC` together with ABC in the selection list.\n\n## Verification\n\n```sh\n./scripts/verify-option-key.sh\npython3 scripts/capture-option-key.py\n```\n\nThe verify script checks the `Default` and `tmux` profiles, both Option keys, Option+Q/I mapping conflicts in both physical-keycode (`Q=12`, `I=34`) and character-code (`q=0x71`, `i=0x69`) form, the active ABC layout, Bun, and the GJC smoke test.\n\nUse the capture tool in a real TTY. In raw mode, Ctrl-C arrives as `0x03` rather than `KeyboardInterrupt`; the tool detects it, restores the terminal, and exits. Confirm with physical key presses in fresh Default and tmux sessions that `Option+Q` and `Option+I` arrive as `ESC q` and `ESC i` and trigger the GJC commands.\n\nRelated live and fixture verification evidence is recorded in `artifacts/option-key-verification.json`.\n",
 	"memory.md": "# Autonomous Memory\n\nWhen enabled, the agent automatically extracts durable knowledge from past sessions and injects a compact summary into each new session. Over time it builds a project-scoped memory store — technical decisions, recurring workflows, pitfalls — that carries forward without manual effort.\n\nDisabled by default. Enable via `/settings` or `config.yml`:\n\n```yaml\nmemories:\n  enabled: true\n```\n\n## Usage\n\n### What gets injected\n\nAt session start, if a memory summary exists for the current project, it is injected into the system prompt as a **Memory Guidance** block. The agent is instructed to:\n\n- Treat memory as heuristic context — useful for process and prior decisions, not authoritative on current repo state.\n- Pair memory-influenced decisions with current-repo evidence before acting.\n- Prefer repo state and user instruction when they conflict with memory; treat conflicting memory as stale.\n\n### Memory artifacts\n\nGenerated local-memory artifacts are private runtime state, not a public tool or URI surface. They may be summarized into the system prompt when local memory is enabled, but users and model-facing tool docs should not rely on direct `memory://` reads. The legacy internal `memory://` resolver remains only for compatibility with existing persisted guidance and is not part of the public coding harness contract; remove it after legacy local-memory prompts no longer reference it.\n### `/memory` slash command\n\n| Subcommand            | Effect                                         |\n| --------------------- | ---------------------------------------------- |\n| `view`                | Show the current memory injection payload      |\n| `clear` / `reset`     | Delete all memory data and generated artifacts |\n| `enqueue` / `rebuild` | Force consolidation to run at next startup     |\n\n## How it works\n\nMemories are built by a background pipeline that runs at startup or when manually triggered via slash command.\n\n**Phase 1 — per-session extraction:** For each past session that has changed since it was last processed, a model reads the session history and extracts durable signal: technical decisions, constraints, resolved failures, recurring workflows. Sessions that are too recent, too old, or currently active are skipped. Each extraction produces a raw memory block and a short synopsis for that session.\n\n**Phase 2 — consolidation:** After extraction, a second model pass reads all per-session extractions and produces three outputs written to disk:\n\n- `MEMORY.md` — a curated long-term memory document\n- `memory_summary.md` — the compact text injected at session start\n- `skills/` — reusable procedural playbooks, each in its own subdirectory\n\nPhase 2 uses a lease to prevent double-running when multiple processes start simultaneously. Stale skill directories from prior runs are pruned automatically.\n\nAll output is scanned for secrets before being written to disk.\n\n### Extraction behavior\n\nMemory extraction and consolidation behavior is driven by static prompt files in `packages/coding-agent/src/prompts/memories/`.\n\n| File                  | Purpose                                     | Variables                                   |\n| --------------------- | ------------------------------------------- | ------------------------------------------- |\n| `stage_one_system.md` | System prompt for per-session extraction    | —                                           |\n| `stage_one_input.md`  | User-turn template wrapping session content | `{{thread_id}}`, `{{response_items_json}}`  |\n| `consolidation.md`    | Prompt for cross-session consolidation      | `{{raw_memories}}`, `{{rollout_summaries}}` |\n| `read_path.md`        | Memory guidance injected into live sessions | `{{memory_summary}}`                        |\n\n### Model selection\n\nMemory piggybacks on the model role system.\n\n| Phase                   | Role                                                                | Purpose                          |\n| ----------------------- | ------------------------------------------------------------------- | -------------------------------- |\n| Phase 1 (extraction)    | `default`                                                           | Per-session knowledge extraction |\n| Phase 2 (consolidation) | `smol` (falls back to `default`, then current/first registry model) | Cross-session synthesis          |\n\nIf the requested memory role is not configured, memory model resolution falls back to the `default` role, then the active session model, then the first model in the registry.\n\n## Configuration\n\n| Setting                               | Default | Description                                               |\n| ------------------------------------- | ------- | --------------------------------------------------------- |\n| `memories.enabled`                    | `false` | Master switch                                             |\n| `memories.maxRolloutAgeDays`          | `30`    | Sessions older than this are not processed                |\n| `memories.minRolloutIdleHours`        | `12`    | Sessions active more recently than this are skipped       |\n| `memories.maxRolloutsPerStartup`      | `64`    | Cap on sessions processed in a single startup             |\n| `memories.summaryInjectionTokenLimit` | `5000`  | Max tokens of the summary injected into the system prompt |\n\nAdditional tuning knobs (concurrency, lease durations, token budgets) are available in config for advanced use.\n\n## Key files\n\n- `packages/coding-agent/src/memories/index.ts` — pipeline orchestration, injection, slash command handling\n- `packages/coding-agent/src/memories/storage.ts` — SQLite-backed job queue and thread registry\n- `packages/coding-agent/src/prompts/memories/` — memory prompt templates\n- `packages/coding-agent/src/internal-urls/memory-protocol.ts` — legacy non-public `memory://` compatibility handler\n",
-	"models.md": "# Model and Provider Configuration (`models.yml`)\n\nThis document describes how the coding-agent currently loads models, applies overrides, resolves credentials, and chooses models at runtime.\n\n## What controls model behavior\n\nPrimary implementation files:\n\n- `src/config/model-registry.ts` — loads built-in + custom models, provider overrides, runtime discovery, auth integration\n- `src/config/model-resolver.ts` — parses model patterns and selects models for the default and agent roles\n- `src/config/settings-schema.ts` — model-related settings (`modelRoles`, provider transport preferences)\n- `src/session/auth-storage.ts` — API key + OAuth resolution order\n- `packages/ai/src/models.ts` and `packages/ai/src/types.ts` — built-in providers/models and `Model`/`compat` types\n\n## Config file location and legacy behavior\n\nDefault config path:\n\n- `~/.gjc/agent/models.yml`\n\nLegacy behavior still present:\n\n- If `models.yml` is missing and `models.json` exists at the same location, it is migrated to `models.yml`.\n- Explicit `.json` / `.jsonc` config paths are still supported when passed programmatically to `ModelRegistry`.\n\n## `models.yml` shape\n\n```yaml\nproviders:\n  <provider-id>:\n    # provider-level config\nequivalence:\n  overrides:\n    <provider-id>/<model-id>: <canonical-model-id>\n  exclude:\n    - <provider-id>/<model-id>\n```\n\n`provider-id` is the canonical provider key used across selection and auth lookup.\n\n`equivalence` is optional and configures canonical model grouping on top of concrete provider models:\n\n- `overrides` maps an exact concrete selector (`provider/modelId`) to an official upstream canonical id\n- `exclude` opts a concrete selector out of canonical grouping\n\n## Provider-level fields\n\n```yaml\nproviders:\n  my-provider:\n    baseUrl: https://api.example.com/v1\n    apiKey: MY_PROVIDER_API_KEY\n    api: openai-completions\n    headers:\n      X-Team: platform\n    authHeader: true\n    auth: apiKey\n    disableStrictTools: false  # set true for Anthropic-compatible endpoints that reject the strict field\n    cacheRetention: short  # none | short | long; model entries and modelOverrides can override this\n    discovery:\n      type: ollama\n    modelOverrides:\n      some-model-id:\n        name: Renamed model\n        cacheRetention: long\n    models:\n      - id: some-model-id\n        name: Some Model\n        api: openai-completions\n        reasoning: false\n        input: [text]\n        cost:\n          input: 0\n          output: 0\n          cacheRead: 0\n          cacheWrite: 0\n        contextWindow: 128000\n        maxTokens: 16384\n        headers:\n          X-Model: value\n        cacheRetention: none\n        thinking:\n          minLevel: low\n          maxLevel: xhigh\n          mode: effort\n          defaultLevel: high\n          levels: [low, medium, high, xhigh]\n        compat:\n          supportsStore: true\n          supportsDeveloperRole: true\n          supportsReasoningEffort: true\n          maxTokensField: max_completion_tokens\n          openRouterRouting:\n            only: [anthropic]\n          vercelGatewayRouting:\n            order: [anthropic, openai]\n          extraBody:\n            gateway: m1-01\n            controller: mlx\nmodelBindings:\n  modelRoles:\n    default: my-provider/some-model-id:high\n  agentModelOverrides:\n    executor: my-provider/some-model-id\n```\n\n### Allowed provider/model `api` values\n\n- `openai-completions`\n- `openai-responses`\n- `openai-codex-responses`\n- `azure-openai-responses`\n- `bedrock-converse-stream`\n- `anthropic-messages`\n- `google-generative-ai`\n- `google-vertex`\n- `google-gemini-cli`\n- `ollama-chat`\n- `cursor-agent`\n\n\n### First-class DeepInfra, Azure OpenAI, and Amazon Bedrock examples\n\nAzure OpenAI uses canonical OpenAI model IDs in GJC and resolves those IDs to Azure deployment names at request time. Set `AZURE_OPENAI_DEPLOYMENT_NAME_MAP` to avoid assuming model id equals deployment name:\n\n```yaml\nproviders:\n  azure-openai:\n    baseUrl: https://my-resource.openai.azure.com/openai/v1\n    apiKeyEnv: AZURE_OPENAI_API_KEY\n    api: azure-openai-responses\n    models:\n      - id: gpt-4.1\n      - id: o3\n```\n\n```sh\nexport AZURE_OPENAI_DEPLOYMENT_NAME_MAP='gpt-4.1=gpt-41-prod,o3=o3-reasoning-prod'\n```\n\nDeepInfra is available as the first-class `deepinfra` provider. It uses DeepInfra's OpenAI-compatible Chat Completions endpoint and reads `DEEPINFRA_API_KEY` when no explicit config key is provided. Set `serviceTier: priority` in GJC config or use the runtime service-tier controls to send DeepInfra's `service_tier: \"priority\"` request field for supported models:\n\n```yaml\nproviders:\n  deepinfra:\n    baseUrl: https://api.deepinfra.com/v1/openai\n    apiKeyEnv: DEEPINFRA_API_KEY\n    api: openai-completions\n    models:\n      - id: deepseek-ai/DeepSeek-V3.2\n```\n\n#### `/fast` provider support\n\n`/fast on` only shows `⚡` when GJC will put a fast/priority field on the selected provider's wire request:\n\n| Provider ID | Wire request | Notes |\n|---|---|---|\n| `openai` | `service_tier: \"priority\"` | OpenAI renamed Priority processing to [Fast mode](https://developers.openai.com/api/docs/guides/fast-mode); `priority` remains an accepted alias. For API-key requests, the response `service_tier` reports the tier actually used and may be `default` after a ramp-rate downgrade. |\n| `openai-codex` | `service_tier: \"priority\"` | ChatGPT-authenticated Codex handles Fast through server-side routing. A final response value of `service_tier: \"default\"` does not show that Fast was ignored or downgraded. |\n| `anthropic` | `speed: \"fast\"` plus `fast-mode-2026-02-01` beta | Direct Claude API only. Anthropic's [Fast mode](https://platform.claude.com/docs/en/build-with-claude/fast-mode) is model- and account-gated; unsupported or unavailable requests can fall back after a provider rejection. Bedrock, Vertex, and Microsoft Foundry do not support it. |\n| `deepinfra` | `service_tier: \"priority\"` | Sent only for the first-class `deepinfra` provider ID and only with the `priority` tier. |\n| `opencodex` | `service_tier: \"priority\"` | First-class OpenCodex discovery opts in automatically; OpenCodex Fast Mode must remain `Auto` for client passthrough. When OpenCodex uses ChatGPT authentication, a final `service_tier: \"default\"` is not downgrade evidence. |\n\nCustom OpenAI-compatible providers remain fail-closed unless their provider or model configuration explicitly sets `compat.supportsServiceTier: true`. Use that opt-in only when the proxy preserves or intentionally realizes OpenAI's `service_tier` contract:\n\n```yaml\nproviders:\n  my-openai-proxy:\n    baseUrl: http://proxy.example/v1\n    api: openai-responses\n    compat:\n      supportsServiceTier: true\n```\n\nWithout that capability, `/fast status` shows `off` even when the session retains an unscoped `priority` intent. The `⚡` indicator means that GJC sends the provider's fast request field. API-key providers may report a downgrade in their response; ChatGPT-authenticated Codex and OpenCodex route Fast server-side and cannot be verified from the final `service_tier` value.\n\nAmazon Bedrock uses the native `bedrock-converse-stream` transport and AWS credential chain auth. Do not put AWS access keys in `models.yml`; configure `AWS_REGION` / `AWS_PROFILE` or standard static AWS credential environment variables instead:\n\n```yaml\nproviders:\n  amazon-bedrock:\n    baseUrl: https://bedrock-runtime.us-east-1.amazonaws.com\n    api: bedrock-converse-stream\n    models:\n      - id: us.anthropic.claude-opus-4-6-v1\n      - id: anthropic.claude-3-5-sonnet-20241022-v2:0\n```\n\n### Coding-plan provider presets\n\nFor supported coding-plan providers, prefer presets so the API type, base URL, environment variable, model catalog, discovery behavior, and compatibility flags are written together:\n\n```sh\ngjc setup provider --preset minimax\ngjc setup provider --preset minimax-cn\ngjc setup provider --preset glm\ngjc setup provider --preset alibaba-token-plan\ngjc setup provider --preset cline-pass\ngjc setup provider --preset commandcode-goat\n```\n\nThe same presets are available inside the TUI:\n\n```text\n/provider add --preset minimax\n/provider add --preset glm\n/provider add zai\n/provider add --preset alibaba-token-plan\n/provider add --preset cline-pass\n/provider add --preset commandcode-goat\n```\n\nPresets only write `models.yml` entries that reference documented environment variable names (`MINIMAX_CODE_API_KEY`, `MINIMAX_CODE_CN_API_KEY`, `ZAI_API_KEY`, `ALIBABA_TOKEN_PLAN_API_KEY`, `CLINE_API_KEY`, or `CMD_API_KEY`); they do not store or validate real credentials. The GLM preset aliases (`glm`, `zai`, `z-ai`) write an OpenAI-compatible custom provider named `glm-proxy` and do not replace the first-class `zai` provider. The Alibaba Token Plan preset (aliases: `alibaba`, `token-plan`) writes an OpenAI-compatible custom provider named `alibaba-token-plan` with per-model API routing. The ClinePass preset (aliases: `clinepass`, `cline`) does not hardcode models: Cline's inference API has no working `/models` route, so GJC follows Cline's own catalog-generation source and fetches the live `cline-pass` provider catalog from `https://models.dev/api.json`. The Command Code GOAT preset (aliases: `commandcode`, `command-code`, `goat`) fetches its live `/provider/v1/models` catalog, routes every current or future `claude-*` model through Anthropic Messages, and routes other models through Chat Completions. Create the corresponding API key in the provider dashboard before inference; plan entitlement is enforced by the provider.\n\n## Model profiles (`--mpreset`)\n\nModel profiles are optional top-level `profiles:` entries in `~/.gjc/agent/models.yml`. A profile can require provider credentials before activation and can map one or more model roles; omitted roles inherit from the active defaults.\n\n> See also: [Cross-vendor role-based profiles](./multi-vendor-profiles.md) — a curated multi-vendor `profiles:` recipe and verified selector notes that build on the mechanism described here.\n\n```yaml\nprofiles:\n  team-standard:\n    required_providers: [openai, anthropic]\n    model_mapping:\n      default: openai/gpt-5.2\n      executor: anthropic/claude-sonnet-5:medium\n      architect: openai/o3:high\n      planner: openai/o3:high\n      critic: openai/o3:high\n```\n\n`model_mapping` keys are role names (`default`, `executor`, `architect`, `planner`, `critic`). Every role accepts either one selector or a non-empty ordered array of selectors; the first entry is primary and later entries are fallback candidates. A selector may be a provider-agnostic bare alias such as `glm-5.2[:effort]` or an explicit `provider/modelId[:effort]` pin, including nested model IDs such as `openrouter/anthropic/claude-sonnet-5`. `required_providers` lists explicit provider prerequisites and may be empty when availability is resolved from bare aliases.\n\n### Fallback chains\n\nPreset `model_mapping` roles, top-level `modelRoles`, and `task.agentModelOverrides` all accept `string | string[]`. Keep one selector per line when a chain needs to be readable:\n\n```yaml\nprofiles:\n  reliable:\n    required_providers: [anthropic, openai]\n    model_mapping:\n      default: [anthropic/claude-sonnet-4-5, openai/gpt-4o-mini]\nmodelBindings:\n  modelRoles:\n    default: [anthropic/claude-sonnet-4-5, openai/gpt-4o-mini]\n  agentModelOverrides:\n    executor: [anthropic/claude-sonnet-4-5, openai/gpt-4o-mini]\n```\n\nResolution-time skips for unavailable, unauthenticated, or unknown entries cost zero attempts and advance immediately. Only request-time retryable failures (such as 429, quota, authentication, or 5xx failures) consume an entry's `fallback.maxAttempts` total attempts (default: `3`). The active default fallback remains sticky for the session; role-override fallback state is fresh for each subagent call. The active model is shown consistently in status and `/model`.\n\nManaged fallback attempts buffer provisional streamed output until an attempt is accepted, so output can appear later than it does for a one-model stream. Current Cursor-agent transports are fail-closed unavailable in retryable fallback chains: resolution rejects them with `Cursor model <selector> requires provider-side tool execution and cannot be used in a retryable fallback chain` because they do not provide a client-side tool-call mode.\n\nCancellation discards provisional output and emits exactly one cancelled `agent_end`; RPC, ACP, and the TUI therefore settle once. On load, the source-aware one-shot migration reads legacy `retry.fallbackChains`, prepends the effective role chain, and writes the ordered, deduplicated result to the corresponding role array; the legacy key is then ignored.\n\nBuilt-in profiles are grouped by provider mix and tier:\n\n- `codex-{eco,medium,pro}` — GPT-5.6 Sol/Terra/Luna role mixes tuned by tier and reasoning effort; `lunamaxxing` — OpenAI Codex Luna-only profile with maximum reasoning on delegated roles\n- `opencodego` — single OpenCode Go preset (Kimi K3 default and planner, DeepSeek executor/architect, MiMo critic)\n- Provider-agnostic open-model profiles are named by the model families they require. Single-family choices are `open-weights-{glm,deepseek,kimi,luna}`; two-family choices are `open-weights-glm-deepseek`, `open-weights-kimi-deepseek`, and `open-weights-kimi-glm`; `open-weights-kimi-glm-deepseek` uses all three open-weight families; `open-weights-all` adds GPT-5.6 Luna. Choose the smallest combination covered by the models available through your configured providers. Every selector is a bare final-segment alias with `required_providers: []`, so each family may come from any authenticated bundled or custom provider under Provider Priority. GPT-5.6 Luna is proprietary despite its inclusion in this group.\n- `claude-opus` — Anthropic OAuth preset centered on `claude-opus-5`\n- Single-provider tiers: `glm-{eco,medium,pro}`, `kimi-coding-plan-{eco,medium,pro}`, `mimo-{eco,medium,pro}`, `grok-{eco,medium,pro}`, `cursor-{eco,medium,pro}`, `minimax-{eco,medium,pro}`\n- Alibaba Token Plan: `alibaba-token-plan-balanced` preserves the established Qwen/DeepSeek V4 Pro/GLM mix; `alibaba-token-plan-pro` raises execution and independent criticism with DeepSeek V4 Flash 0731 max and GLM xhigh; `alibaba-token-plan-qwenmaxxing` stays Qwen-only; `alibaba-token-plan-qwen-deepseek` keeps Qwen 3.8 Max (`qwen3.8-max`) on the expensive default (high)/architect (xhigh)/critic (xhigh) roles and spends DeepSeek V4 Flash 0731 on the cheap planner (max) and executor (high) roles; `alibaba-token-plan-glm-deepseek` does the same with GLM 5.2 (`glm-5.2`) as the expensive model\n- Combos: `opus-codex`, `codex-opencodego`, and `fable-opus-codex`\n\nThe `eco`, `medium`, and `pro` Codex profile mappings are current product judgments: Eco assigns Terra low/Luna low/Luna high/Terra xhigh/Terra high to default/executor/planner/critic/architect; Medium assigns Sol low/Terra low/Terra high/Sol xhigh/Sol high; Pro assigns Sol medium/Terra medium/Sol high/Sol max/Sol xhigh; and LunaMaxxing assigns Luna medium/Luna xhigh/Luna max/Luna max/Luna max. `opus-codex` retains the Medium Codex executor, critic, and architect roles but uses `anthropic/claude-sonnet-5` for planner; `codex-opencodego` retains the Medium Codex default and architect roles; and `fable-opus-codex` uses the Pro Codex executor and architect roles with `anthropic/claude-opus-5:medium` for planner. The descriptive repeated local exact-edit evidence informs only selected executor-style TypeScript tasks; it does not evaluate or prove default, planner, architect, or critic performance. See [GPT-5.6 Codex preset benchmark](./gpt-5.6-codex-preset-benchmark.md). The Alibaba Pro role evidence and its limits are recorded separately in [Alibaba Token Plan Pro profile benchmark](./alibaba-token-plan-pro-profile-benchmark.md). Cursor Eco uses Composer 2.5 for every role; Medium keeps standard Composer for default/planning and spends the Fast premium on execution, criticism, and architecture; Pro uses Composer 2.5 Fast throughout. Composer does not expose a strength value through the current Cursor RPC, so these profiles use exact model IDs without inert generic effort suffixes. See [Cursor Composer profile tiers](./cursor-composer-profile-tiers.md). Effort suffixes are clamped to each model's supported thinking range at preview and activation time. Single-provider tiers pin each provider's current flagship (`zai/glm-5.2`, `kimi-code/kimi-k2.7-code`, `xiaomi/mimo-v2.5-pro`, `xai/grok-4.3`, `cursor/composer-2.5`, `minimax-code/MiniMax-M3`). User-defined profiles override built-ins by exact profile name.\n\n\nUse `gjc --mpreset <name>` to activate a profile for the current session only. Activation hard-blocks when any provider listed in `required_providers` lacks credentials. Add `--default` to persist the selected profile as `modelProfile.default` in `config.yml`, so it applies at startup:\n\n```sh\ngjc --mpreset codex-medium\ngjc --mpreset opencodego --default\n```\n\n### Routing built-in presets through a proxy (`modelProfile.proxyProvider`)\n\nBuilt-in preset selectors pin a direct provider endpoint (`xai/grok-4.3`, `xiaomi/mimo-v2.5-pro`, …). To serve those models through your own OpenAI-compatible gateway (LiteLLM, OpenRouter, or a custom proxy) instead of each vendor's endpoint, configure the proxy provider id and routing mode in `config.yml`:\n\n```yaml\nmodelProfile:\n  proxyProvider: litellm\n  proxyMode: always # use fallback to keep directly authenticated providers direct\n```\n\nThe proxy provider is a normal `providers:` entry. Add it with `gjc setup provider --preset litellm --base-url <url>` or the generic `gjc setup provider --preset openai-compatible-proxy --base-url <url>` (both presets require `--base-url` and use live model discovery). The configured proxy must be authenticated and expose every routed model. Activation rewrites each selected built-in preset selector from `<direct-provider>/<model>` to `<proxy>/<direct-provider>/<model>` (for example `xai/grok-4.3` → `litellm/xai/grok-4.3`), matching the proxy's catalog entry for the model. The rules:\n\n- Routing applies to **built-in presets only**. User-defined `profiles:` entries always keep their exact selectors — set them explicitly if you want them proxied.\n- `proxyMode: fallback` (the default) routes only selectors whose direct provider is unauthenticated. `proxyMode: always` routes every proxy-routable built-in selector through the configured proxy, including selectors with direct credentials.\n- The proxy id must name a configured provider. `proxyMode: always` requires `proxyProvider` and a usable proxy credential; activation fails closed when a required proxy is unset or unauthenticated. `auth: none` proxies count as authenticated.\n- Only providers the bundled preset catalog treats as routable are rewritten; providers outside that set (for example a custom `acme-private`) keep the direct credential error.\n- A routed selector must have exactly one matching proxy catalog model. Exact `<direct-provider>/<model>` proxy ids win over suffix matches; missing or ambiguous matches fail activation before any role can run.\n\nThe `/model` command opens to a preset landing view: presets are grouped by provider with live auth marks (✓/✗), highlighting a group expands its tiers, and selecting a tier shows the full role→model preview before applying for the session or as default. Typing jumps straight to model search, and `Browse all models` opens the classic tabbed model selector. In `/login`, `Add custom provider` is the first option for configuring credentials needed by custom or profile-required providers; after a successful provider login, the matching preset is recommended automatically. Custom providers participate in provider-agnostic alias resolution but require manual preset selection.\n\nExternal SDK/ACP clients (e.g. the Paseo TUI) can select profiles like ordinary models: the SDK `models.list/current` (Q10) catalog exposes every usable profile as a synthetic `gajae-code/<profile>` entry (e.g. `gajae-code/codex-eco`), and selecting one through `model.set` (or the ACP Model picker) activates the profile for the live session only. Persisting a profile remains an explicit TUI choice, mirroring `gjc --mpreset <name> --default`. See [SDK model profiles](./sdk.md#model-profiles-as-synthetic-models-gajae-codeprofile).\n\nMiniMax's OpenAI-compatible endpoint rejects multiple system messages and emits thinking in `reasoning_content`, so pin the public-safe compatibility fields when hand-authoring a custom provider:\n\n```yaml\nproviders:\n  minimax-custom:\n    baseUrl: https://api.minimax.io/v1\n    apiKeyEnv: MINIMAX_API_KEY\n    api: openai-completions\n    compat:\n      supportsStore: false\n      supportsDeveloperRole: false\n      supportsReasoningEffort: false\n      reasoningContentField: reasoning_content\n    models:\n      - id: MiniMax-M2.5\n```\n\nGLM via z.ai is available as the first-class `zai` provider. For a private GLM-compatible proxy, keep secrets in an env var and disable OpenAI-only request fields as needed:\n\n```yaml\nproviders:\n  glm-proxy:\n    baseUrl: https://api.z.ai/api/paas/v4\n    apiKeyEnv: ZAI_API_KEY\n    api: openai-completions\n    compat:\n      supportsDeveloperRole: false\n      supportsReasoningEffort: false\n    models:\n      - id: glm-4.6\n```\n\n### JetBrains AI (Junie)\n\n`jetbrains-junie` is a first-class provider serving JetBrains-hosted models through the documented\nIngrazzio gateway (`https://ingrazzio-cloud-prod.labs.jb.gg`).\n\nAuthenticate with an access token generated at [junie.jetbrains.com/cli](https://junie.jetbrains.com/cli):\n\n```sh\nexport JUNIE_API_KEY=...\n```\n\nThe token is sent as `Authorization: Bearer` — JetBrains AI rejects requests that also carry `x-api-key`, so\nthis provider never lets the Anthropic SDK attach one. Usage is billed against your JetBrains AI\nsubscription, so bundled per-token costs are zero. There is no OAuth login flow; the environment variable is\nthe only supported credential source.\n\nThe gateway multiplexes transports by model family:\n\n| Family | Models | Transport | Prompt limit |\n| --- | --- | --- | --- |\n| Claude | `claude-sonnet-4-6` (default), `claude-sonnet-5`, `claude-opus-4-6`, `claude-opus-4-7`, `claude-opus-4-8`, `claude-opus-5`, `claude-fable-5` | `anthropic-messages` | 1M |\n| GPT | `gpt-5-2025-08-07`, `gpt-5.2-2025-12-11`, `gpt-5.4`, `gpt-5.5`, `gpt-5.6-luna`, `gpt-5.6-sol`, `gpt-5.6-terra` | `openai-completions` | 922K |\n| GPT (Responses-only) | `gpt-5.3-codex` | `openai-responses` | 272K |\n\nAll models cap output at 128K. Junie also exposes Gemini and Grok, but those ride a proprietary Grazie\ntranslation protocol that GJC does not implement, so they are deliberately not bundled. The bare\n`opus`/`sonnet`/`gpt`/`grok` aliases are Junie CLI shorthands the gateway itself rejects.\n\n### Allowed auth/discovery values\n\n- `auth`: `apiKey` (default), `none`, or `oauth`; for `models.yml` custom models, `oauth` is accepted by schema but does not waive the `apiKey` requirement\n- `models.yml` is strict: unknown provider/model keys fail validation before provider dispatch, so stale keys such as `requestTransform` or `wireModelId` only work where this document lists them.\n- `discovery.type`: `ollama`, `llama.cpp`, `lm-studio`, or `openai-models-list`\n- `cacheRetention`: `none`, `short`, or `long`; request-time options win over model/modelOverride values, then provider values, then `GJC_CACHE_RETENTION`, then the runtime default. The runtime default is `short` for most providers, but the Anthropic provider defaults to `long` because the ~5m cache is fragile for long-running subagent workflows. Canonical Anthropic models use top-level automatic caching and emit `ttl: \"1h\"` when long retention is supported. Claude-family models on non-canonical Anthropic-compatible endpoints default to explicit block markers because compatible proxies commonly inject, rewrite, or reject top-level cache controls; they omit `ttl` unless `compat.supportsLongCacheRetention: true` opts the endpoint into 1-hour retention. For OpenAI Responses, this controls `prompt_cache_retention` only; it does not disable `prompt_cache_key` when a stable session id exists.\n\n## OpenAI-compatible proxy configuration\n\nOpenAI-compatible proxy providers should use schema-supported provider keys first:\nThe first-class way to add a proxy provider is `gjc setup provider --preset litellm --base-url <url>` (LiteLLM) or `gjc setup provider --preset openai-compatible-proxy --base-url <url>` (any OpenAI-compatible gateway); both presets require `--base-url` and configure live model discovery. Proxy providers can also be used to route built-in model-preset selectors — see [Routing built-in presets through a proxy](#routing-built-in-presets-through-a-proxy-modelprofileproxyprovider). The YAML below shows the equivalent hand-written provider config:\n\n```yaml\nproviders:\n  proxy-provider:\n    baseUrl: https://api.proxy.example/v1\n    apiKeyEnv: PROXY_API_KEY\n    api: openai-completions\n    auth: apiKey\n    headers:\n      User-Agent: curl/8.7.1\n    models:\n      - id: local-gpt\n        name: Local GPT\n        reasoning: true\n        input: [text]\n        cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 }\n        contextWindow: 400000\n        maxTokens: 128000\n```\n\nUse provider-level `headers` for proxy-required headers. Keep the provider `api` set to `openai-completions` when the proxy exposes Chat Completions-compatible `/v1/chat/completions` semantics. `auth: apiKey` sends the resolved token as bearer auth; use `auth: none` only for trusted local/no-auth endpoints.\n\n`auth` selects the transport scheme only; it never supplies a credential. A provider that declares `models:` must therefore also declare where its key comes from, and `models.yml` validation rejects the config before model discovery otherwise:\n\n| Intent | Required keys |\n| --- | --- |\n| Authenticated proxy (recommended) | `auth: apiKey` (default) + `apiKeyEnv: MY_TOKEN` |\n| Authenticated proxy, key inline | `auth: apiKey` (default) + `apiKey: sk-…` (less safe; stored in plaintext) |\n| Genuinely unauthenticated endpoint | `auth: none`, no key |\n\nOmitting both `apiKey` and `apiKeyEnv` while leaving `auth` at its `apiKey` default fails with `Provider <name>: custom models need a credential source, but none is configured.` — the fix is to add one of the rows above, not to change `api` or `baseUrl`.\n\n`input` is the model modality list GJC uses to decide whether image content is forwarded. When a custom model omits `input`, GJC defaults to `[text]` (unless a bundled model with the same id contributes a reference). Vision-capable upstream models therefore need an explicit `input: [text, image]`; otherwise `read`/tool images are stripped before the request and replaced with `[image omitted: model does not support vision]`, even if the remote model can see images.\n\n```yaml\nproviders:\n  ali:\n    baseUrl: https://token-plan.ap-southeast-1.maas.aliyuncs.com/compatible-mode/v1\n    apiKeyEnv: ALI_API_KEY\n    api: openai-completions\n    auth: apiKey\n    models:\n      # id-only → text-only; images will be omitted\n      - id: some-text-model\n      # vision-capable hosted model must declare image input\n      - id: qwen3.8-max-preview\n        name: Qwen3.8 Max Preview\n        reasoning: true\n        input: [text, image]\n```\n\n`requestTransform` and `wireModelId` remain supported for request-body shaping, but they are not needed for ordinary OpenAI-compatible proxies whose local model id is already the upstream wire id. Unknown config keys fail validation before a provider request is sent.\n\nWhen request shaping is needed:\n\n- `requestTransform.profile: openai-proxy` strips OpenAI SDK/Stainless telemetry and beta headers at final fetch time and sets a generic GJC user agent.\n- `stripHeaders` replaces the preset strip list when provided.\n- `setHeaders` is applied after stripping; use `null` to remove a header.\n- `extraBody` is shallow-merged into the JSON request body after provider compatibility fields; core transport keys such as `model`, `messages`/`input`, `stream`, `tools`, and `tool_choice` are protected and ignored.\n- Model-level `requestTransform` overrides provider-level fields and shallow-merges `setHeaders`/`extraBody`.\n- `wireModelId` changes only the upstream request body model id; local selection still uses `provider/id`.\n\n### Layofflabs-style proxy example\n\n```yaml\nproviders:\n  layofflabs:\n    baseUrl: https://api.layofflabs.com/v1\n    apiKeyEnv: OPENAI_API_KEY\n    api: openai-completions\n    auth: apiKey\n    headers:\n      User-Agent: curl/8.7.1\n    models:\n      - id: gpt-5.5\n        name: GPT 5.5 via Layofflabs\n        reasoning: true\n        thinking:\n          minLevel: low\n          maxLevel: xhigh\n          mode: effort\n          defaultLevel: high\n          levels: [low, medium, high, xhigh]\n        input: [text]\n        cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 }\n        contextWindow: 400000\n        maxTokens: 128000\n\nmodelBindings:\n  modelRoles:\n    default: layofflabs/gpt-5.5:high\n  agentModelOverrides:\n    executor: layofflabs/gpt-5.5:high\n```\n\n## Validation rules (current)\n\n### Full custom provider (`models` is non-empty)\n\nRequired:\n\n- `baseUrl`\n- A credential source: `apiKeyEnv` or `apiKey`. `auth` selects the scheme, not the credential, so `auth: apiKey` (the default) still needs one of them. Exempt: `auth: none`, and `api: bedrock-converse-stream`, which resolves AWS credentials from its own chain.\n- `api` at provider level or each model\n\n### Override-only provider (`models` missing or empty)\n\nMust define at least one of:\n\n- `baseUrl`\n- `headers`\n- `compat`\n- `requestTransform`\n- `disableStrictTools`\n- `modelOverrides`\n- `discovery`\n\n### Discovery\n\n- `discovery` requires provider-level `api`.\n\n### Model value checks\n\n- `id` required\n- `contextWindow` and `maxTokens` must be positive if provided\n- unknown provider, model, override, and request-transform keys fail schema validation; remove stale keys instead of relying on them being ignored.\n\n## Merge and override order\n\nModelRegistry pipeline (on refresh):\n\n1. Load built-in providers/models from `@gajae-code/ai`.\n2. Load `models.yml` custom config.\n3. Apply provider overrides (`baseUrl`, `headers`, `requestTransform`, `disableStrictTools`, `cacheRetention`) to built-in models.\n4. Apply `modelOverrides` (per provider + model id).\n5. Merge custom `models`:\n   - same `provider + id` replaces existing\n   - otherwise append\n6. Load cached/runtime-discovered models (Ollama, llama.cpp, LM Studio, plus built-in provider managers), then re-apply model overrides.\n\n### Provider-model cache and static fingerprint\n\nCached per-provider model lists are persisted in the model-cache SQLite\ndatabase (schema v3) with a `static_fingerprint` column that hashes the\nstatic catalog slice merged into the row. When `resolveProviderModels`\nskips the network fetch and the fingerprint of the in-memory static\ncatalog matches the cached one, the cached rows are returned verbatim —\nthe static + dynamic merge is bypassed entirely. The fingerprint is\nmemoized per process via a WeakMap keyed by the static-models array\nreference, so repeated cold-start calls do not re-hash.\n\n## Canonical model equivalence and coalescing\n\nThe registry keeps every concrete provider model and then builds a canonical layer above them.\n\nCanonical ids are official upstream ids only, for example:\n\n- `anthropic-model-opus-4-6`\n- `anthropic-model-haiku-4-5`\n- `gpt-5.3-openai-code`\n\n### `models.yml` equivalence config\n\nExample:\n\n```yaml\nproviders:\n  zenmux:\n    baseUrl: https://api.zenmux.example/v1\n    apiKey: ZENMUX_API_KEY\n    api: openai-codex-responses\n    models:\n      - id: openai-code\n        name: Zenmux OpenAI code\n        reasoning: true\n        input: [text]\n        cost:\n          input: 0\n          output: 0\n          cacheRead: 0\n          cacheWrite: 0\n        contextWindow: 200000\n        maxTokens: 32768\n\nequivalence:\n  overrides:\n    zenmux/openai-code: gpt-5.3-openai-code\n    p-openai-code/openai-code: gpt-5.3-openai-code\n  exclude:\n    - demo/openai-code-preview\n```\n\nBuild order for canonical grouping:\n\n1. exact user override from `equivalence.overrides`\n2. bundled official-id matches from built-in model metadata\n3. conservative heuristic normalization for gateway/provider variants\n4. fallback to the concrete model's own id\n\nCurrent heuristics are intentionally narrow:\n\n- embedded upstream prefixes can be stripped when present, for example `anthropic/...` or `openai/...`\n- dotted and dashed version variants can normalize only when they map to an existing official id, for example `4.6 -> 4-6`\n- ambiguous families or versions are not merged without a bundled match or explicit override\n\n### Canonical and preset-equivalent resolution behavior\n\nWhen multiple concrete variants are eligible for automatic resolution, the global provider policy uses this order:\n\n1. explicit `config.yml` `modelProviderOrder` entries, in their saved order\n2. omitted providers whose effective credential came from OAuth\n3. omitted providers using a manual API key, unknown credential provenance, or keyless access\n4. vision capability, exact canonical identity, canonical source quality, lowest `cost.input + cost.cacheRead`, stable registry model order, then concrete selector order\n\nThe explicit provider list may be partial. A listed API-key provider beats every omitted OAuth provider. Resetting Provider Priority clears the explicit list and restores OAuth-first plus deterministic fallback. Saved providers that are not currently available remain visible and persisted, but runtime resolution skips them.\n\nModel-profile and preset assignments support a lookup-only alias when the assignment does not explicitly name a provider. For example, a bare preset assignment `gpt-5` can select any available concrete variant whose final model-id segment is `gpt-5`, ranked by the same global policy. An assignment such as `openai/gpt-5` is an explicit provider pin: if that exact model is unavailable, activation reports it unavailable instead of switching providers. Alias lookup never rewrites the selected model's concrete provider, full model id, or `wireModelId`; a known bare alias with no eligible variant is unavailable and does not fall through to a different fuzzy match. Direct model selection remains unchanged.\nRuntime custom providers participate in the same lookup automatically. A custom provider model such as `hosted/glm-5.2` contributes the alias `glm-5.2` when the provider is registered, authenticated, and available; Provider Priority can rank that custom provider ahead of bundled providers without changing the preset. Custom IDs whose final segment differs (for example `glm-5.2-special`) do not join the alias.\n\nA session that resolves a canonical or alias selector keeps its concrete variant across provider-priority edits and discovery refreshes. Updated priority applies to new or unpinned resolutions. The session re-ranks only after explicit reselection or when the sticky variant becomes unavailable. Session state and transcripts continue to record the concrete provider/model that executed the turn.\n\nProvider defaults vs per-model overrides:\n\n- Provider `headers` are baseline.\n- Model `headers` override provider header keys.\n- `modelOverrides` can override model metadata (`name`, `reasoning`, `input`, `cost`, `contextWindow`, `maxTokens`, `headers`, `compat`, `contextPromotionTarget`).\n- `compat` is deep-merged for nested routing blocks (`openRouterRouting`, `vercelGatewayRouting`, `extraBody`).\n\n## Runtime discovery integration\n\n### Implicit Ollama discovery\n\nIf `ollama` is not explicitly configured, registry adds an implicit discoverable provider:\n\n- provider: `ollama`\n- api: `openai-responses`\n- base URL: `OLLAMA_BASE_URL` or `http://127.0.0.1:11434`\n- auth mode: keyless (`auth: none` behavior)\n\nRuntime discovery calls Ollama endpoints and normalizes discovered OpenAI-compatible models to `openai-responses`.\n\n### Implicit llama.cpp discovery\n\nIf `llama.cpp` is not explicitly configured, registry adds an implicit discoverable provider:\n\n- provider: `llama.cpp`\n- api: `openai-responses`\n- base URL: `LLAMA_CPP_BASE_URL` or `http://127.0.0.1:8080`\n- auth mode: keyless (`auth: none` behavior)\n\nRuntime discovery calls llama.cpp model endpoints and synthesizes model entries with local defaults.\n\n### Implicit LM Studio discovery\n\nIf `lm-studio` is not explicitly configured, registry adds an implicit discoverable provider:\n\n- provider: `lm-studio`\n- api: `openai-completions`\n- base URL: `LM_STUDIO_BASE_URL` or `http://127.0.0.1:1234/v1`\n- auth mode: keyless (`auth: none` behavior)\n\nRuntime discovery fetches models (`GET /models`) and synthesizes model entries with local defaults.\n\n### Explicit provider discovery\n\nYou can configure discovery yourself:\n\n```yaml\nproviders:\n  ollama:\n    baseUrl: http://127.0.0.1:11434\n    api: openai-responses\n    auth: none\n    discovery:\n      type: ollama\n\n  llama.cpp:\n    baseUrl: http://127.0.0.1:8080\n    api: openai-responses\n    auth: none\n    discovery:\n      type: llama.cpp\n```\n\n### Extension provider registration\n\nExtensions can register providers at runtime (`pi.registerProvider(...)`), including:\n\n- model replacement/append for a provider\n- custom stream handler registration for new API IDs\n- custom OAuth provider registration\n\n## Auth and API key resolution order\n\nWhen requesting a key for a provider, effective order is:\n\n1. Runtime override (CLI `--api-key`)\n2. Stored API key credential in `agent.db`\n3. Stored OAuth credential in `agent.db` (with refresh)\n4. Environment variable mapping (`OPENAI_API_KEY`, `ANTHROPIC_API_KEY`, etc.)\n5. ModelRegistry fallback resolver (provider `apiKey` from `models.yml`, env-name-or-literal semantics)\n\n`models.yml` `apiKey` behavior:\n\n- Value is first treated as an environment variable name.\n- If no env var exists, the literal string is used as the token.\n\nIf `authHeader: true` and provider `apiKey` is set, models get:\n\n- `Authorization: Bearer <resolved-key>` header injected.\n\nKeyless providers:\n\n- Providers marked `auth: none` are treated as available without credentials.\n- `getApiKey*` returns `kNoAuth` for them.\n\n### Broker mode\n\nWhen `GJC_AUTH_BROKER_URL` (or `auth.broker.url`) is set, the local SQLite credential store is replaced by `RemoteAuthCredentialStore`. Layers 2 and 3 above (stored API key / OAuth in `agent.db`) are served from a broker-supplied snapshot whose `refresh` tokens are redacted; expiry triggers `POST /v1/credential/:id/refresh` on the broker rather than a local refresh.\n\n`AuthStorage.setConfigApiKey` lets a `models.yml` `apiKey` win over a broker-resolved OAuth token without overriding a runtime `--api-key`. See [`auth-broker-gateway.md`](./auth-broker-gateway.md) for the full broker / gateway design and env surface (`GJC_AUTH_BROKER_URL`, `GJC_AUTH_BROKER_TOKEN`, `auth.broker.url`, `auth.broker.token`).\n\n## Model availability vs all models\n\n- `getAll()` returns the loaded model registry (built-in + merged custom + discovered).\n- `getAvailable()` filters to models that are keyless or have resolvable auth.\n\nSo a model can exist in registry but not be selectable until auth is available.\n\n## Runtime model resolution\n\n### CLI and pattern parsing\n\n`model-resolver.ts` supports:\n\n- exact `provider/modelId`\n- exact canonical model id\n- exact model id (provider inferred)\n- fuzzy/substring matching\n- glob scope patterns in `--models` (e.g. `openai/*`, `*sonnet*`)\n- optional `:thinkingLevel` suffix (`off|minimal|low|medium|high|xhigh`)\n\n`--provider` is legacy; `--model` is preferred.\n\nResolution precedence for exact selectors:\n\n1. exact `provider/modelId` bypasses coalescing\n2. exact canonical id resolves through the canonical index\n3. exact bare concrete id still works\n4. fuzzy and glob matching run after the exact paths\n\nThinking suffixes are split once from the final `:` only after the complete selector does not resolve. This preserves concrete OpenRouter route IDs such as `openrouter/z-ai/glm-4.7:nitro`; `:high` can follow that route suffix. Multiple suffixes are not recursively consumed. A complete `provider/modelId` selector is exact-only: it never falls back to fuzzy, substring, glob, or another provider when that concrete selector is absent. Exact-case provider/model entries resolve deterministically for custom replacement semantics; a case-insensitive selector that remains ambiguous does not guess.\n\nPreset/profile activation may use an unqualified assignment as a final-segment lookup alias after exact resolution fails. Provider-qualified assignments remain exact pins, and this opt-in does not apply to CLI or direct concrete model selection.\n\n### Initial model selection priority\n\n`findInitialModel(...)` uses this order:\n\n1. explicit CLI provider+model\n2. first scoped model (if not resuming)\n3. saved default provider/model\n4. known provider defaults (e.g. OpenAI/Anthropic/etc.) among available models\n5. first available model\n\n### Role aliases and settings\n\nSupported model roles:\n\n- `default` plus the agent assignment targets `executor`, `architect`, `planner`, `critic`\n\nRole aliases like `pi/default` expand through `settings.modelRoles`. Each role value can also append a thinking selector such as `:minimal`, `:low`, `:medium`, or `:high`.\n\nIf a role points at another role, the target model still inherits normally and any explicit suffix on the referring role wins for that role-specific use.\n\nRelated settings:\n\n- `modelRoles` (record)\n- `enabledModels` (scoped pattern list)\n- `modelProviderOrder` (global automatic provider precedence; editable under Settings → Providers with add/remove, move up/down, unavailable-entry retention, and Reset)\n- `providers.kimiApiFormat` (`openai` or `anthropic` request format)\n- `providers.openaiWebsockets` (`auto|off|on` websocket preference for OpenAI code provider transport)\n\n`modelRoles` may store either:\n\n- `provider/modelId` to pin a concrete provider variant\n- a canonical id such as `gpt-5.3-openai-code` to allow provider coalescing\n\nFor `enabledModels` and CLI `--models`:\n\n- exact canonical ids expand to all concrete variants in that canonical group\n- explicit `provider/modelId` entries stay exact\n- globs and fuzzy matches still operate on concrete models\n\nGlobal `enabledModels` and `disabledProviders` entries may also be scoped to a path prefix:\n\n```yaml\nenabledModels:\n  - anthropic-model-sonnet-4-5\n  - path: ~/work\n    models:\n      - anthropic/anthropic-model-opus-4-5\ndisabledProviders:\n  - ollama\n  - path: ~/private\n    providers:\n      - anthropic\n```\n\nString entries apply everywhere. Scoped entries apply when the current working directory is the configured path or one of its subdirectories. Use `path`, `paths`, `pathPrefix`, or `pathPrefixes`; use `models` for `enabledModels`, `providers` for `disabledProviders`, or `values` for either.\n\n## `/model` and `--list-models`\n\nBoth surfaces keep provider-prefixed models visible and selectable.\n\nThey now also expose canonical/coalesced models:\n\n- `/model` includes a canonical view alongside provider tabs\n- `--list-models` prints a canonical section plus the concrete provider rows\n\nSelecting a canonical entry stores the canonical selector. Selecting a provider row stores the explicit `provider/modelId`.\n\n## Context promotion (model-level fallback chains)\n\nContext promotion is an overflow recovery mechanism for small-context variants (for example `*-spark`) that automatically promotes to a larger-context sibling when the API rejects a request with a context length error. It is **off by default** (`contextPromotion.enabled` is `false`); opt in to enable it.\n\n### Trigger and order\n\nWhen a turn fails with a context overflow error (e.g. `context_length_exceeded`), `AgentSession` attempts promotion **before** falling back to compaction:\n\n1. If `contextPromotion.enabled` is true, resolve a promotion target (see below).\n2. If a target is found, switch to it and retry the request — no compaction needed.\n3. If no target is available, fall through to auto-compaction on the current model.\n\n### Target selection\n\nSelection is model-driven, not role-driven:\n\n1. `currentModel.contextPromotionTarget` (if configured)\n2. smallest larger-context model on the same provider + API\n\nCandidates are ignored unless credentials resolve (`ModelRegistry.getApiKey(...)`).\n\n### OpenAI code provider websocket handoff\n\nIf switching from/to `openai-codex-responses`, session provider state key `openai-codex-responses` is closed before model switch. This drops websocket transport state so the next turn starts clean on the promoted model.\n\n### Persistence behavior\n\nPromotion uses temporary switching (`setModelTemporary`):\n\n- recorded as a temporary `model_change` in session history\n- does not rewrite saved role mapping\n\n### Configuring explicit fallback chains\n\nConfigure fallback directly in model metadata via `contextPromotionTarget`.\n\n`contextPromotionTarget` accepts either:\n\n- `provider/model-id` (explicit)\n- `model-id` (resolved within current provider)\n\nExample (`models.yml`) for Spark -> non-Spark on the same provider:\n\n```yaml\nproviders:\n  openai-code:\n    modelOverrides:\n      gpt-5.3-openai-code-spark:\n        contextPromotionTarget: openai-code/gpt-5.3-openai-code\n```\n\nThe built-in model generator also assigns this automatically for `*-spark` models when a same-provider base model exists.\n\n## Compatibility and routing fields\n\nThe `compat` block on a provider or model overrides the URL-based auto-detection in `packages/ai/src/providers/openai-completions-compat.ts`. It is validated by `OpenAICompatSchema` in `packages/coding-agent/src/config/model-registry.ts` and consumed by every `openai-completions` transport (`packages/ai/src/providers/openai-completions.ts`). The canonical type is `OpenAICompat` in `packages/ai/src/types.ts`.\n\n`models.yml` accepts the following keys (all optional; unset falls back to URL detection):\n\nRequest shaping:\n\n- `supportsStore` — emit `store: false` on requests. Default: auto (off for non-standard endpoints).\n- `supportsDeveloperRole` — use the `developer` system role for reasoning models instead of `system`. Default: auto.\n- `sendSessionHeaders` — forward the agent session id as `session_id` and `x-session-id` request headers so OpenAI-compatible relays/proxies can do session-affinity routing and reuse a server-side prompt cache. Default: `false`. Caller-set `headers`/`requestTransform` values are never overwritten.\n- `supportsResponsesSessionAffinity` — for `openai-responses`, opt in to forwarding `session_id` and `x-client-request-id` affinity headers to a custom OpenAI-compatible relay. Canonical OpenAI routing remains automatic; known non-OpenAI provider IDs are rejected. Default: `false`.\n- `supportsUsageInStreaming` — send `stream_options: { include_usage: true }` to receive token usage on streaming responses. Default: `true`.\n- `maxTokensField` — `\"max_completion_tokens\"` or `\"max_tokens\"`. Default: auto.\n- `supportsToolChoice` — emit the `tool_choice` parameter when the caller forces a specific tool. Default: `true`. Set `false` for endpoints that 400 on `tool_choice` (e.g. DeepSeek when reasoning is on).\n- `disableReasoningOnForcedToolChoice` — drop `reasoning_effort` / OpenRouter `reasoning` whenever `tool_choice` forces a call. Default: auto (Kimi/Anthropic-fronted endpoints).\n- `extraBody` — extra top-level fields merged into every request body (gateway hints, controller selectors, etc.).\n\nReasoning / thinking:\n\n- `supportsReasoningEffort` — accept `reasoning_effort`. Default: auto (off for Grok and zAI).\n- `reasoningEffortMap` — partial map from internal effort levels (`minimal|low|medium|high|xhigh`) to provider-specific strings (e.g. DeepSeek maps `xhigh -> \"max\"`).\n- `thinkingFormat` — request shape for thinking: `\"openai\"` (`reasoning_effort`), `\"openrouter\"` (`reasoning: { effort }`), `\"zai\"` (`thinking: { type: \"enabled\" }`), `\"qwen\"` (top-level `enable_thinking`), or `\"qwen-chat-template\"` (`chat_template_kwargs.enable_thinking`). Default: `\"openai\"`.\n- `reasoningContentField` — assistant field carrying chain-of-thought: `\"reasoning_content\"`, `\"reasoning\"`, or `\"reasoning_text\"`. Default: auto.\n- `requiresReasoningContentForToolCalls` — assistant tool-call turns must round-trip the reasoning field (DeepSeek-R1, Kimi, OpenRouter when reasoning is on). Default: `false`.\n- `requiresAssistantContentForToolCalls` — assistant tool-call turns must include non-empty text content (Kimi). Default: `false`.\n\nTool / message normalization:\n\n- `requiresToolResultName` — tool-result messages need a `name` field (Mistral). Default: auto.\n- `requiresAssistantAfterToolResult` — a user message after a tool result needs an assistant turn in between. Default: auto.\n- `requiresThinkingAsText` — convert thinking blocks to text wrapped in `<thinking>` delimiters (Mistral). Default: auto.\n- `requiresMistralToolIds` — normalize tool-call ids to exactly 9 alphanumeric chars. Default: auto.\n- `supportsStrictMode` — accept the per-tool `strict` field on tool schemas. Default: conservative auto-detect per provider/baseUrl.\n- `toolStrictMode` — `\"all_strict\"` forces strict on every tool, `\"none\"` forces it off; unset keeps the existing per-tool mixed behavior.\n\nGateway routing (only applied when `baseUrl` matches the gateway):\n\n- `openRouterRouting.only` / `openRouterRouting.order` — provider routing on `openrouter.ai` (see <https://openrouter.ai/docs/provider-routing>).\n- `vercelGatewayRouting.only` / `vercelGatewayRouting.order` — provider routing on `ai-gateway.vercel.sh` (see <https://vercel.com/docs/ai-gateway/models-and-providers/provider-options>).\n\nProvider-level `compat` is the baseline; per-model `compat` is deep-merged on top, with `openRouterRouting`, `vercelGatewayRouting`, and `extraBody` merged as nested objects.\n\n### Anthropic compatibility (`anthropic-messages`)\n\nFor `anthropic-messages` models, `compat.promptCacheMode` and `compat.supportsLongCacheRetention` are configurable at provider, model, and `modelOverrides` levels. Provider-level `compat` is the baseline; model and override values merge on top.\n\nPrompt-cache modes:\n\n- `automatic` — emit one top-level `cache_control` marker and let the Anthropic-compatible endpoint advance the breakpoint as the conversation grows.\n- `explicit` — emit block-level breakpoints instead. Use this for endpoints that reject top-level `cache_control` but support Anthropic's explicit content-block markers.\n- `none` — emit no generated Anthropic cache controls. Per-request or configured `cacheRetention: none` also disables generated caching.\n\nWithout an explicit mode, canonical Anthropic endpoints default to `automatic`, Claude-family model ids on non-canonical compatible endpoints default to `explicit`, and unknown non-Claude compatible endpoints default to `none`. Non-canonical endpoints get the default ~5m lifetime unless they opt into `supportsLongCacheRetention: true`. Set `promptCacheMode: automatic` only when a gateway is known to pass through Anthropic's top-level cache control without adding conflicting block markers.\n\nIf a gateway attaches enough cache markers of its own that ours push the request past Anthropic's four-breakpoint limit, Anthropic rejects it with `A maximum of 4 blocks with cache_control may be provided.` Those extra markers are not visible in the request GJC builds, so the limit is handled at runtime rather than predicted. Because the rejection means \"too many\" rather than \"none allowed\", recovery reduces the generated breakpoints one step at a time: `explicit` mode normally emits two markers (a conversation-prefix anchor and a current-turn refresh point), so the first retry keeps only the prefix anchor, and generated caching is disabled entirely only if that is rejected too. The reduced setting persists for the rest of the provider session, so an endpoint with one free slot keeps caching its conversation prefix instead of losing caching altogether. Set `promptCacheMode: none` on a gateway that never has a free slot to skip the wasted attempts.\n\n```yaml\nproviders:\n  corp-anthropic:\n    baseUrl: https://proxy.example.com/anthropic\n    apiKeyEnv: CORP_ANTHROPIC_API_KEY\n    api: anthropic-messages\n    compat:\n      promptCacheMode: explicit\n      supportsLongCacheRetention: false\n    models:\n      - id: claude-sonnet-4-5\n        contextWindow: 200000\n        maxTokens: 8192\n```\n\nOther Anthropic-side compatibility knobs such as `disableAdaptiveThinking` and `supportsEagerToolInputStreaming` remain built-in catalog metadata rather than `models.yml` fields. `disableStrictTools` stays a provider-level setting (below).\n\n### Strict tool schemas (`disableStrictTools`)\n\nAnthropic's API supports a `strict` field on tool definitions that forces the model to always follow the provided schema exactly. This is enabled by default for all `anthropic-messages` providers because it guarantees schema conformance in agentic systems.\n\nThird-party providers that front the Anthropic API (AWS Bedrock, Azure, self-hosted proxies) do not always implement this field and will reject requests that include it. Set `disableStrictTools: true` at the provider level to opt out:\n\n```yaml\nproviders:\n  bedrock-anthropic:\n    baseUrl: https://bedrock-runtime.us-east-1.amazonaws.com/anthropic\n    apiKey: AWS_BEARER_TOKEN\n    api: anthropic-messages\n    disableStrictTools: true\n    models:\n      - id: anthropic-model-sonnet-4-20250514\n        name: Anthropic model Sonnet 4 (Bedrock)\n        input: [text, image]\n        contextWindow: 200000\n        maxTokens: 16384\n        cost:\n          input: 3.00\n          output: 15.00\n          cacheRead: 0.30\n          cacheWrite: 3.75\n```\n\n`disableStrictTools` is a provider-level flag that applies to all models in the provider.\n\nTool schemas going on the wire are normalized by the unified flow in\n`packages/ai/src/utils/schema/normalize.ts` (Google/CCA/MCP dispatchers\nplus the OpenAI strict-mode sanitize+enforce pipeline). See\n[`ai-schema-normalize.md`](./ai-schema-normalize.md) for the strict-mode\nedge cases (local `$ref` inlining, single-item `allOf` collapse,\n`anyOf`-wrapper description hoist, enum/const primitive-type inference)\nand the per-provider dispatcher mapping.\n## Practical examples\n\n### Local OpenAI-compatible endpoint (no auth)\n\n```yaml\nproviders:\n  local-openai:\n    baseUrl: http://127.0.0.1:8000/v1\n    auth: none\n    api: openai-completions\n    models:\n      - id: Qwen/Qwen2.5-Coder-32B-Instruct\n        name: Qwen 2.5 Coder 32B (local)\n```\n\n### Hosted proxy with env-based key\n\n```yaml\nproviders:\n  anthropic-proxy:\n    baseUrl: https://proxy.example.com/anthropic\n    apiKey: ANTHROPIC_PROXY_API_KEY\n    api: anthropic-messages\n    authHeader: true\n    disableStrictTools: true  # if the proxy doesn't support strict tool schemas\n    models:\n      - id: anthropic-model-sonnet-4-20250514\n        name: Anthropic model Sonnet 4 (Proxy)\n        reasoning: true\n        input: [text, image]\n```\n\n### Override built-in provider route + model metadata\n\n```yaml\nproviders:\n  openrouter:\n    baseUrl: https://my-proxy.example.com/v1\n    headers:\n      X-Team: platform\n    modelOverrides:\n      anthropic/anthropic-model-sonnet-4:\n        name: Sonnet 4 (Corp)\n        compat:\n          openRouterRouting:\n            only: [anthropic]\n```\n\n## Legacy consumer caveat\n\nMost model configuration now flows through `models.yml` via `ModelRegistry`. Explicit `.json` / `.jsonc` paths remain supported only when passed programmatically to `ModelRegistry`; the default user config is `~/.gjc/agent/models.yml`.\n\n## Failure mode\n\nIf `models.yml` fails schema or validation checks:\n\n- registry keeps operating with built-in models\n- error is exposed via `ModelRegistry.getError()` and surfaced in UI/notifications\n",
+	"models.md": "# Model and Provider Configuration (`models.yml`)\n\nThis document describes how the coding-agent currently loads models, applies overrides, resolves credentials, and chooses models at runtime.\n\n## What controls model behavior\n\nPrimary implementation files:\n\n- `src/config/model-registry.ts` — loads built-in + custom models, provider overrides, runtime discovery, auth integration\n- `src/config/model-resolver.ts` — parses model patterns and selects models for the default and agent roles\n- `src/config/settings-schema.ts` — model-related settings (`modelRoles`, provider transport preferences)\n- `src/session/auth-storage.ts` — API key + OAuth resolution order\n- `packages/ai/src/models.ts` and `packages/ai/src/types.ts` — built-in providers/models and `Model`/`compat` types\n\n## Config file location and legacy behavior\n\nDefault config path:\n\n- `~/.gjc/agent/models.yml`\n\nLegacy behavior still present:\n\n- If `models.yml` is missing and `models.json` exists at the same location, it is migrated to `models.yml`.\n- Explicit `.json` / `.jsonc` config paths are still supported when passed programmatically to `ModelRegistry`.\n\n## `models.yml` shape\n\n```yaml\nproviders:\n  <provider-id>:\n    # provider-level config\nequivalence:\n  overrides:\n    <provider-id>/<model-id>: <canonical-model-id>\n  exclude:\n    - <provider-id>/<model-id>\n```\n\n`provider-id` is the canonical provider key used across selection and auth lookup.\n\n`equivalence` is optional and configures canonical model grouping on top of concrete provider models:\n\n- `overrides` maps an exact concrete selector (`provider/modelId`) to an official upstream canonical id\n- `exclude` opts a concrete selector out of canonical grouping\n\n## Provider-level fields\n\n```yaml\nproviders:\n  my-provider:\n    baseUrl: https://api.example.com/v1\n    apiKey: MY_PROVIDER_API_KEY\n    api: openai-completions\n    headers:\n      X-Team: platform\n    authHeader: true\n    auth: apiKey\n    disableStrictTools: false  # set true for Anthropic-compatible endpoints that reject the strict field\n    cacheRetention: short  # none | short | long; model entries and modelOverrides can override this\n    discovery:\n      type: ollama\n    modelOverrides:\n      some-model-id:\n        name: Renamed model\n        cacheRetention: long\n    models:\n      - id: some-model-id\n        name: Some Model\n        api: openai-completions\n        reasoning: false\n        input: [text]\n        cost:\n          input: 0\n          output: 0\n          cacheRead: 0\n          cacheWrite: 0\n        contextWindow: 128000\n        maxTokens: 16384\n        headers:\n          X-Model: value\n        cacheRetention: none\n        thinking:\n          minLevel: low\n          maxLevel: xhigh\n          mode: effort\n          defaultLevel: high\n          levels: [low, medium, high, xhigh]\n        compat:\n          supportsStore: true\n          supportsDeveloperRole: true\n          supportsReasoningEffort: true\n          maxTokensField: max_completion_tokens\n          openRouterRouting:\n            only: [anthropic]\n          vercelGatewayRouting:\n            order: [anthropic, openai]\n          extraBody:\n            gateway: m1-01\n            controller: mlx\nmodelBindings:\n  modelRoles:\n    default: my-provider/some-model-id:high\n  agentModelOverrides:\n    executor: my-provider/some-model-id\n```\n\n### Allowed provider/model `api` values\n\n- `openai-completions`\n- `openai-responses`\n- `openai-codex-responses`\n- `azure-openai-responses`\n- `bedrock-converse-stream`\n- `anthropic-messages`\n- `google-generative-ai`\n- `google-vertex`\n- `google-gemini-cli`\n- `ollama-chat`\n- `cursor-agent`\n\n\n### First-class DeepInfra, Azure OpenAI, and Amazon Bedrock examples\n\nAzure OpenAI uses canonical OpenAI model IDs in GJC and resolves those IDs to Azure deployment names at request time. Set `AZURE_OPENAI_DEPLOYMENT_NAME_MAP` to avoid assuming model id equals deployment name:\n\n```yaml\nproviders:\n  azure-openai:\n    baseUrl: https://my-resource.openai.azure.com/openai/v1\n    apiKeyEnv: AZURE_OPENAI_API_KEY\n    api: azure-openai-responses\n    models:\n      - id: gpt-4.1\n      - id: o3\n```\n\n```sh\nexport AZURE_OPENAI_DEPLOYMENT_NAME_MAP='gpt-4.1=gpt-41-prod,o3=o3-reasoning-prod'\n```\n\nDeepInfra is available as the first-class `deepinfra` provider. It uses DeepInfra's OpenAI-compatible Chat Completions endpoint and reads `DEEPINFRA_API_KEY` when no explicit config key is provided. Set `serviceTier: priority` in GJC config or use the runtime service-tier controls to send DeepInfra's `service_tier: \"priority\"` request field for supported models:\n\n```yaml\nproviders:\n  deepinfra:\n    baseUrl: https://api.deepinfra.com/v1/openai\n    apiKeyEnv: DEEPINFRA_API_KEY\n    api: openai-completions\n    models:\n      - id: deepseek-ai/DeepSeek-V3.2\n```\n\n#### `/fast` provider support\n\n`/fast on` only shows `⚡` when GJC will put a fast/priority field on the selected provider's wire request:\n\n| Provider ID | Wire request | Notes |\n|---|---|---|\n| `openai` | `service_tier: \"priority\"` | OpenAI renamed Priority processing to [Fast mode](https://developers.openai.com/api/docs/guides/fast-mode); `priority` remains an accepted alias. For API-key requests, the response `service_tier` reports the tier actually used and may be `default` after a ramp-rate downgrade. |\n| `openai-codex` | `service_tier: \"priority\"` | ChatGPT-authenticated Codex handles Fast through server-side routing. A final response value of `service_tier: \"default\"` does not show that Fast was ignored or downgraded. |\n| `anthropic` | `speed: \"fast\"` plus `fast-mode-2026-02-01` beta | Direct Claude API only. Anthropic's [Fast mode](https://platform.claude.com/docs/en/build-with-claude/fast-mode) is model- and account-gated; unsupported or unavailable requests can fall back after a provider rejection. Bedrock, Vertex, and Microsoft Foundry do not support it. |\n| `deepinfra` | `service_tier: \"priority\"` | Sent only for the first-class `deepinfra` provider ID and only with the `priority` tier. |\n| `opencodex` | `service_tier: \"priority\"` | First-class OpenCodex discovery opts in automatically; OpenCodex Fast Mode must remain `Auto` for client passthrough. When OpenCodex uses ChatGPT authentication, a final `service_tier: \"default\"` is not downgrade evidence. |\n\nCustom OpenAI-compatible providers remain fail-closed unless their provider or model configuration explicitly sets `compat.supportsServiceTier: true`. Use that opt-in only when the proxy preserves or intentionally realizes OpenAI's `service_tier` contract:\n\n```yaml\nproviders:\n  my-openai-proxy:\n    baseUrl: http://proxy.example/v1\n    api: openai-responses\n    compat:\n      supportsServiceTier: true\n```\n\nWithout that capability, `/fast status` shows `off` even when the session retains an unscoped `priority` intent. The `⚡` indicator means that GJC sends the provider's fast request field. API-key providers may report a downgrade in their response; ChatGPT-authenticated Codex and OpenCodex route Fast server-side and cannot be verified from the final `service_tier` value.\n\nAmazon Bedrock uses the native `bedrock-converse-stream` transport and AWS credential chain auth. Do not put AWS access keys in `models.yml`; configure `AWS_REGION` / `AWS_PROFILE` or standard static AWS credential environment variables instead:\n\n```yaml\nproviders:\n  amazon-bedrock:\n    baseUrl: https://bedrock-runtime.us-east-1.amazonaws.com\n    api: bedrock-converse-stream\n    models:\n      - id: us.anthropic.claude-opus-4-6-v1\n      - id: anthropic.claude-3-5-sonnet-20241022-v2:0\n```\n\n### Coding-plan provider presets\n\nFor supported coding-plan providers, prefer presets so the API type, base URL, environment variable, model catalog, discovery behavior, and compatibility flags are written together:\n\n```sh\ngjc setup provider --preset minimax\ngjc setup provider --preset minimax-cn\ngjc setup provider --preset glm\ngjc setup provider --preset alibaba-token-plan\ngjc setup provider --preset cline-pass\ngjc setup provider --preset commandcode-goat\n```\n\nThe same presets are available inside the TUI:\n\n```text\n/provider add --preset minimax\n/provider add --preset glm\n/provider add zai\n/provider add --preset alibaba-token-plan\n/provider add --preset cline-pass\n/provider add --preset commandcode-goat\n```\n\nPresets only write `models.yml` entries that reference documented environment variable names (`MINIMAX_CODE_API_KEY`, `MINIMAX_CODE_CN_API_KEY`, `ZAI_API_KEY`, `ALIBABA_TOKEN_PLAN_API_KEY`, `CLINE_API_KEY`, or `CMD_API_KEY`); they do not store or validate real credentials. The GLM preset aliases (`glm`, `zai`, `z-ai`) write an OpenAI-compatible custom provider named `glm-proxy` and do not replace the first-class `zai` provider. The Alibaba Token Plan preset (aliases: `alibaba`, `token-plan`) writes an OpenAI-compatible custom provider named `alibaba-token-plan` with per-model API routing. The ClinePass preset (aliases: `clinepass`, `cline`) does not hardcode models: Cline's inference API has no working `/models` route, so GJC follows Cline's own catalog-generation source and fetches the live `cline-pass` provider catalog from `https://models.dev/api.json`. The Command Code GOAT preset (aliases: `commandcode`, `command-code`, `goat`) fetches its live `/provider/v1/models` catalog, routes every current or future `claude-*` model through Anthropic Messages, and routes other models through Chat Completions. Create the corresponding API key in the provider dashboard before inference; plan entitlement is enforced by the provider.\n\n## Model profiles (`--mpreset`)\n\nModel profiles are optional top-level `profiles:` entries in `~/.gjc/agent/models.yml`. A profile can require provider credentials before activation and can map one or more model roles; omitted roles inherit from the active defaults.\n\n> See also: [Cross-vendor role-based profiles](./multi-vendor-profiles.md) — a curated multi-vendor `profiles:` recipe and verified selector notes that build on the mechanism described here.\n\n```yaml\nprofiles:\n  team-standard:\n    required_providers: [openai, anthropic]\n    model_mapping:\n      default: openai/gpt-5.2\n      executor: anthropic/claude-sonnet-5:medium\n      architect: openai/o3:high\n      planner: openai/o3:high\n      critic: openai/o3:high\n```\n\n`model_mapping` keys are role names (`default`, `executor`, `architect`, `planner`, `critic`). Every role accepts either one selector or a non-empty ordered array of selectors; the first entry is primary and later entries are fallback candidates. A selector may be a provider-agnostic bare alias such as `glm-5.2[:effort]` or an explicit `provider/modelId[:effort]` pin, including nested model IDs such as `openrouter/anthropic/claude-sonnet-5`. `required_providers` lists explicit provider prerequisites and may be empty when availability is resolved from bare aliases.\n\n### Fallback chains\n\nPreset `model_mapping` roles, top-level `modelRoles`, and `task.agentModelOverrides` all accept `string | string[]`. Keep one selector per line when a chain needs to be readable:\n\n```yaml\nprofiles:\n  reliable:\n    required_providers: [anthropic, openai]\n    model_mapping:\n      default: [anthropic/claude-sonnet-4-5, openai/gpt-4o-mini]\nmodelBindings:\n  modelRoles:\n    default: [anthropic/claude-sonnet-4-5, openai/gpt-4o-mini]\n  agentModelOverrides:\n    executor: [anthropic/claude-sonnet-4-5, openai/gpt-4o-mini]\n```\n\nResolution-time skips for unavailable, unauthenticated, or unknown entries cost zero attempts and advance immediately. Only request-time retryable failures (such as 429, quota, authentication, or 5xx failures) consume an entry's `fallback.maxAttempts` total attempts (default: `3`). The active default fallback remains sticky for the session; role-override fallback state is fresh for each subagent call. The active model is shown consistently in status and `/model`.\n\nManaged fallback attempts buffer provisional streamed output until an attempt is accepted, so output can appear later than it does for a one-model stream. Current Cursor-agent transports are fail-closed unavailable in retryable fallback chains: resolution rejects them with `Cursor model <selector> requires provider-side tool execution and cannot be used in a retryable fallback chain` because they do not provide a client-side tool-call mode.\n\nCancellation discards provisional output and emits exactly one cancelled `agent_end`; RPC, ACP, and the TUI therefore settle once. On load, the source-aware one-shot migration reads legacy `retry.fallbackChains`, prepends the effective role chain, and writes the ordered, deduplicated result to the corresponding role array; the legacy key is then ignored.\n\nBuilt-in profiles are grouped by provider mix and tier:\n\n- `codex-{eco,medium,pro}` — GPT-5.6 Sol/Terra/Luna role mixes tuned by tier and reasoning effort; `lunamaxxing` — OpenAI Codex Luna-only profile with maximum reasoning on delegated roles\n- `opencodego` — single OpenCode Go preset (Kimi K3 default and planner, DeepSeek executor/architect, MiMo critic)\n- Provider-agnostic open-model profiles are named by the model families they require. Single-family choices are `open-weights-{glm,deepseek,kimi,luna}`; two-family choices are `open-weights-glm-deepseek`, `open-weights-kimi-deepseek`, and `open-weights-kimi-glm`; `open-weights-kimi-glm-deepseek` uses all three open-weight families; `open-weights-all` adds GPT-5.6 Luna. Choose the smallest combination covered by the models available through your configured providers. Every selector is a bare final-segment alias with `required_providers: []`, so each family may come from any authenticated bundled or custom provider under Provider Priority. GPT-5.6 Luna is proprietary despite its inclusion in this group.\n- `claude-opus` — Anthropic OAuth preset centered on `claude-opus-5`\n- Single-provider tiers: `glm-{eco,medium,pro}`, `kimi-coding-plan-{eco,medium,pro}`, `mimo-{eco,medium,pro}`, `grok-{eco,medium,pro}`, `grok-45-{eco,medium,pro}`, `grok-46-{eco,medium,pro}`, `cursor-{eco,medium,pro}`, `minimax-{eco,medium,pro}`\n- Alibaba Token Plan: `alibaba-token-plan-balanced` preserves the established Qwen/DeepSeek V4 Pro/GLM mix; `alibaba-token-plan-pro` raises execution and independent criticism with DeepSeek V4 Flash 0731 max and GLM xhigh; `alibaba-token-plan-qwenmaxxing` stays Qwen-only; `alibaba-token-plan-qwen-deepseek` keeps Qwen 3.8 Max (`qwen3.8-max`) on the expensive default (high)/architect (xhigh)/critic (xhigh) roles and spends DeepSeek V4 Flash 0731 on the cheap planner (max) and executor (high) roles; `alibaba-token-plan-glm-deepseek` does the same with GLM 5.2 (`glm-5.2`) as the expensive model\n- Combos: `opus-codex`, `codex-opencodego`, and `fable-opus-codex`\n\nThe `eco`, `medium`, and `pro` Codex profile mappings are current product judgments: Eco assigns Terra low/Luna low/Luna high/Terra xhigh/Terra high to default/executor/planner/critic/architect; Medium assigns Sol low/Terra low/Terra high/Sol xhigh/Sol high; Pro assigns Sol medium/Terra medium/Sol high/Sol max/Sol xhigh; and LunaMaxxing assigns Luna medium/Luna xhigh/Luna max/Luna max/Luna max. `opus-codex` retains the Medium Codex executor, critic, and architect roles but uses `anthropic/claude-sonnet-5` for planner; `codex-opencodego` retains the Medium Codex default and architect roles; and `fable-opus-codex` uses the Pro Codex executor and architect roles with `anthropic/claude-opus-5:medium` for planner. The descriptive repeated local exact-edit evidence informs only selected executor-style TypeScript tasks; it does not evaluate or prove default, planner, architect, or critic performance. See [GPT-5.6 Codex preset benchmark](./gpt-5.6-codex-preset-benchmark.md). The Alibaba Pro role evidence and its limits are recorded separately in [Alibaba Token Plan Pro profile benchmark](./alibaba-token-plan-pro-profile-benchmark.md). Cursor Eco uses Composer 2.5 for every role; Medium keeps standard Composer for default/planning and spends the Fast premium on execution, criticism, and architecture; Pro uses Composer 2.5 Fast throughout. Composer does not expose a strength value through the current Cursor RPC, so these profiles use exact model IDs without inert generic effort suffixes. See [Cursor Composer profile tiers](./cursor-composer-profile-tiers.md). Effort suffixes are clamped to each model's supported thinking range at preview and activation time. Single-provider tiers pin each provider's current flagship (`zai/glm-5.2`, `kimi-code/kimi-k2.7-code`, `xiaomi/mimo-v2.5-pro`, `xai/grok-4.3`, `cursor/composer-2.5`, `minimax-code/MiniMax-M3`). User-defined profiles override built-ins by exact profile name.\n\n\nUse `gjc --mpreset <name>` to activate a profile for the current session only. Activation hard-blocks when any provider listed in `required_providers` lacks credentials. Add `--default` to persist the selected profile as `modelProfile.default` in `config.yml`, so it applies at startup:\n\n```sh\ngjc --mpreset codex-medium\ngjc --mpreset opencodego --default\n```\n\n### Routing built-in presets through a proxy (`modelProfile.proxyProvider`)\n\nBuilt-in preset selectors pin a direct provider endpoint (`xai/grok-4.3`, `xiaomi/mimo-v2.5-pro`, …). To serve those models through your own OpenAI-compatible gateway (LiteLLM, OpenRouter, or a custom proxy) instead of each vendor's endpoint, configure the proxy provider id and routing mode in `config.yml`:\n\n```yaml\nmodelProfile:\n  proxyProvider: litellm\n  proxyMode: always # use fallback to keep directly authenticated providers direct\n```\n\nThe proxy provider is a normal `providers:` entry. Add it with `gjc setup provider --preset litellm --base-url <url>` or the generic `gjc setup provider --preset openai-compatible-proxy --base-url <url>` (both presets require `--base-url` and use live model discovery). The configured proxy must be authenticated and expose every routed model. Activation rewrites each selected built-in preset selector from `<direct-provider>/<model>` to `<proxy>/<direct-provider>/<model>` (for example `xai/grok-4.3` → `litellm/xai/grok-4.3`), matching the proxy's catalog entry for the model. The rules:\n\n- Routing applies to **built-in presets only**. User-defined `profiles:` entries always keep their exact selectors — set them explicitly if you want them proxied.\n- `proxyMode: fallback` (the default) routes only selectors whose direct provider is unauthenticated. `proxyMode: always` routes every proxy-routable built-in selector through the configured proxy, including selectors with direct credentials.\n- The proxy id must name a configured provider. `proxyMode: always` requires `proxyProvider` and a usable proxy credential; activation fails closed when a required proxy is unset or unauthenticated. `auth: none` proxies count as authenticated.\n- Only providers the bundled preset catalog treats as routable are rewritten; providers outside that set (for example a custom `acme-private`) keep the direct credential error.\n- A routed selector must have exactly one matching proxy catalog model. Exact `<direct-provider>/<model>` proxy ids win over suffix matches; missing or ambiguous matches fail activation before any role can run.\n\nThe `/model` command opens to a preset landing view: presets are grouped by provider with live auth marks (✓/✗), highlighting a group expands its tiers, and selecting a tier shows the full role→model preview before applying for the session or as default. Typing jumps straight to model search, and `Browse all models` opens the classic tabbed model selector. In `/login`, `Add custom provider` is the first option for configuring credentials needed by custom or profile-required providers; after a successful provider login, the matching preset is recommended automatically. Custom providers participate in provider-agnostic alias resolution but require manual preset selection.\n\nExternal SDK/ACP clients (e.g. the Paseo TUI) can select profiles like ordinary models: the SDK `models.list/current` (Q10) catalog exposes every usable profile as a synthetic `gajae-code/<profile>` entry (e.g. `gajae-code/codex-eco`), and selecting one through `model.set` (or the ACP Model picker) activates the profile for the live session only. Persisting a profile remains an explicit TUI choice, mirroring `gjc --mpreset <name> --default`. See [SDK model profiles](./sdk.md#model-profiles-as-synthetic-models-gajae-codeprofile).\n\nMiniMax's OpenAI-compatible endpoint rejects multiple system messages and emits thinking in `reasoning_content`, so pin the public-safe compatibility fields when hand-authoring a custom provider:\n\n```yaml\nproviders:\n  minimax-custom:\n    baseUrl: https://api.minimax.io/v1\n    apiKeyEnv: MINIMAX_API_KEY\n    api: openai-completions\n    compat:\n      supportsStore: false\n      supportsDeveloperRole: false\n      supportsReasoningEffort: false\n      reasoningContentField: reasoning_content\n    models:\n      - id: MiniMax-M2.5\n```\n\nGLM via z.ai is available as the first-class `zai` provider. For a private GLM-compatible proxy, keep secrets in an env var and disable OpenAI-only request fields as needed:\n\n```yaml\nproviders:\n  glm-proxy:\n    baseUrl: https://api.z.ai/api/paas/v4\n    apiKeyEnv: ZAI_API_KEY\n    api: openai-completions\n    compat:\n      supportsDeveloperRole: false\n      supportsReasoningEffort: false\n    models:\n      - id: glm-4.6\n```\n\n### JetBrains AI (Junie)\n\n`jetbrains-junie` is a first-class provider serving JetBrains-hosted models through the documented\nIngrazzio gateway (`https://ingrazzio-cloud-prod.labs.jb.gg`).\n\nAuthenticate with an access token generated at [junie.jetbrains.com/cli](https://junie.jetbrains.com/cli):\n\n```sh\nexport JUNIE_API_KEY=...\n```\n\nThe token is sent as `Authorization: Bearer` — JetBrains AI rejects requests that also carry `x-api-key`, so\nthis provider never lets the Anthropic SDK attach one. Usage is billed against your JetBrains AI\nsubscription, so bundled per-token costs are zero. There is no OAuth login flow; the environment variable is\nthe only supported credential source.\n\nThe gateway multiplexes transports by model family:\n\n| Family | Models | Transport | Prompt limit |\n| --- | --- | --- | --- |\n| Claude | `claude-sonnet-4-6` (default), `claude-sonnet-5`, `claude-opus-4-6`, `claude-opus-4-7`, `claude-opus-4-8`, `claude-opus-5`, `claude-fable-5` | `anthropic-messages` | 1M |\n| GPT | `gpt-5-2025-08-07`, `gpt-5.2-2025-12-11`, `gpt-5.4`, `gpt-5.5`, `gpt-5.6-luna`, `gpt-5.6-sol`, `gpt-5.6-terra` | `openai-completions` | 922K |\n| GPT (Responses-only) | `gpt-5.3-codex` | `openai-responses` | 272K |\n\nAll models cap output at 128K. Junie also exposes Gemini and Grok, but those ride a proprietary Grazie\ntranslation protocol that GJC does not implement, so they are deliberately not bundled. The bare\n`opus`/`sonnet`/`gpt`/`grok` aliases are Junie CLI shorthands the gateway itself rejects.\n\n### Allowed auth/discovery values\n\n- `auth`: `apiKey` (default), `none`, or `oauth`; for `models.yml` custom models, `oauth` is accepted by schema but does not waive the `apiKey` requirement\n- `models.yml` is strict: unknown provider/model keys fail validation before provider dispatch, so stale keys such as `requestTransform` or `wireModelId` only work where this document lists them.\n- `discovery.type`: `ollama`, `llama.cpp`, `lm-studio`, or `openai-models-list`\n- `cacheRetention`: `none`, `short`, or `long`; request-time options win over model/modelOverride values, then provider values, then `GJC_CACHE_RETENTION`, then the runtime default. The runtime default is `short` for most providers, but the Anthropic provider defaults to `long` because the ~5m cache is fragile for long-running subagent workflows. Canonical Anthropic models use top-level automatic caching and emit `ttl: \"1h\"` when long retention is supported. Claude-family models on non-canonical Anthropic-compatible endpoints default to explicit block markers because compatible proxies commonly inject, rewrite, or reject top-level cache controls; they omit `ttl` unless `compat.supportsLongCacheRetention: true` opts the endpoint into 1-hour retention. For OpenAI Responses, this controls `prompt_cache_retention` only; it does not disable `prompt_cache_key` when a stable session id exists.\n\n## OpenAI-compatible proxy configuration\n\nOpenAI-compatible proxy providers should use schema-supported provider keys first:\nThe first-class way to add a proxy provider is `gjc setup provider --preset litellm --base-url <url>` (LiteLLM) or `gjc setup provider --preset openai-compatible-proxy --base-url <url>` (any OpenAI-compatible gateway); both presets require `--base-url` and configure live model discovery. Proxy providers can also be used to route built-in model-preset selectors — see [Routing built-in presets through a proxy](#routing-built-in-presets-through-a-proxy-modelprofileproxyprovider). The YAML below shows the equivalent hand-written provider config:\n\n```yaml\nproviders:\n  proxy-provider:\n    baseUrl: https://api.proxy.example/v1\n    apiKeyEnv: PROXY_API_KEY\n    api: openai-completions\n    auth: apiKey\n    headers:\n      User-Agent: curl/8.7.1\n    models:\n      - id: local-gpt\n        name: Local GPT\n        reasoning: true\n        input: [text]\n        cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 }\n        contextWindow: 400000\n        maxTokens: 128000\n```\n\nUse provider-level `headers` for proxy-required headers. Keep the provider `api` set to `openai-completions` when the proxy exposes Chat Completions-compatible `/v1/chat/completions` semantics. `auth: apiKey` sends the resolved token as bearer auth; use `auth: none` only for trusted local/no-auth endpoints.\n\n`auth` selects the transport scheme only; it never supplies a credential. A provider that declares `models:` must therefore also declare where its key comes from, and `models.yml` validation rejects the config before model discovery otherwise:\n\n| Intent | Required keys |\n| --- | --- |\n| Authenticated proxy (recommended) | `auth: apiKey` (default) + `apiKeyEnv: MY_TOKEN` |\n| Authenticated proxy, key inline | `auth: apiKey` (default) + `apiKey: sk-…` (less safe; stored in plaintext) |\n| Genuinely unauthenticated endpoint | `auth: none`, no key |\n\nOmitting both `apiKey` and `apiKeyEnv` while leaving `auth` at its `apiKey` default fails with `Provider <name>: custom models need a credential source, but none is configured.` — the fix is to add one of the rows above, not to change `api` or `baseUrl`.\n\n`input` is the model modality list GJC uses to decide whether image content is forwarded. When a custom model omits `input`, GJC defaults to `[text]` (unless a bundled model with the same id contributes a reference). Vision-capable upstream models therefore need an explicit `input: [text, image]`; otherwise `read`/tool images are stripped before the request and replaced with `[image omitted: model does not support vision]`, even if the remote model can see images.\n\n```yaml\nproviders:\n  ali:\n    baseUrl: https://token-plan.ap-southeast-1.maas.aliyuncs.com/compatible-mode/v1\n    apiKeyEnv: ALI_API_KEY\n    api: openai-completions\n    auth: apiKey\n    models:\n      # id-only → text-only; images will be omitted\n      - id: some-text-model\n      # vision-capable hosted model must declare image input\n      - id: qwen3.8-max-preview\n        name: Qwen3.8 Max Preview\n        reasoning: true\n        input: [text, image]\n```\n\n`requestTransform` and `wireModelId` remain supported for request-body shaping, but they are not needed for ordinary OpenAI-compatible proxies whose local model id is already the upstream wire id. Unknown config keys fail validation before a provider request is sent.\n\nWhen request shaping is needed:\n\n- `requestTransform.profile: openai-proxy` strips OpenAI SDK/Stainless telemetry and beta headers at final fetch time and sets a generic GJC user agent.\n- `stripHeaders` replaces the preset strip list when provided.\n- `setHeaders` is applied after stripping; use `null` to remove a header.\n- `extraBody` is shallow-merged into the JSON request body after provider compatibility fields; core transport keys such as `model`, `messages`/`input`, `stream`, `tools`, and `tool_choice` are protected and ignored.\n- Model-level `requestTransform` overrides provider-level fields and shallow-merges `setHeaders`/`extraBody`.\n- `wireModelId` changes only the upstream request body model id; local selection still uses `provider/id`.\n\n### Layofflabs-style proxy example\n\n```yaml\nproviders:\n  layofflabs:\n    baseUrl: https://api.layofflabs.com/v1\n    apiKeyEnv: OPENAI_API_KEY\n    api: openai-completions\n    auth: apiKey\n    headers:\n      User-Agent: curl/8.7.1\n    models:\n      - id: gpt-5.5\n        name: GPT 5.5 via Layofflabs\n        reasoning: true\n        thinking:\n          minLevel: low\n          maxLevel: xhigh\n          mode: effort\n          defaultLevel: high\n          levels: [low, medium, high, xhigh]\n        input: [text]\n        cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 }\n        contextWindow: 400000\n        maxTokens: 128000\n\nmodelBindings:\n  modelRoles:\n    default: layofflabs/gpt-5.5:high\n  agentModelOverrides:\n    executor: layofflabs/gpt-5.5:high\n```\n\n## Validation rules (current)\n\n### Full custom provider (`models` is non-empty)\n\nRequired:\n\n- `baseUrl`\n- A credential source: `apiKeyEnv` or `apiKey`. `auth` selects the scheme, not the credential, so `auth: apiKey` (the default) still needs one of them. Exempt: `auth: none`, and `api: bedrock-converse-stream`, which resolves AWS credentials from its own chain.\n- `api` at provider level or each model\n\n### Override-only provider (`models` missing or empty)\n\nMust define at least one of:\n\n- `baseUrl`\n- `headers`\n- `compat`\n- `requestTransform`\n- `disableStrictTools`\n- `modelOverrides`\n- `discovery`\n\n### Discovery\n\n- `discovery` requires provider-level `api`.\n\n### Model value checks\n\n- `id` required\n- `contextWindow` and `maxTokens` must be positive if provided\n- unknown provider, model, override, and request-transform keys fail schema validation; remove stale keys instead of relying on them being ignored.\n\n## Merge and override order\n\nModelRegistry pipeline (on refresh):\n\n1. Load built-in providers/models from `@gajae-code/ai`.\n2. Load `models.yml` custom config.\n3. Apply provider overrides (`baseUrl`, `headers`, `requestTransform`, `disableStrictTools`, `cacheRetention`) to built-in models.\n4. Apply `modelOverrides` (per provider + model id).\n5. Merge custom `models`:\n   - same `provider + id` replaces existing\n   - otherwise append\n6. Load cached/runtime-discovered models (Ollama, llama.cpp, LM Studio, plus built-in provider managers), then re-apply model overrides.\n\n### Provider-model cache and static fingerprint\n\nCached per-provider model lists are persisted in the model-cache SQLite\ndatabase (schema v3) with a `static_fingerprint` column that hashes the\nstatic catalog slice merged into the row. When `resolveProviderModels`\nskips the network fetch and the fingerprint of the in-memory static\ncatalog matches the cached one, the cached rows are returned verbatim —\nthe static + dynamic merge is bypassed entirely. The fingerprint is\nmemoized per process via a WeakMap keyed by the static-models array\nreference, so repeated cold-start calls do not re-hash.\n\n## Canonical model equivalence and coalescing\n\nThe registry keeps every concrete provider model and then builds a canonical layer above them.\n\nCanonical ids are official upstream ids only, for example:\n\n- `anthropic-model-opus-4-6`\n- `anthropic-model-haiku-4-5`\n- `gpt-5.3-openai-code`\n\n### `models.yml` equivalence config\n\nExample:\n\n```yaml\nproviders:\n  zenmux:\n    baseUrl: https://api.zenmux.example/v1\n    apiKey: ZENMUX_API_KEY\n    api: openai-codex-responses\n    models:\n      - id: openai-code\n        name: Zenmux OpenAI code\n        reasoning: true\n        input: [text]\n        cost:\n          input: 0\n          output: 0\n          cacheRead: 0\n          cacheWrite: 0\n        contextWindow: 200000\n        maxTokens: 32768\n\nequivalence:\n  overrides:\n    zenmux/openai-code: gpt-5.3-openai-code\n    p-openai-code/openai-code: gpt-5.3-openai-code\n  exclude:\n    - demo/openai-code-preview\n```\n\nBuild order for canonical grouping:\n\n1. exact user override from `equivalence.overrides`\n2. bundled official-id matches from built-in model metadata\n3. conservative heuristic normalization for gateway/provider variants\n4. fallback to the concrete model's own id\n\nCurrent heuristics are intentionally narrow:\n\n- embedded upstream prefixes can be stripped when present, for example `anthropic/...` or `openai/...`\n- dotted and dashed version variants can normalize only when they map to an existing official id, for example `4.6 -> 4-6`\n- ambiguous families or versions are not merged without a bundled match or explicit override\n\n### Canonical and preset-equivalent resolution behavior\n\nWhen multiple concrete variants are eligible for automatic resolution, the global provider policy uses this order:\n\n1. explicit `config.yml` `modelProviderOrder` entries, in their saved order\n2. omitted providers whose effective credential came from OAuth\n3. omitted providers using a manual API key, unknown credential provenance, or keyless access\n4. vision capability, exact canonical identity, canonical source quality, lowest `cost.input + cost.cacheRead`, stable registry model order, then concrete selector order\n\nThe explicit provider list may be partial. A listed API-key provider beats every omitted OAuth provider. Resetting Provider Priority clears the explicit list and restores OAuth-first plus deterministic fallback. Saved providers that are not currently available remain visible and persisted, but runtime resolution skips them.\n\nModel-profile and preset assignments support a lookup-only alias when the assignment does not explicitly name a provider. For example, a bare preset assignment `gpt-5` can select any available concrete variant whose final model-id segment is `gpt-5`, ranked by the same global policy. An assignment such as `openai/gpt-5` is an explicit provider pin: if that exact model is unavailable, activation reports it unavailable instead of switching providers. Alias lookup never rewrites the selected model's concrete provider, full model id, or `wireModelId`; a known bare alias with no eligible variant is unavailable and does not fall through to a different fuzzy match. Direct model selection remains unchanged.\nRuntime custom providers participate in the same lookup automatically. A custom provider model such as `hosted/glm-5.2` contributes the alias `glm-5.2` when the provider is registered, authenticated, and available; Provider Priority can rank that custom provider ahead of bundled providers without changing the preset. Custom IDs whose final segment differs (for example `glm-5.2-special`) do not join the alias.\n\nA session that resolves a canonical or alias selector keeps its concrete variant across provider-priority edits and discovery refreshes. Updated priority applies to new or unpinned resolutions. The session re-ranks only after explicit reselection or when the sticky variant becomes unavailable. Session state and transcripts continue to record the concrete provider/model that executed the turn.\n\nProvider defaults vs per-model overrides:\n\n- Provider `headers` are baseline.\n- Model `headers` override provider header keys.\n- `modelOverrides` can override model metadata (`name`, `reasoning`, `input`, `cost`, `contextWindow`, `maxTokens`, `headers`, `compat`, `contextPromotionTarget`).\n- `compat` is deep-merged for nested routing blocks (`openRouterRouting`, `vercelGatewayRouting`, `extraBody`).\n\n## Runtime discovery integration\n\n### Implicit Ollama discovery\n\nIf `ollama` is not explicitly configured, registry adds an implicit discoverable provider:\n\n- provider: `ollama`\n- api: `openai-responses`\n- base URL: `OLLAMA_BASE_URL` or `http://127.0.0.1:11434`\n- auth mode: keyless (`auth: none` behavior)\n\nRuntime discovery calls Ollama endpoints and normalizes discovered OpenAI-compatible models to `openai-responses`.\n\n### Implicit llama.cpp discovery\n\nIf `llama.cpp` is not explicitly configured, registry adds an implicit discoverable provider:\n\n- provider: `llama.cpp`\n- api: `openai-responses`\n- base URL: `LLAMA_CPP_BASE_URL` or `http://127.0.0.1:8080`\n- auth mode: keyless (`auth: none` behavior)\n\nRuntime discovery calls llama.cpp model endpoints and synthesizes model entries with local defaults.\n\n### Implicit LM Studio discovery\n\nIf `lm-studio` is not explicitly configured, registry adds an implicit discoverable provider:\n\n- provider: `lm-studio`\n- api: `openai-completions`\n- base URL: `LM_STUDIO_BASE_URL` or `http://127.0.0.1:1234/v1`\n- auth mode: keyless (`auth: none` behavior)\n\nRuntime discovery fetches models (`GET /models`) and synthesizes model entries with local defaults.\n\n### Explicit provider discovery\n\nYou can configure discovery yourself:\n\n```yaml\nproviders:\n  ollama:\n    baseUrl: http://127.0.0.1:11434\n    api: openai-responses\n    auth: none\n    discovery:\n      type: ollama\n\n  llama.cpp:\n    baseUrl: http://127.0.0.1:8080\n    api: openai-responses\n    auth: none\n    discovery:\n      type: llama.cpp\n```\n\n### Extension provider registration\n\nExtensions can register providers at runtime (`pi.registerProvider(...)`), including:\n\n- model replacement/append for a provider\n- custom stream handler registration for new API IDs\n- custom OAuth provider registration\n\n## Auth and API key resolution order\n\nWhen requesting a key for a provider, effective order is:\n\n1. Runtime override (CLI `--api-key`)\n2. Stored API key credential in `agent.db`\n3. Stored OAuth credential in `agent.db` (with refresh)\n4. Environment variable mapping (`OPENAI_API_KEY`, `ANTHROPIC_API_KEY`, etc.)\n5. ModelRegistry fallback resolver (provider `apiKey` from `models.yml`, env-name-or-literal semantics)\n\n`models.yml` `apiKey` behavior:\n\n- Value is first treated as an environment variable name.\n- If no env var exists, the literal string is used as the token.\n\nIf `authHeader: true` and provider `apiKey` is set, models get:\n\n- `Authorization: Bearer <resolved-key>` header injected.\n\nKeyless providers:\n\n- Providers marked `auth: none` are treated as available without credentials.\n- `getApiKey*` returns `kNoAuth` for them.\n\n### Broker mode\n\nWhen `GJC_AUTH_BROKER_URL` (or `auth.broker.url`) is set, the local SQLite credential store is replaced by `RemoteAuthCredentialStore`. Layers 2 and 3 above (stored API key / OAuth in `agent.db`) are served from a broker-supplied snapshot whose `refresh` tokens are redacted; expiry triggers `POST /v1/credential/:id/refresh` on the broker rather than a local refresh.\n\n`AuthStorage.setConfigApiKey` lets a `models.yml` `apiKey` win over a broker-resolved OAuth token without overriding a runtime `--api-key`. See [`auth-broker-gateway.md`](./auth-broker-gateway.md) for the full broker / gateway design and env surface (`GJC_AUTH_BROKER_URL`, `GJC_AUTH_BROKER_TOKEN`, `auth.broker.url`, `auth.broker.token`).\n\n## Model availability vs all models\n\n- `getAll()` returns the loaded model registry (built-in + merged custom + discovered).\n- `getAvailable()` filters to models that are keyless or have resolvable auth.\n\nSo a model can exist in registry but not be selectable until auth is available.\n\n## Runtime model resolution\n\n### CLI and pattern parsing\n\n`model-resolver.ts` supports:\n\n- exact `provider/modelId`\n- exact canonical model id\n- exact model id (provider inferred)\n- fuzzy/substring matching\n- glob scope patterns in `--models` (e.g. `openai/*`, `*sonnet*`)\n- optional `:thinkingLevel` suffix (`off|minimal|low|medium|high|xhigh`)\n\n`--provider` is legacy; `--model` is preferred.\n\nResolution precedence for exact selectors:\n\n1. exact `provider/modelId` bypasses coalescing\n2. exact canonical id resolves through the canonical index\n3. exact bare concrete id still works\n4. fuzzy and glob matching run after the exact paths\n\nThinking suffixes are split once from the final `:` only after the complete selector does not resolve. This preserves concrete OpenRouter route IDs such as `openrouter/z-ai/glm-4.7:nitro`; `:high` can follow that route suffix. Multiple suffixes are not recursively consumed. A complete `provider/modelId` selector is exact-only: it never falls back to fuzzy, substring, glob, or another provider when that concrete selector is absent. Exact-case provider/model entries resolve deterministically for custom replacement semantics; a case-insensitive selector that remains ambiguous does not guess.\n\nPreset/profile activation may use an unqualified assignment as a final-segment lookup alias after exact resolution fails. Provider-qualified assignments remain exact pins, and this opt-in does not apply to CLI or direct concrete model selection.\n\n### Initial model selection priority\n\n`findInitialModel(...)` uses this order:\n\n1. explicit CLI provider+model\n2. first scoped model (if not resuming)\n3. saved default provider/model\n4. known provider defaults (e.g. OpenAI/Anthropic/etc.) among available models\n5. first available model\n\n### Role aliases and settings\n\nSupported model roles:\n\n- `default` plus the agent assignment targets `executor`, `architect`, `planner`, `critic`\n\nRole aliases like `pi/default` expand through `settings.modelRoles`. Each role value can also append a thinking selector such as `:minimal`, `:low`, `:medium`, or `:high`.\n\nIf a role points at another role, the target model still inherits normally and any explicit suffix on the referring role wins for that role-specific use.\n\nRelated settings:\n\n- `modelRoles` (record)\n- `enabledModels` (scoped pattern list)\n- `modelProviderOrder` (global automatic provider precedence; editable under Settings → Providers with add/remove, move up/down, unavailable-entry retention, and Reset)\n- `providers.kimiApiFormat` (`openai` or `anthropic` request format)\n- `providers.openaiWebsockets` (`auto|off|on` websocket preference for OpenAI code provider transport)\n\n`modelRoles` may store either:\n\n- `provider/modelId` to pin a concrete provider variant\n- a canonical id such as `gpt-5.3-openai-code` to allow provider coalescing\n\nFor `enabledModels` and CLI `--models`:\n\n- exact canonical ids expand to all concrete variants in that canonical group\n- explicit `provider/modelId` entries stay exact\n- globs and fuzzy matches still operate on concrete models\n\nGlobal `enabledModels` and `disabledProviders` entries may also be scoped to a path prefix:\n\n```yaml\nenabledModels:\n  - anthropic-model-sonnet-4-5\n  - path: ~/work\n    models:\n      - anthropic/anthropic-model-opus-4-5\ndisabledProviders:\n  - ollama\n  - path: ~/private\n    providers:\n      - anthropic\n```\n\nString entries apply everywhere. Scoped entries apply when the current working directory is the configured path or one of its subdirectories. Use `path`, `paths`, `pathPrefix`, or `pathPrefixes`; use `models` for `enabledModels`, `providers` for `disabledProviders`, or `values` for either.\n\n## `/model` and `--list-models`\n\nBoth surfaces keep provider-prefixed models visible and selectable.\n\nThey now also expose canonical/coalesced models:\n\n- `/model` includes a canonical view alongside provider tabs\n- `--list-models` prints a canonical section plus the concrete provider rows\n\nSelecting a canonical entry stores the canonical selector. Selecting a provider row stores the explicit `provider/modelId`.\n\n## Context promotion (model-level fallback chains)\n\nContext promotion is an overflow recovery mechanism for small-context variants (for example `*-spark`) that automatically promotes to a larger-context sibling when the API rejects a request with a context length error. It is **off by default** (`contextPromotion.enabled` is `false`); opt in to enable it.\n\n### Trigger and order\n\nWhen a turn fails with a context overflow error (e.g. `context_length_exceeded`), `AgentSession` attempts promotion **before** falling back to compaction:\n\n1. If `contextPromotion.enabled` is true, resolve a promotion target (see below).\n2. If a target is found, switch to it and retry the request — no compaction needed.\n3. If no target is available, fall through to auto-compaction on the current model.\n\n### Target selection\n\nSelection is model-driven, not role-driven:\n\n1. `currentModel.contextPromotionTarget` (if configured)\n2. smallest larger-context model on the same provider + API\n\nCandidates are ignored unless credentials resolve (`ModelRegistry.getApiKey(...)`).\n\n### OpenAI code provider websocket handoff\n\nIf switching from/to `openai-codex-responses`, session provider state key `openai-codex-responses` is closed before model switch. This drops websocket transport state so the next turn starts clean on the promoted model.\n\n### Persistence behavior\n\nPromotion uses temporary switching (`setModelTemporary`):\n\n- recorded as a temporary `model_change` in session history\n- does not rewrite saved role mapping\n\n### Configuring explicit fallback chains\n\nConfigure fallback directly in model metadata via `contextPromotionTarget`.\n\n`contextPromotionTarget` accepts either:\n\n- `provider/model-id` (explicit)\n- `model-id` (resolved within current provider)\n\nExample (`models.yml`) for Spark -> non-Spark on the same provider:\n\n```yaml\nproviders:\n  openai-code:\n    modelOverrides:\n      gpt-5.3-openai-code-spark:\n        contextPromotionTarget: openai-code/gpt-5.3-openai-code\n```\n\nThe built-in model generator also assigns this automatically for `*-spark` models when a same-provider base model exists.\n\n## Compatibility and routing fields\n\nThe `compat` block on a provider or model overrides the URL-based auto-detection in `packages/ai/src/providers/openai-completions-compat.ts`. It is validated by `OpenAICompatSchema` in `packages/coding-agent/src/config/model-registry.ts` and consumed by every `openai-completions` transport (`packages/ai/src/providers/openai-completions.ts`). The canonical type is `OpenAICompat` in `packages/ai/src/types.ts`.\n\n`models.yml` accepts the following keys (all optional; unset falls back to URL detection):\n\nRequest shaping:\n\n- `supportsStore` — emit `store: false` on requests. Default: auto (off for non-standard endpoints).\n- `supportsDeveloperRole` — use the `developer` system role for reasoning models instead of `system`. Default: auto.\n- `sendSessionHeaders` — forward the agent session id as `session_id` and `x-session-id` request headers so OpenAI-compatible relays/proxies can do session-affinity routing and reuse a server-side prompt cache. Default: `false`. Caller-set `headers`/`requestTransform` values are never overwritten.\n- `supportsResponsesSessionAffinity` — for `openai-responses`, opt in to forwarding `session_id` and `x-client-request-id` affinity headers to a custom OpenAI-compatible relay. Canonical OpenAI routing remains automatic; known non-OpenAI provider IDs are rejected. Default: `false`.\n- `supportsUsageInStreaming` — send `stream_options: { include_usage: true }` to receive token usage on streaming responses. Default: `true`.\n- `maxTokensField` — `\"max_completion_tokens\"` or `\"max_tokens\"`. Default: auto.\n- `supportsToolChoice` — emit the `tool_choice` parameter when the caller forces a specific tool. Default: `true`. Set `false` for endpoints that 400 on `tool_choice` (e.g. DeepSeek when reasoning is on).\n- `disableReasoningOnForcedToolChoice` — drop `reasoning_effort` / OpenRouter `reasoning` whenever `tool_choice` forces a call. Default: auto (Kimi/Anthropic-fronted endpoints).\n- `extraBody` — extra top-level fields merged into every request body (gateway hints, controller selectors, etc.).\n\nReasoning / thinking:\n\n- `supportsReasoningEffort` — accept `reasoning_effort`. Default: auto (off for Grok and zAI).\n- `reasoningEffortMap` — partial map from internal effort levels (`minimal|low|medium|high|xhigh`) to provider-specific strings (e.g. DeepSeek maps `xhigh -> \"max\"`).\n- `thinkingFormat` — request shape for thinking: `\"openai\"` (`reasoning_effort`), `\"openrouter\"` (`reasoning: { effort }`), `\"zai\"` (`thinking: { type: \"enabled\" }`), `\"qwen\"` (top-level `enable_thinking`), or `\"qwen-chat-template\"` (`chat_template_kwargs.enable_thinking`). Default: `\"openai\"`.\n- `reasoningContentField` — assistant field carrying chain-of-thought: `\"reasoning_content\"`, `\"reasoning\"`, or `\"reasoning_text\"`. Default: auto.\n- `requiresReasoningContentForToolCalls` — assistant tool-call turns must round-trip the reasoning field (DeepSeek-R1, Kimi, OpenRouter when reasoning is on). Default: `false`.\n- `requiresAssistantContentForToolCalls` — assistant tool-call turns must include non-empty text content (Kimi). Default: `false`.\n\nTool / message normalization:\n\n- `requiresToolResultName` — tool-result messages need a `name` field (Mistral). Default: auto.\n- `requiresAssistantAfterToolResult` — a user message after a tool result needs an assistant turn in between. Default: auto.\n- `requiresThinkingAsText` — convert thinking blocks to text wrapped in `<thinking>` delimiters (Mistral). Default: auto.\n- `requiresMistralToolIds` — normalize tool-call ids to exactly 9 alphanumeric chars. Default: auto.\n- `supportsStrictMode` — accept the per-tool `strict` field on tool schemas. Default: conservative auto-detect per provider/baseUrl.\n- `toolStrictMode` — `\"all_strict\"` forces strict on every tool, `\"none\"` forces it off; unset keeps the existing per-tool mixed behavior.\n\nGateway routing (only applied when `baseUrl` matches the gateway):\n\n- `openRouterRouting.only` / `openRouterRouting.order` — provider routing on `openrouter.ai` (see <https://openrouter.ai/docs/provider-routing>).\n- `vercelGatewayRouting.only` / `vercelGatewayRouting.order` — provider routing on `ai-gateway.vercel.sh` (see <https://vercel.com/docs/ai-gateway/models-and-providers/provider-options>).\n\nProvider-level `compat` is the baseline; per-model `compat` is deep-merged on top, with `openRouterRouting`, `vercelGatewayRouting`, and `extraBody` merged as nested objects.\n\n### Anthropic compatibility (`anthropic-messages`)\n\nFor `anthropic-messages` models, `compat.promptCacheMode` and `compat.supportsLongCacheRetention` are configurable at provider, model, and `modelOverrides` levels. Provider-level `compat` is the baseline; model and override values merge on top.\n\nPrompt-cache modes:\n\n- `automatic` — emit one top-level `cache_control` marker and let the Anthropic-compatible endpoint advance the breakpoint as the conversation grows.\n- `explicit` — emit block-level breakpoints instead. Use this for endpoints that reject top-level `cache_control` but support Anthropic's explicit content-block markers.\n- `none` — emit no generated Anthropic cache controls. Per-request or configured `cacheRetention: none` also disables generated caching.\n\nWithout an explicit mode, canonical Anthropic endpoints default to `automatic`, Claude-family model ids on non-canonical compatible endpoints default to `explicit`, and unknown non-Claude compatible endpoints default to `none`. Non-canonical endpoints get the default ~5m lifetime unless they opt into `supportsLongCacheRetention: true`. Set `promptCacheMode: automatic` only when a gateway is known to pass through Anthropic's top-level cache control without adding conflicting block markers.\n\nIf a gateway attaches enough cache markers of its own that ours push the request past Anthropic's four-breakpoint limit, Anthropic rejects it with `A maximum of 4 blocks with cache_control may be provided.` Those extra markers are not visible in the request GJC builds, so the limit is handled at runtime rather than predicted. Because the rejection means \"too many\" rather than \"none allowed\", recovery reduces the generated breakpoints one step at a time: `explicit` mode normally emits two markers (a conversation-prefix anchor and a current-turn refresh point), so the first retry keeps only the prefix anchor, and generated caching is disabled entirely only if that is rejected too. The reduced setting persists for the rest of the provider session, so an endpoint with one free slot keeps caching its conversation prefix instead of losing caching altogether. Set `promptCacheMode: none` on a gateway that never has a free slot to skip the wasted attempts.\n\n```yaml\nproviders:\n  corp-anthropic:\n    baseUrl: https://proxy.example.com/anthropic\n    apiKeyEnv: CORP_ANTHROPIC_API_KEY\n    api: anthropic-messages\n    compat:\n      promptCacheMode: explicit\n      supportsLongCacheRetention: false\n    models:\n      - id: claude-sonnet-4-5\n        contextWindow: 200000\n        maxTokens: 8192\n```\n\nOther Anthropic-side compatibility knobs such as `disableAdaptiveThinking` and `supportsEagerToolInputStreaming` remain built-in catalog metadata rather than `models.yml` fields. `disableStrictTools` stays a provider-level setting (below).\n\n### Strict tool schemas (`disableStrictTools`)\n\nAnthropic's API supports a `strict` field on tool definitions that forces the model to always follow the provided schema exactly. This is enabled by default for all `anthropic-messages` providers because it guarantees schema conformance in agentic systems.\n\nThird-party providers that front the Anthropic API (AWS Bedrock, Azure, self-hosted proxies) do not always implement this field and will reject requests that include it. Set `disableStrictTools: true` at the provider level to opt out:\n\n```yaml\nproviders:\n  bedrock-anthropic:\n    baseUrl: https://bedrock-runtime.us-east-1.amazonaws.com/anthropic\n    apiKey: AWS_BEARER_TOKEN\n    api: anthropic-messages\n    disableStrictTools: true\n    models:\n      - id: anthropic-model-sonnet-4-20250514\n        name: Anthropic model Sonnet 4 (Bedrock)\n        input: [text, image]\n        contextWindow: 200000\n        maxTokens: 16384\n        cost:\n          input: 3.00\n          output: 15.00\n          cacheRead: 0.30\n          cacheWrite: 3.75\n```\n\n`disableStrictTools` is a provider-level flag that applies to all models in the provider.\n\nTool schemas going on the wire are normalized by the unified flow in\n`packages/ai/src/utils/schema/normalize.ts` (Google/CCA/MCP dispatchers\nplus the OpenAI strict-mode sanitize+enforce pipeline). See\n[`ai-schema-normalize.md`](./ai-schema-normalize.md) for the strict-mode\nedge cases (local `$ref` inlining, single-item `allOf` collapse,\n`anyOf`-wrapper description hoist, enum/const primitive-type inference)\nand the per-provider dispatcher mapping.\n## Practical examples\n\n### Local OpenAI-compatible endpoint (no auth)\n\n```yaml\nproviders:\n  local-openai:\n    baseUrl: http://127.0.0.1:8000/v1\n    auth: none\n    api: openai-completions\n    models:\n      - id: Qwen/Qwen2.5-Coder-32B-Instruct\n        name: Qwen 2.5 Coder 32B (local)\n```\n\n### Hosted proxy with env-based key\n\n```yaml\nproviders:\n  anthropic-proxy:\n    baseUrl: https://proxy.example.com/anthropic\n    apiKey: ANTHROPIC_PROXY_API_KEY\n    api: anthropic-messages\n    authHeader: true\n    disableStrictTools: true  # if the proxy doesn't support strict tool schemas\n    models:\n      - id: anthropic-model-sonnet-4-20250514\n        name: Anthropic model Sonnet 4 (Proxy)\n        reasoning: true\n        input: [text, image]\n```\n\n### Override built-in provider route + model metadata\n\n```yaml\nproviders:\n  openrouter:\n    baseUrl: https://my-proxy.example.com/v1\n    headers:\n      X-Team: platform\n    modelOverrides:\n      anthropic/anthropic-model-sonnet-4:\n        name: Sonnet 4 (Corp)\n        compat:\n          openRouterRouting:\n            only: [anthropic]\n```\n\n## Legacy consumer caveat\n\nMost model configuration now flows through `models.yml` via `ModelRegistry`. Explicit `.json` / `.jsonc` paths remain supported only when passed programmatically to `ModelRegistry`; the default user config is `~/.gjc/agent/models.yml`.\n\n## Failure mode\n\nIf `models.yml` fails schema or validation checks:\n\n- registry keeps operating with built-in models\n- error is exposed via `ModelRegistry.getError()` and surfaced in UI/notifications\n",
 	"multi-vendor-profiles.md": "# Choosing models in GJC: role-based profiles\n\nA practical guide to picking models for GJC's roles, for every subscription situation — one vendor, two vendors, or the full multi-vendor set. It adds curated cross-vendor `profiles:` for `~/.gjc/agent/models.yml` and verified selector notes on top of the mechanism in [Model profiles](./models.md#model-profiles---mpreset). Everything here is **user config**; it complements the built-in `--mpreset` presets and overrides a built-in only when it shares its exact name.\n\n> Selectors, prices, and \"axis leaders\" are catalog- and time-sensitive (selectors and prices observed 2026-07 on the current bundled catalog; the measured latency and single-message-limit notes below were observed 2026-06 on `claude-opus-4-8` and have not been re-measured on `claude-opus-5`). Re-verify any selector with `gjc -p --no-session --no-tools --model <selector> \"Reply OK\"`.\n\n## The five roles\n\n`default` runs the main loop and most turns; `executor` / `architect` / `planner` / `critic` are the four bundled task agents, delegated only when the work calls for it.\n\n| Role | What it optimizes for |\n| --- | --- |\n| `default` | tool-calling reliability + honesty (it routes — its quality bounds the whole system) |\n| `executor` | real coding (SWE-bench Verified) |\n| `planner` | reasoning + sequencing (GPQA / ARC-AGI-2) |\n| `architect` | large-context + multimodal review |\n| `critic` | independent adversarial review (different family from what it reviews) |\n\n## Pick by what you subscribe to\n\n| You have | Use |\n| --- | --- |\n| **One vendor** | the built-in preset for that vendor — `claude-opus` (Anthropic), `codex-{eco,medium,pro}` (OpenAI/Codex), `opencodego` (OpenCode Go), or a single-vendor flagship tier (`zai/glm-5.2`, `kimi-code/...`, `xiaomi/...`, `xai/grok-4.3`, `minimax-code/...`). These already map all five roles inside one vendor. |\n| **Claude + Codex** | the built-in `opus-codex` (Claude main loop + Codex support roles). |\n| **Three or more / all five** | the cross-vendor profiles below — each role on its axis leader, `critic` kept cross-family. |\n\nThe single guiding rule across all of these: **keep `default` on the strongest router you have** (Anthropic Opus when available). A weak `default` caps quality regardless of the delegated models.\n\n## Cross-vendor profiles (3+ vendors)\n\nNo single vendor leads every axis, so these put each role on its axis leader and keep `critic` on a different family from the `executor` it reviews.\n\n```yaml\nprofiles:\n\n  daily:                 # everyday balance\n    required_providers: [anthropic, openai-codex, google-antigravity, xai]\n    model_mapping:\n      default:   anthropic/claude-opus-5:medium\n      executor:  openai-codex/gpt-5.4:high\n      planner:   google-antigravity/gemini-3.1-pro-low:high\n      architect: google-antigravity/gemini-3.1-pro-low:high\n      critic:    xai/grok-4.3:medium\n\n  ultimate:              # cost-no-object, best per role\n    required_providers: [anthropic, openai-codex, google-antigravity, xai]\n    model_mapping:\n      default:   anthropic/claude-opus-5:high\n      executor:  anthropic/claude-opus-5:max\n      planner:   openai-codex/gpt-5.5:xhigh\n      architect: google-antigravity/gemini-3.1-pro-low:high\n      critic:    xai/grok-4.3:high\n\n  eco:                   # cheapest delegated work; main loop stays on Opus\n    required_providers: [anthropic, opencode-go, google-antigravity, xai]\n    model_mapping:\n      default:   anthropic/claude-opus-5:low\n      executor:  opencode-go/deepseek-v4-flash\n      planner:   xai/grok-4-1-fast:high\n      architect: google-antigravity/gemini-3.1-pro-low\n      critic:    google-antigravity/gemini-3.5-flash\n\n  monorepo:              # huge codebases (openai-codex excluded: 372k context cap)\n    required_providers: [anthropic, google-antigravity, opencode-go]\n    model_mapping:\n      default:   anthropic/claude-opus-5:medium\n      executor:  anthropic/claude-opus-5:high\n      planner:   google-antigravity/gemini-3.1-pro-low:high\n      architect: anthropic/claude-opus-5:high\n      critic:    opencode-go/glm-5.2\n\n  reviewer:              # review/audit stance — the author-mode role split, inverted\n    required_providers: [anthropic, openai-codex, google-antigravity]\n    model_mapping:\n      default:   anthropic/claude-opus-5:high                 # aggregator restraint: preserve raw reviewer verdicts\n      executor:  openai-codex/gpt-5.5:high                      # support — repro PoCs, failing tests, harnesses\n      planner:   google-antigravity/gemini-3.1-pro-low:high     # review checklists / audit scoping\n      architect: anthropic/claude-opus-5:high                 # lead 1 — primary code-review judge (effective long-context)\n      critic:    openai-codex/gpt-5.5:high                      # lead 2 — merge gate, cross-family vs Claude-authored code\n```\n\n## Reviewer stance and the external review gate\n\nThe profiles above assume an **authoring** stance: `executor` is the lead and `architect`/`critic` verify its work. In a session whose primary job is reviewing or auditing (not writing) code, the roles invert — `architect`/`critic` become the leads and `executor` is support (reproduction PoCs, failing tests). The `reviewer` profile encodes that inversion, with one generalized provenance rule: **the reviewing model family must differ from the family that authored the code under review**, not merely from the session's own executor.\n\nA verified use is the cross-session final review gate: the authoring session launches a fresh, stateless reviewer sub-session so the finished diff is judged without the authoring context:\n\n```sh\n# the one-shot gate needs only a cross-family --model; add --mpreset reviewer as an\n# optional enhancement AFTER installing this profile in ~/.gjc/agent/models.yml:\ngjc -p --no-session --model openai-codex/gpt-5.5:xhigh --tools read,search,find \"<review prompt: diff + spec paths, severity findings, final line VERDICT: APPROVE|REQUEST_CHANGES>\"\n```\n\nThe `--tools` allowlist is part of the contract: it enforces the reviewer's read-only boundary for the built-in tool surface instead of trusting the prompt (the runtime still injects the session `goal` tool unless `goal.enabled` is off — disabling it for the reviewer invocation is **mandatory**, via a dedicated gate directory outside the repo so the reviewed checkout stays clean, see the template — plus `generate_image` when an image credential exists). In this one-shot form the session's `default` model authors the verdict — a tool-restricted print session cannot delegate to the profile's `critic`/`architect` roles — so the explicit cross-family `--model` carries provenance, and the `reviewer` profile itself serves the interactive review-session case (activate it with `--mpreset reviewer` only after copying it into `models.yml`; otherwise activation fails with an unknown-profile error). Profile names in this document live in the user namespace — a user profile overrides a builtin preset only on an exact name match, and a future builtin with the same name would be silently shadowed by your copy.\n\nSee [Extragoal local skill template](./extragoal-skill-template.md) for the full gate workflow (verdict contract, findings triage, bounded re-sign loop, secret-scan and injection guards) built on this recipe.\n\n## Model cheatsheet (by need)\n\nCurrent axis leaders and the cheaper second option, with metered price ($/1M in/out; Gemini via Antigravity runs on the Google AI subscription):\n\n| Need | First pick | Cheaper option |\n| --- | --- | --- |\n| Router / tool-calling (`default`) | `anthropic/claude-opus-5` (5/25) | `anthropic/claude-sonnet-5` (3/15) |\n| Coding (`executor`) | `anthropic/claude-opus-5` (5/25) — the prior `claude-opus-4-8` scored SWE-bench Verified ~88.6; no Opus 5 measurement yet | `openai-codex/gpt-5.4` (2.5/15) · `opencode-go/deepseek-v4-flash` (0.14/0.28) |\n| Reasoning (`planner`) | `openai-codex/gpt-5.5` (ARC-AGI-2) / `google-antigravity/gemini-3.1-pro-low:high` (GPQA) | `xai/grok-4-1-fast` (0.2/0.5) |\n| Large context (`architect`) | `anthropic/claude-opus-5` (effective long-context) | `xai/grok-4-fast` (2M nominal, 0.2/0.5) |\n| Multimodal review (`architect`) | `google-antigravity/gemini-3.1-pro-low:high` | `google-antigravity/gemini-3.5-flash` |\n| Independent critic | `xai/grok-4.3` (1.25/2.5) | `opencode-go/glm-5.2` · `google-antigravity/gemini-3.5-flash` |\n\nOn standard tasks, all current frontier models in the catalog are accurate; **pick by cost, latency, and role fit, not by raw accuracy on easy prompts.** As an indicative GJC-routed latency reference (`gjc -p`, identical coding + reasoning prompts, all correct): `grok-4.3` and `glm-5.2` ≈ 2–3s, `deepseek-v4-pro` ≈ 3–4s, `claude-opus-4-8` / `gpt-5.5` ≈ 4–7s, `gemini-3.1-pro-low:high` ≈ 7s. `claude-opus-5` shares Opus 4.8's published context/output envelope but has not been latency-measured here.\n\n## Verified selector notes (current catalog)\n\nObserved via live `gjc -p` calls; useful when wiring the profiles above:\n\n- **Antigravity Gemini, high reasoning** → use `google-antigravity/gemini-3.1-pro-low:high`. The id `gemini-3.1-pro-high` returns HTTP 400 (no matching backend model); `thinkingLevel` is a per-request parameter, so raising it on `gemini-3.1-pro-low` invokes the model's native high-reasoning mode rather than a degraded one.\n- **openai-codex on a ChatGPT account** serves base GPT only (`gpt-5.5`, `gpt-5.4`). Standalone `-codex` variants (`gpt-5.3-codex`, `gpt-5.2-codex`, `gpt-5.1-codex-max` / `-mini`) return `not supported when using Codex with a ChatGPT account`.\n- **Single-message input limit is separate from the context window.** Measured on `claude-opus-4-8` (not yet re-measured on `claude-opus-5`, which publishes the same 1M window): the model runs with a 1M window via multi-turn accumulation, but a single `@file` message above ~400k tokens returns 400 on `anthropic` / `google-antigravity`; `xai` / `opencode-go` accept larger single messages. Chunk very large inputs across turns instead of pasting one block.\n- **Some selectors come from a provider's live catalog, not the bundled snapshot.** `opencode-go/glm-5.2` and `google-antigravity/gemini-3.5-flash` resolved in `gjc -p` tests but are **not** in `packages/ai/src/models.json`; they appear only after the provider's online model discovery has populated the registry. `required_providers` verifies credentials at activation — it does **not** guarantee fresh, non-stale discovery — so activation can still fail with `selector did not resolve` until discovery runs (re-login or retry to refresh). If you hit that, substitute a bundled id: `opencode-go/deepseek-v4-pro` for the critic, or `zai/glm-5.2` (add `zai` to `required_providers`) for GLM 5.2.\n\n## Activation\n\n```bash\ngjc --mpreset daily               # this session only\ngjc --mpreset ultimate --default  # persist as the startup default (config.yml)\n```\n\nActivation hard-blocks when any provider in `required_providers` lacks credentials, so log in first: `/login anthropic`, `/login openai-codex`, `/login google-antigravity`, `/login xai` (and `opencode-go` via `OPENCODE_API_KEY`).\n\n### Serving cross-vendor profiles through one OpenAI-compatible proxy\n\nWhen a single gateway (LiteLLM, OpenRouter, or a custom proxy) fronts several vendors, you do not need to configure every `required_providers` entry directly. Add the gateway as a provider — `gjc setup provider --preset litellm --base-url <url>` or `gjc setup provider --preset openai-compatible-proxy --base-url <url>` — and point `modelProfile.proxyProvider` at it in `config.yml`:\n\n```yaml\nmodelProfile:\n  proxyProvider: litellm\n  proxyMode: always # route all supported built-in preset selectors through the gateway\n```\n\n`proxyMode: fallback` is the default and uses the gateway only when the direct provider is unauthenticated. Set `proxyMode: always` when the gateway must be the single audit, quota, or spend-control surface: it routes all proxy-routable **built-in** preset selectors through the proxy even when direct credentials exist. The configured proxy must be authenticated and expose every routed model; activation fails closed for missing or ambiguous models. User-defined profiles are never rewritten — set their selectors to `litellm/…` explicitly if you want them proxied. Routing and fail-closed behavior are documented in [Routing built-in presets through a proxy](./models.md#routing-built-in-presets-through-a-proxy-modelprofileproxyprovider).\n",
 	"native-ffi-optimization-policy.md": "# ADR: Native FFI Optimization Policy\n\n- Status: Accepted\n- Scope: `crates/pi-natives` algorithmic ports proposed for performance reasons\n- Related: [`porting-to-natives.md`](./porting-to-natives.md), [`natives-architecture.md`](./natives-architecture.md), [`natives-binding-contract.md`](./natives-binding-contract.md), [`cpu-hotspot-map.json`](./cpu-hotspot-map.json), [`hotspot-map-successor.md`](./hotspot-map-successor.md)\n\n## Decision\n\nA new native (Rust N-API / FFI) port proposed **to optimize a leftover hot path** does not land unless **all** of the following gates pass:\n\n1. **Corpus evidence** — a profiling-corpus trace shows the path has user-visible latency or RSS impact on a representative workload (not just a static complexity argument).\n2. **Self-time attribution** — a `profilerSelfTime` artifact identifies the proposed hotspot, **or** fallback-toggle evidence proves an end-to-end benefit without byte changes. Wall-clock proxy timing alone is never sufficient.\n3. **Measured FFI overhead** — the N-API call/marshalling overhead is measured against the JS/TS baseline, not assumed away.\n4. **Representative win** — a representative p50/p95 win exists on realistic inputs, not only microbenchmark seed results.\n5. **Byte parity** — a byte-identical corpus covers rendered, persisted, and provider-visible bytes for the changed path.\n6. **Operational cost** — fallback, packaging, and rollback costs are documented.\n\nThis policy governs **speculative algorithmic ports**. It does **not** re-litigate already-native platform/system surfaces (see [Scope boundary](#scope-boundary)).\n\n## Context\n\nThe CPU/memory hotspot program (Optimization Suites v1–v3, tracked in [`cpu-hotspot-map.json`](./cpu-hotspot-map.json)) is closed out. Its prioritization was a **static structural ranking** (algorithmic complexity × trigger frequency), and the map's own `method` field records that real CPU self-time was \"to be measured by the agreed profiling corpus during optimization.\" That corpus is being built separately; until its evidence exists, new native ports for leftover hotspots would repeat the same evidence gap.\n\nThe suites already produced concrete decisions that this policy codifies so they are not re-discovered:\n\n- **v2 (#530)** measured and **rejected the five remaining Rust port candidates** per the FFI cost gates after shipping only `diffLines` (H03) natively. Native overhead did not beat the JS/TS baseline for those candidates on realistic inputs.\n- **v3 (#558) rejected a native word-diff (H04)** \"without a fresh FFI gate\" — the TS fast paths were retained instead; a native port would need to re-clear gates 1–6 above.\n- **Hunt-Szymanski LCS (H05)** was implemented as a native/algorithmic replacement, then **reverted** because it produced byte-different rendered diffs (reproduced by red-team). Byte parity is the gate, not raw speed.\n- **The custom JSON length counter (H08)** was implemented, made exact, then **deleted** — an exact JS reimplementation was not faster than native `JSON.stringify`. \"More native\" is not automatically \"faster.\"\n\nThese four precedents share a root cause: a plausible algorithmic/native win that failed a real gate (cost, byte parity, or end-to-end benefit). The policy makes those gates a precondition rather than a post-hoc discovery.\n\n## Evidence taxonomy\n\nNative-port claims must classify their evidence using the same separated classes as the profiling corpus. These classes must never be conflated:\n\n- **`wallClockPhase`** — elapsed timing around a phase or operation. Useful for perceived-latency and regression detection; **insufficient** to confirm CPU self-time or to justify a port on its own.\n- **`processCpuUsage`** — `process.cpuUsage()` user/system deltas, optionally normalized by elapsed time. Indicates process-level CPU pressure; **cannot** attribute self-time to a specific hotspot.\n- **`profilerSelfTime`** — profiler (or equivalent sampled/trace) attribution of self-time to a function, module, or native symbol. **Required** before a hotspot may be called \"CPU-self-time confirmed.\"\n\nA native-optimization proposal that cites only `wallClockPhase` or `processCpuUsage` is **not** CPU-self-time confirmed and does not clear gate 2.\n\n## Approval checklist\n\nBefore opening a native-optimization PR, confirm and attach evidence for each:\n\n- [ ] Corpus trace shows user-visible latency or RSS impact for the path (gate 1).\n- [ ] `profilerSelfTime` artifact identifies the hotspot, **or** fallback-toggle before/after evidence proves end-to-end benefit without byte changes (gate 2).\n- [ ] FFI/marshalling overhead measured vs the JS/TS baseline in the same benchmark run (gate 3).\n- [ ] Representative p50/p95 win on realistic inputs, not only seeded microbench results (gate 4).\n- [ ] Byte-identical corpus covers rendered, persisted, and provider-visible bytes (gate 5).\n- [ ] Fallback, packaging (platform variants / embedded addon), and rollback costs documented (gate 6).\n\nIf any box is unchecked, keep the work in TypeScript or hold it as a tracked candidate; do not switch callsites. This mirrors the existing **Rule of thumb** in [`porting-to-natives.md`](./porting-to-natives.md): if native is not faster *and* behavior-compatible, do not switch callsites.\n\n## Scope boundary\n\nThis policy targets **speculative algorithmic ports**, not the established native surface. The following are **already native** by design and are explicitly out of scope (see `alreadyNativeExcluded` in [`cpu-hotspot-map.json`](./cpu-hotspot-map.json)):\n\n`grep`, `fd`/`glob`, text width/wrap/truncate/slice, syntax highlighting, HTML→Markdown, AST, summary, process/PTY/shell, SIXEL, clipboard, `Bun.hash.xxHash32/64`, and `JSON.parse`/`JSON.stringify`.\n\nThese are native because they are I/O, OS/process integration, or platform primitives — the criteria in [`porting-to-natives.md`](./porting-to-natives.md#when-to-port). Distinguishing them from algorithmic ports matters: a leftover algorithmic hotspot must clear gates 1–6, whereas adding a new OS/process/native-primitive binding follows the standard porting guide.\n\n## Consequences\n\n- New native algorithmic ports require profiling-corpus evidence and a measured cost gate before review; this slows speculative optimization but prevents byte-parity regressions and dead native code.\n- The default answer for a leftover hotspot is \"keep it in TypeScript\" until the corpus proves it matters.\n- Already-native platform/system primitives and new OS/process bindings are unaffected; they follow [`porting-to-natives.md`](./porting-to-natives.md) as before.\n- Reviewers can reject a native-optimization PR purely on a missing gate, citing this ADR, without re-deriving the rationale.\n\n## Follow-ups\n\n- Held native candidates (H04 word-diff, H05 LCS, and other v2-rejected candidates) stay held unless a future PR clears gates 1–6 with fresh corpus evidence.\n- When the profiling corpus lands, link its threshold/evidence ledger here so native-port proposals can cite concrete corpus artifacts.\n",
 	"natives-addon-loader-runtime.md": "# Natives Addon Loader Runtime\n\nThis document covers the runtime loader shipped by `@gajae-code/natives`: how `native/index.js` decides which `.node` file to require, how compiled-binary embedded payloads are extracted, and what startup failures report.\n\n## Implementation files\n\n- `packages/natives/native/index.js`\n- `packages/natives/native/loader-state.js`\n- `packages/natives/native/embedded-addon.js`\n- `packages/natives/scripts/embed-native.ts`\n- `packages/natives/package.json`\n\n## Scope and responsibility\n\nThe loader is intentionally narrow:\n\n- Build a platform/CPU-aware candidate list for addon filenames and directories.\n- Treat an embedded-addon manifest as the authoritative compiled-binary signal when present.\n- Optionally materialize an embedded addon into a versioned per-user cache directory.\n- Attempt candidates in deterministic order and return the first addon that `require(...)` loads.\n\nThe current loader does **not** run a separate `validateNative(...)` export-presence gate. API shape is provided by the generated N-API binding file (`native/index.d.ts`) and the loaded addon itself. A stale binary therefore normally fails as a missing property or native load error rather than as a custom \"missing exports\" validation error.\n\n## Runtime inputs and derived state\n\nAt module initialization, `native/index.js` computes:\n\n- **Platform tag**: `${process.platform}-${process.arch}` (for example `darwin-arm64`).\n- **Package version**: from `packages/natives/package.json`.\n- **Core directories**:\n  - `nativeDir`: package-local `packages/natives/native`.\n  - `execDir`: directory containing `process.execPath`.\n  - `versionedDir`: `<getNativesDir()>/<packageVersion>`.\n  - `userDataDir` fallback:\n    - Windows: `%LOCALAPPDATA%/gjc` or `%USERPROFILE%/AppData/Local/gjc`.\n    - Non-Windows: `~/.local/bin`.\n- **Natives cache root** (`getNativesDir()`):\n  - if `$XDG_DATA_HOME/gjc` exists, `$XDG_DATA_HOME/gjc/natives`;\n  - otherwise `~/.gjc/natives`.\n- **Compiled-binary mode** (`detectCompiledBinary`): true if any of:\n  - embedded-addon manifest is non-null,\n  - `GJC_COMPILED` env var is set,\n  - `import.meta.url` contains Bun embedded markers (`$bunfs`, `~BUN`, `%7EBUN`).\n- **Variant override**: `GJC_NATIVE_VARIANT` (`modern`/`baseline` only; invalid values ignored).\n- **Selected variant**: explicit override, otherwise runtime AVX2 detection on x64 (`modern` if AVX2, else `baseline`).\n\n## Platform support and tag resolution\n\n`SUPPORTED_PLATFORMS` is fixed to:\n\n- `linux-x64`\n- `linux-arm64`\n- `darwin-arm64`\n- `win32-x64`\n\nUnsupported platforms are not rejected before probing. The loader first tries the computed candidate paths. If all fail and `platformTag` is unsupported, it throws an unsupported-platform error listing supported tags.\n\n## Variant selection (`modern` / `baseline` / default)\n\n### x64 behavior\n\n1. `GJC_NATIVE_VARIANT=modern|baseline` wins when valid.\n2. Otherwise AVX2 support is detected:\n   - Linux: scan `/proc/cpuinfo` for `avx2`.\n   - macOS: `sysctl -n machdep.cpu.leaf7_features`, then `machdep.cpu.features`.\n   - Windows: PowerShell `[System.Runtime.Intrinsics.X86.Avx2]::IsSupported`.\n3. AVX2 selects `modern`; unavailable or undetectable AVX2 selects `baseline`.\n\n### Non-x64 behavior\n\nNo variant suffix is used; the filename is `pi_natives.<platform>-<arch>.node`.\n\n### Filename construction\n\n`loader-state.js#getAddonFilenames` returns:\n\n- Non-x64 or no variant: `pi_natives.<tag>.node`\n- x64 + `modern`:\n  1. `pi_natives.<tag>-modern.node`\n  2. `pi_natives.<tag>-baseline.node`\n  3. `pi_natives.<tag>.node`\n- x64 + `baseline`:\n  1. `pi_natives.<tag>-baseline.node`\n  2. `pi_natives.<tag>.node`\n\nThe default unsuffixed fallback remains part of the x64 candidate list.\n\n## Candidate path construction and fallback ordering\n\n`resolveLoaderCandidates(...)` expands every filename across directories, then de-duplicates while preserving first occurrence order.\n\n### Non-compiled runtime\n\nFor each filename, candidates are:\n\n1. `<nativeDir>/<filename>`\n2. `<execDir>/<filename>`\n\n### Compiled runtime\n\nFor each filename, candidates are:\n\n1. `<versionedDir>/<filename>`\n2. `<userDataDir>/<filename>`\n3. `<nativeDir>/<filename>`\n4. `<execDir>/<filename>`\n\nAt load time, an extracted embedded candidate, when produced, is prepended ahead of these de-duplicated candidates.\n\n## Embedded addon extraction lifecycle\n\n`embedded-addon.js` is generated by `scripts/embed-native.ts`. The reset stub exports `embeddedAddon = null`. A populated manifest has:\n\n- `platformTag`\n- `version`\n- `files[]` entries with `variant`, `filename`, and `filePath`\n\nExtraction (`maybeExtractEmbeddedAddon`) runs only when:\n\n1. compiled-binary mode is true,\n2. `embeddedAddon` is non-null,\n3. manifest `platformTag` equals the runtime platform tag,\n4. manifest `version` equals the package version,\n5. a variant-appropriate embedded file exists.\n\nVariant file selection:\n\n- Non-x64: prefer `default`, then first available file.\n- x64 + `modern`: prefer `modern`, fallback to `baseline`.\n- x64 + `baseline`: require `baseline`.\n\nMaterialization:\n\n1. Ensure `<versionedDir>` exists.\n2. Reuse `<versionedDir>/<selected filename>` if it already exists.\n3. Otherwise read `selectedEmbeddedFile.filePath` and write the target path.\n4. Return the target path as the first candidate.\n\nDirectory creation or write failures are appended to the loader error list; probing continues through normal candidates.\n\n## Lifecycle and state transitions\n\n```text\nInit\n  -> Load package metadata and embedded-addon manifest\n  -> Compute platform/version/variant/filenames/candidate paths\n  -> (compiled + embedded manifest matches?)\n       yes -> try extract to versionedDir (record errors, continue)\n       no  -> skip extraction\n  -> For each runtime candidate in order:\n       require(candidate)\n       -> success: return addon exports (READY)\n       -> failure: record error, continue\n  -> none loaded:\n       if unsupported platform tag -> throw Unsupported platform\n       else -> throw Failed to load (tried-path diagnostics + hints)\n```\n\n## Failure behavior and diagnostics\n\n### Unsupported platform\n\nIf all candidates fail and `platformTag` is not supported, the loader throws:\n\n- `Unsupported platform: <tag>`\n- supported platform list\n- issue-reporting guidance\n\n### No loadable candidate\n\nIf the platform is supported but no candidate can be loaded, the final error includes:\n\n- `Failed to load pi_natives native addon for <platformTag>` or `<platformTag> (<variant>)`\n- every attempted path with the corresponding `require(...)` error\n- mode-specific remediation hints\n\n### Compiled-binary startup failures\n\nCompiled mode diagnostics include:\n\n- expected versioned cache target paths (`<versionedDir>/<filename>`),\n- remediation to delete the versioned cache and rerun,\n- direct release download `curl` commands for each expected filename.\n\n### Non-compiled startup failures\n\nNormal package/runtime diagnostics include:\n\n- reinstall hint (`bun install @gajae-code/natives`),\n- local rebuild command (`bun --cwd=packages/natives run build`),\n- optional x64 variant build hint (`TARGET_VARIANT=baseline|modern bun --cwd=packages/natives run build`).\n",
@@ -75,6 +75,7 @@ export const EMBEDDED_DOCS: Readonly<Record<string, string>> = {
 	"prompt-architect-reports/tool-prompts.raw.md": "# ToolPrompts recovered raw report\n\nThe original `agent://0-ToolPrompts` result surfaced as failed, but inspecting the subagent JSONL context recovered 34 structured `report_finding` entries before the session died on stalls/429.\n\nCanonical recovered artifacts:\n\n- `recovered-context/0-ToolPrompts.recovered.md`\n- `recovered-context/0-ToolPrompts.findings.json`\n- `recovery-summary.md`\n\nRecovered severity breakdown: P1 = 4, P2 = 18, P3 = 12. No final `yield` or grade was emitted.\n",
 	"provider-streaming-internals.md": "# Provider streaming internals\n\nThis document explains how token/tool streaming is normalized in `@gajae-code/ai`, then propagated through `@gajae-code/agent-core` and `coding-agent` session events.\n\n## End-to-end flow\n\n1. `streamSimple()` (`packages/ai/src/stream.ts`) maps generic options and dispatches to a provider stream function.\n2. Provider stream functions translate provider-native stream events into the unified `AssistantMessageEvent` sequence. Current built-ins include Anthropic, OpenAI Responses/Completions/OpenAI code/Azure Responses, Google Gemini/Gemini CLI/Vertex, Bedrock Converse, Ollama, Cursorand GitLab Duo/Kimi wrappers.\n3. Each provider pushes events into `AssistantMessageEventStream` (`packages/ai/src/utils/event-stream.ts`), which throttles delta events and exposes:\n   - async iteration for incremental updates\n   - `result()` for final `AssistantMessage`\n4. `agentLoop` (`packages/agent/src/agent-loop.ts`) consumes those events, mutates in-flight assistant state, and emits `message_update` events carrying the raw `assistantMessageEvent`.\n5. `AgentSession` (`packages/coding-agent/src/session/agent-session.ts`) subscribes to agent events, persists messages, and applies session behaviors (retry, compaction, TTSR, streaming-edit abort checks).\n\n## Unified stream contract in `@gajae-code/ai`\n\nAll providers emit the same shape (`AssistantMessageEvent` in `packages/ai/src/types.ts`):\n\n- `start`\n- content block lifecycle triplets:\n  - text: `text_start` → `text_delta`\\* → `text_end`\n  - thinking: `thinking_start` → `thinking_delta`\\* → `thinking_end`\n  - tool call: `toolcall_start` → `toolcall_delta`\\* → `toolcall_end`\n- terminal event:\n  - `done` with `reason: \"stop\" | \"length\" | \"toolUse\"`\n  - or `error` with `reason: \"aborted\" | \"error\"`\n\n`AssistantMessageEventStream` guarantees:\n\n- final result is resolved by terminal event (`done` or `error`)\n- deltas are batched/throttled (~50ms)\n- buffered deltas are flushed before non-delta events and before completion\n\n## Delta throttling and harmonization behavior\n\n`AssistantMessageEventStream` treats `text_delta`, `thinking_delta`, and `toolcall_delta` as mergeable events:\n\n- buffered deltas are merged only when **type + contentIndex** match\n- merge keeps the latest `partial` snapshot\n- non-delta events force immediate flush\n\nThis smooths high-frequency provider streams for TUI/event consumers, but is not provider backpressure: providers still produce at full speed, while the local stream buffers.\n\n## Provider normalization details\n\n## Anthropic (`anthropic-messages`)\n\nSource: `packages/ai/src/providers/anthropic.ts`\n\nNormalization points:\n\n- `message_start` initializes usage (input/output/cache tokens)\n- `content_block_start` maps to text/thinking/toolcall starts\n- `content_block_delta` maps:\n  - `text_delta` → `text_delta`\n  - `thinking_delta` → `thinking_delta`\n  - `input_json_delta` → `toolcall_delta`\n  - `signature_delta` updates `thinkingSignature` only (no event)\n- `content_block_stop` emits corresponding `*_end`\n- `message_delta.stop_reason` maps via `mapStopReason()`\n\nTool-call argument streaming:\n\n- each tool block carries internal `partialJson`\n- every JSON delta appends to `partialJson`\n- `arguments` are reparsed on each delta via `parseStreamingJson()`\n- `toolcall_end` reparses once more, then strips `partialJson`\n\n## OpenAI Responses family (`openai-responses`, `openai-code-responses`, `azure-openai-responses`)\n\nSources: `packages/ai/src/providers/openai-responses.ts`, `openai-code-responses.ts`, and `azure-openai-responses.ts`\n\nNormalization points:\n\n- `response.output_item.added` starts reasoning/text/function-call blocks\n- reasoning summary events (`response.reasoning_summary_text.delta`) become `thinking_delta`\n- output/refusal deltas become `text_delta`\n- `response.function_call_arguments.delta` becomes `toolcall_delta`\n- `response.output_item.done` emits `thinking_end` / `text_end` / `toolcall_end`\n- `response.completed` maps status to stop reason and usage\n\nTool-call argument streaming:\n\n- same `partialJson` accumulation pattern as Anthropic\n- providers that send only `response.function_call_arguments.done` still populate final args\n- tool call IDs are normalized as `\"<call_id>|<item_id>\"`\n\n## Google Generative AI (`google-generative-ai`)\n\nSource: `packages/ai/src/providers/google.ts`\n\nNormalization points:\n\n- iterates `candidate.content.parts`\n- text parts are split into thinking vs text by `isThinkingPart(part)`\n- block transitions close previous block before starting a new one\n- `part.functionCall` is treated as a complete tool call (start/delta/end emitted immediately)\n- finish reason mapped by `mapStopReason()` from `google-shared.ts`\n\nTool-call argument streaming:\n\n- function call args arrive as structured object, not incremental JSON text\n- implementation emits one synthetic `toolcall_delta` containing `JSON.stringify(arguments)`\n- no partial JSON parser needed for Google in this path\n\n## Partial tool-call JSON accumulation and recovery\n\nShared behavior for Anthropic/OpenAI Responses uses `parseStreamingJson()` (`packages/ai/src/utils/json-parse.ts`):\n\n1. try `JSON.parse`\n2. fallback to `partial-json` parser for incomplete fragments\n3. if both fail, return `{}`\n\nImplications:\n\n- malformed or truncated argument deltas do not crash stream processing immediately\n- in-progress `arguments` may temporarily be `{}`\n- later valid deltas can recover structured arguments because parsing is retried on every append\n- final `toolcall_end` performs one more parse attempt before emission\n\n## Stop reasons vs transport/runtime errors\n\nProvider stop reasons are mapped to normalized `stopReason`:\n\n- Anthropic: `end_turn`→`stop`, `max_tokens`→`length`, `tool_use`→`toolUse`, safety/refusal cases→`error`\n- OpenAI Responses: `completed`→`stop`, `incomplete`→`length`, `failed/cancelled`→`error`\n- Google: `STOP`→`stop`, `MAX_TOKENS`→`length`, safety/prohibited/malformed-function-call classes→`error`\n\nError semantics are split in two stages:\n\n1. **Model completion semantics** (provider reported finish reason/status)\n2. **Transport/runtime failure** (network/client/parser/abort exceptions)\n\nIf provider stream throws or signals failure, each provider wrapper catches and emits terminal `error` event with:\n\n- `stopReason = \"aborted\"` when abort signal is set\n- otherwise `stopReason = \"error\"`\n- `errorMessage = formatErrorMessageWithRetryAfter(error)`\n\n## Malformed chunk / SSE parse failure behavior\n\nFor these provider paths, chunk/SSE framing is handled by vendor SDK streams (Anthropic SDK, OpenAI SDK, Google SDK). This code does not implement a custom SSE decoder here.\n\nObserved behavior in current implementation:\n\n- malformed chunk/SSE parsing at SDK level surfaces as an exception or stream `error` event\n- provider wrapper converts that into unified terminal `error` event\n- no provider-specific resume/retry inside the stream function itself\n- higher-level retries are handled in `AgentSession` auto-retry logic (message-level retry, not stream-chunk replay)\n\n## Cancellation boundaries\n\nCancellation is layered:\n\n- AI provider request: `options.signal` is passed into provider client stream call.\n- Provider wrapper: after stream loop, aborted signal forces error path (`\"Request was aborted\"`).\n- Agent loop: checks `signal.aborted` before handling each provider event and can synthesize an aborted assistant message from the latest partial.\n- Session/agent controls: `AgentSession.abort()` -> `agent.abort()` -> shared abort controller cancellation.\n\nTool execution cancellation is separate from model stream cancellation:\n\n- tool runners use `AbortSignal.any([agentSignal, steeringAbortSignal])`\n- steering interrupts can abort remaining tool execution while preserving already-produced tool results\n\n## Backpressure boundaries\n\nThere is no hard backpressure mechanism between provider SDK stream and downstream consumers:\n\n- `EventStream` uses in-memory queues with no max size\n- throttling reduces UI update rate but does not slow provider intake\n- if consumers lag significantly, queued events can grow until completion\n\nCurrent design favors responsiveness and simple ordering over bounded-buffer flow control.\n\n## How stream events surface as agent/session events\n\n`agentLoop.streamAssistantResponse()` bridges `AssistantMessageEvent` to `AgentEvent`:\n\n- on `start`: pushes placeholder assistant message and emits `message_start`\n- on block events (`text_*`, `thinking_*`, `toolcall_*`): updates last assistant message, emits `message_update` with raw `assistantMessageEvent`\n- on terminal (`done`/`error`): resolves final message from `response.result()`, emits `message_end`\n\n`AgentSession` then consumes those events for session-level behaviors:\n\n- TTSR watches `message_update.assistantMessageEvent` for `text_delta`, `thinking_delta`, and `toolcall_delta`\n- streaming edit guard inspects `toolcall_delta`/`toolcall_end` on `edit` calls and can abort early\n- persistence writes finalized messages at `message_end`\n- auto-retry examines assistant `stopReason === \"error\"` plus `errorMessage` heuristics\n\n## Unified vs provider-specific responsibilities\n\nUnified (common contract):\n\n- event shape (`AssistantMessageEvent`)\n- final result extraction (`done`/`error`)\n- delta throttling + merge rules\n- agent/session event propagation model\n\nProvider-specific (not fully abstracted):\n\n- upstream event taxonomies and mapping logic\n- stop-reason translation tables\n- tool-call ID conventions\n- reasoning/thinking block semantics and signatures\n- usage token semantics and availability timing\n- message conversion constraints per API\n\n## Implementation files\n\n- [`../../ai/src/stream.ts`](../packages/ai/src/stream.ts) — provider dispatch, option mapping, API key/session plumbing, custom API dispatch, and provider-specific credential handling.\n- [`../../ai/src/utils/event-stream.ts`](../packages/ai/src/utils/event-stream.ts) — generic stream queue + assistant delta throttling.\n- [`../../ai/src/utils/json-parse.ts`](../packages/ai/src/utils/json-parse.ts) — partial JSON parsing for streamed tool arguments.\n- [`../../ai/src/providers/anthropic.ts`](../packages/ai/src/providers/anthropic.ts) — Anthropic event translation and tool JSON delta accumulation.\n- [`../../ai/src/providers/openai-responses.ts`](../packages/ai/src/providers/openai-responses.ts), [`openai-code-responses.ts`](../packages/ai/src/providers/openai-code-responses.ts), [`azure-openai-responses.ts`](../packages/ai/src/providers/azure-openai-responses.ts) — Responses-family event translation and status mapping.\n- [`../../ai/src/providers/google.ts`](../packages/ai/src/providers/google.ts), [`google-gemini-cli.ts`](../packages/ai/src/providers/google-gemini-cli.ts), [`google-vertex.ts`](../packages/ai/src/providers/google-vertex.ts) — Gemini stream chunk-to-block translation variants.\n- [`../../ai/src/providers/google-shared.ts`](../packages/ai/src/providers/google-shared.ts) — Gemini finish-reason mapping and shared conversion rules.\n- [`../../ai/src/providers/amazon-bedrock.ts`](../packages/ai/src/providers/amazon-bedrock.ts), [`openai-completions.ts`](../packages/ai/src/providers/openai-completions.ts), [`ollama.ts`](../packages/ai/src/providers/ollama.ts), [`cursor.ts`](../packages/ai/src/providers/cursor.ts) — additional built-in stream adapters using the same event contract.\n- [`../../agent/src/agent-loop.ts`](../packages/agent/src/agent-loop.ts) — provider stream consumption and `message_update` bridging.\n- [`../src/session/agent-session.ts`](../packages/coding-agent/src/session/agent-session.ts) — session-level handling of streaming updates, abort, retry, and persistence.\n",
 	"python-repl.md": "# Eval Tool Python Backend\n\nThis document describes the Python execution stack in `packages/coding-agent`.\nIt covers tool behavior, runner lifecycle, environment handling, execution semantics, output rendering, supported magics, and operational failure modes.\n\n## Scope and Key Files\n\n- Tool surface: `src/tools/eval.ts`\n- Session/per-call kernel orchestration: `src/eval/py/executor.ts`\n- Subprocess kernel client: `src/eval/py/kernel.ts`\n- Python wrapper / NDJSON server: `src/eval/py/runner.py`\n- Prelude helpers loaded into every kernel: `src/eval/py/prelude.py`\n- MIME bundle renderer (text + structured outputs): `src/eval/py/display.ts`\n- Interactive-mode renderer for user-triggered Python runs: `src/modes/components/eval-execution.ts`\n- Runtime/env filtering and Python resolution: `src/eval/py/runtime.ts`\n\n## What eval's Python backend is\n\nThe `eval` tool executes one or more Python cells inside a long-lived `python3` subprocess that speaks NDJSON over stdin/stdout. No Jupyter, no kernel gateway, no extra pip dependencies — a vanilla Python 3.8+ interpreter is enough. Rich `display()` output (PIL, pandas, plotly, matplotlib figures) keeps working because the wrapper reimplements the MIME-bundle dispatch that IPython previously provided.\n\nTool params:\n\n```ts\n{\n  cells: Array<{ code: string; title?: string }>;\n  timeout?: number; // seconds, clamped to 1..600, default 30\n  reset?: boolean; // reset selected runtime before the first cell only\n}\n```\n\nThe tool is `concurrency = \"exclusive\"` for a session, so calls do not overlap.\n\n## Kernel lifecycle\n\nEach kernel is a single Python subprocess: `python -u <runner.py>`. The bundled runner is materialized once per GJC process in a process-private temporary directory and file, then reused only by subsequent spawns within that process.\n\nKernel startup sequence:\n\n1. Availability check (`checkPythonKernelAvailability`) — verifies that a Python interpreter resolves and runs.\n2. Spawn `python -u runner.py` with filtered env and `cwd`.\n3. Send an init request that runs `os.chdir(cwd)`, injects env entries, and adds `cwd` to `sys.path`.\n4. Execute `PYTHON_PRELUDE` (idempotent — only initializes once per process).\n\nKernel shutdown:\n\n- Send `{\"type\": \"exit\"}` over stdin.\n- Wait for process exit with `SHUTDOWN_GRACE_MS` budget.\n- Escalate to `SIGTERM` and finally `SIGKILL` if the process does not exit in time.\n\n## Wire protocol (NDJSON, host ↔ runner)\n\nOne JSON object per line, UTF-8, `\\n` terminated.\n\nHost → runner:\n\n```jsonc\n{\"id\": \"<reqId>\", \"code\": \"<source>\", \"silent\": false, \"storeHistory\": true}\n{\"type\": \"exit\"}\n```\n\nRunner → host:\n\n```jsonc\n{\"type\": \"started\",  \"id\": \"<reqId>\"}\n{\"type\": \"stdout\",   \"id\": \"<reqId>\", \"data\": \"...\"}\n{\"type\": \"stderr\",   \"id\": \"<reqId>\", \"data\": \"...\"}\n{\"type\": \"display\",  \"id\": \"<reqId>\", \"bundle\": {<mime>: <value>}}\n{\"type\": \"result\",   \"id\": \"<reqId>\", \"bundle\": {<mime>: <value>}}\n{\"type\": \"error\",    \"id\": \"<reqId>\", \"ename\": \"...\", \"evalue\": \"...\", \"traceback\": [\"...\"]}\n{\"type\": \"done\",     \"id\": \"<reqId>\", \"status\": \"ok\"|\"error\", \"executionCount\": N, \"cancelled\": false}\n```\n\nStatus events the prelude emits (e.g. `_emit_status(\"find\", count=…)`) ship inside display bundles under `application/x-gjc-status` so the existing TUI status renderer keeps working.\n\n## Magics\n\nThe runner's source transformer rewrites IPython-style magics to plain Python calls before parsing. Supported set:\n\n| Magic | Effect |\n| --- | --- |\n| `%pip <args>` | `python -m pip <args>` with live streaming output. Newly installed packages are evicted from `sys.modules` so the next `import` picks up the fresh install. |\n| `%cd <path>` | `os.chdir(path)` (with `~` expansion); emits status event. |\n| `%pwd` | Returns `os.getcwd()`. |\n| `%ls [path]` | Returns `sorted(os.listdir(path))`. |\n| `%env [KEY[=VAL]]` | List, read, or set env vars (matches prelude `env()` semantics). |\n| `%set_env KEY VALUE` | Set `os.environ[KEY]`. |\n| `%time <expr>` / `%timeit <expr>` | Time the expression; emits status event with elapsed ms. |\n| `%who` / `%whos` | List user-namespace names. |\n| `%reset` | Clear user globals and re-inject prelude. |\n| `%load <path>` | Read a file into a fresh cell and execute. |\n| `%run <path>` | `runpy.run_path` and merge globals back. |\n| `%%bash` / `%%sh` | Run the cell body via `bash`/`sh`. |\n| `%%capture [name]` | Run body with stdout/stderr captured into `name`. |\n| `%%timeit` | Time the cell body. |\n| `%%writefile <path>` | Write body to file. |\n| `!cmd` / `var = !cmd` | Run command via subprocess shell; returns an SList-style result with `.n` / `.s` helpers. |\n| `var = %name args` | Assignment forms work for line magics and `!cmd`. |\n\nUnknown magic names raise `NameError: UsageError: ...` inside the cell.\n\n## Session persistence semantics\n\n`python.kernelMode` controls retained kernel reuse:\n\n- `session` (default)\n  - Reuses kernel sessions keyed by session file plus cwd when a session file exists; otherwise by cwd.\n  - Execution is serialized per session via a queue.\n  - Idle sessions are evicted after 5 minutes.\n  - At most 4 sessions; oldest is evicted on overflow.\n  - Heartbeat checks detect dead kernels.\n  - Auto-restart allowed once; repeated crash ⇒ hard failure.\n- `per-call`\n  - Spawns a fresh subprocess for each request.\n  - Shuts the subprocess down after the request.\n  - No cross-call state persistence.\n\n### Multi-cell behavior in a single tool call\n\nCells run sequentially in the same kernel instance for that tool call.\n\nIf an intermediate cell fails:\n\n- Earlier cell state remains in memory.\n- Tool returns a targeted error indicating which cell failed.\n- Later cells are not executed.\n\n`reset=true` only applies to the first cell execution in that call.\n\n## Environment filtering and runtime resolution\n\nEnvironment is filtered before launching the runner:\n\n- Allowlist includes core vars like `PATH`, `HOME`, locale vars, `VIRTUAL_ENV`, `PYTHONPATH`, etc.\n- Allow-prefixes: `LC_`, `XDG_`, `GJC_`\n- Denylist strips common API keys (OpenAI/Anthropic/Gemini/etc.)\n\nRuntime selection order:\n\n1. Active/located venv (`VIRTUAL_ENV`, then `<cwd>/.venv`, `<cwd>/venv`)\n2. Managed venv at `~/.gjc/python-env`\n3. `python` or `python3` on PATH\n\nWhen a venv is selected, its bin/Scripts path is prepended to `PATH`.\n\nThe runner additionally receives `PYTHONUNBUFFERED=1` and `PYTHONIOENCODING=utf-8` so streamed output reaches the host promptly.\n\n## Tool availability and mode selection\n\n`eval.py` / `eval.js` (both default `true`) plus optional `GJC_PY` override controls eval backend exposure:\n\n- Python backend only (`eval.py=true`, `eval.js=false`)\n- JavaScript backend only (`eval.py=false`, `eval.js=true`)\n- both backends\n\n`GJC_PY` accepted values:\n\n- `0` / `bash` → JavaScript backend only\n- `1` / `py` → Python backend only\n- `mix` / `both` → both backends\n\nIf Python preflight fails and `eval.js` is enabled, `eval` remains available and dispatches to JavaScript unless `language: \"python\"` is explicitly requested.\n\n## Execution flow and cancellation/timeout\n\n### Tool-level timeout\n\n`eval` timeout is in seconds, default 30, clamped to `1..600`. The tool combines caller abort signal and timeout signal with `AbortSignal.any(...)`.\n\n### Kernel execution cancellation\n\nOn abort/timeout:\n\n- The host sends `kill(\"SIGINT\")` to the runner subprocess.\n- The runner's exec-time signal handler raises `KeyboardInterrupt` inside the user code.\n- Result includes `cancelled=true`; timeout path annotates output as `Command timed out after <n> seconds`.\n- Between requests the runner installs `SIG_IGN` for SIGINT so a stray cancel does not tear down the kernel.\n\nIf a second cancel is required (runner stuck in C code), the host escalates to `SIGTERM` and the session restarts on the next call.\n\n### stdin behavior\n\nInteractive stdin is not supported. The runner does not forward `input()` prompts; user code that calls `input()` blocks until cancellation.\n\n## Output capture and rendering\n\n### Captured output classes\n\nFrom runner frames:\n\n- `stdout` / `stderr` → plain text chunks\n- `display` / `result` → rich display handling (MIME bundle)\n- `error` → traceback text\n- `application/x-gjc-status` MIME inside `display` → structured status events\n\nDisplay MIME precedence:\n\n1. `text/markdown`\n2. `text/plain`\n3. `text/html` (converted to basic markdown)\n\nAdditionally captured as structured outputs:\n\n- `application/json` → JSON tree data\n- `image/png` / `image/jpeg` → image payloads\n- `application/x-gjc-status` → status events\n\n### Matplotlib\n\nThe runner sets `MPLBACKEND=Agg` as an environ default so figures render off-screen. After every cell, `pyplot.get_fignums()` is iterated; each figure is saved to PNG, emitted as an `image/png` display, and closed.\n\n### Storage and truncation\n\nOutput is streamed through `OutputSink` and may be persisted to artifact storage. Tool results can include truncation metadata and `artifact://<id>` for full output recovery.\n\n### Renderer behavior\n\n- Tool renderer (`eval.ts`):\n  - shows code-cell blocks with per-cell status\n  - collapsed preview defaults to 10 lines\n  - supports expanded mode for full output and richer status detail\n- Interactive renderer (`eval-execution.ts`):\n  - used for user-triggered Python execution in TUI\n  - collapsed preview defaults to 20 lines\n  - clamps very long individual lines to 4000 chars for display safety\n  - shows cancellation/error/truncation notices\n\n## Operational troubleshooting\n\n- **Python backend not available** — Check `eval.py`, `GJC_PY`, and that `python`/`python3` is on PATH. If preflight fails and `eval.js` is enabled, omit `language` or pass `language: \"js\"` to use JavaScript.\n- **No Python on PATH** — Install a system Python 3.8+ or place a venv at `~/.gjc/python-env`. `gjc setup python --check` reports the resolved interpreter.\n- **Execution hangs then times out** — Increase tool `timeout` (max 600s) if workload is legitimate. For stuck native code, cancellation triggers `SIGINT` first then escalates; the session restarts on the next request.\n- **stdin/input prompts in Python code** — `input()` is not supported; pass data programmatically.\n- **Working directory errors** — Tool validates `cwd` exists and is a directory before execution.\n\n## Relevant environment variables\n\n- `GJC_PY` — tool exposure override\n- `GJC_PYTHON_SKIP_CHECK=1` — bypass Python preflight/warm checks\n- `GJC_PYTHON_INTEGRATION=1` — enable gated integration tests that spawn a real Python\n- `GJC_PYTHON_IPC_TRACE=1` — log NDJSON frames exchanged with the runner subprocess\n",
+	"release-0.13.3-integration.md": "# Release 0.13.3 integration record\n\n## Authority and immutable base\n\n- Integration branch: `release/0.13.3`\n- Authorized repository: `Yeachan-Heo/gajae-code`\n- Exact base: `origin/main` at `5666472818b71a1c37615408d9b4d3b5a77b7fa3`\n- Delivery mode: coherent cherry-pick groups pushed directly to the existing release branch\n- Explicitly not authorized here: version bump, `main` mutation, PR creation, tag, publish, or release execution\n\n## Included scope and dependency decisions\n\n### Original hotfixes\n\n| Candidate | Decision | Dependency boundary |\n| --- | --- | --- |\n| `0bd4898d` / #4437 | Include | Isolated Rust text wrapping termination fix and regression tests. No provider, auth, SDK lifecycle, or settings dependency. |\n| `38f3b407` / supplied #4509 anchor (commit subject references #4481) | Include | Bounded TUI overlay geometry and dedicated capture regressions. Changes stay in TUI rendering plus diagnostic scripts. |\n| `b5ef23f5` / #4446 | Include | Print-mode process termination and postmortem draining. Embedding and ACP ownership are explicitly preserved by the change. |\n\n### Bounded utilities\n\n| Candidate | Decision | Dependency boundary |\n| --- | --- | --- |\n| `6080b983` / #4453 | Include | Narrow `todo_write` positional-handle diagnostics and tests. No routing or model-contract change. |\n| `8eb8127e` / #4424 | Include | Narrow terminal image scrollback preservation. Shares `tui.ts` with the overlay hotfix, so it is integrated before the geometry hardening and tested together. |\n| `c06f4f5` / #4451 | Skip | Candidate directly modifies `src/autoresearch/dashboard.ts`; autoresearch is explicitly excluded. Its render-cache work is not cherry-picked partially because the commit couples the cache contract to excluded autoresearch and several controller/component migrations. |\n\n### Session, storage, retry, and crash resilience\n\n| Candidate | Decision | Dependency boundary |\n| --- | --- | --- |\n| `c81614fc` / #4396 | Include | Managed-output publication/reaping plus the minimum native path-identity binding needed by that storage implementation. No SDK notification/lifecycle architecture change. |\n| `8d6784cb` / #4411 | Include | Managed transcript size proactive/reactive recovery and compaction signal. Changes are limited to agent compaction and coding-agent session persistence. |\n| `515f1c7c` / #4421 | Include | Missing managed transcript recovery in session manager. |\n| `ffa07d6c` / #4450 | Include | Missing predecessor recovery in managed session storage. Applied after #4396 because both touch the same storage implementation and #4450 is the later semantic correction. |\n| `0a736b17` / #4470 | Include | Crash-index compaction/read compatibility fix. Applied before #4495 so the later broader recovery work retains this invariant. |\n| `7a6b0d13` / #4495 | Include | Crash journal/index recovery hardening. Scoped to crash persistence and record loading. |\n\n### Pet features\n\n| Candidate | Decision | Dependency boundary |\n| --- | --- | --- |\n| `d2f6e8a4` / #4468 | Include with strict pet-only settings allowance | Ouroboros pet/theme requires a bounded settings schema enum, pet/theme selectors, theme registration, and UI verification updates. These are accepted only as the minimum cohesive dependency of the pet feature; no general settings migration, MCP, customization migration, provider, model, or auth contract is admitted. |\n| `dfb1081c` / #4499 | Include | iTerm2 pet rendering is applied after #4468 because both update the pet widget and TUI pet renderer. It does not alter auth/provider/model or SDK lifecycle contracts. |\n\n## Owner-authorized expansion after the original audit\n\nThe owner subsequently authorized a bounded provider/preset and stability expansion on the existing release branch. The following exact changes were added without importing the broader `dev` architecture:\n\n### Stability and security\n\n- #4385 Synthetic login validation: use the provider models endpoint instead of a retired hard-coded Kimi probe (`6de4096025`, changelog `e14189ec82`).\n- #4302 image-generation provider error hardening: redact credentials and bound response bodies (`08e3e4de22`, `b3ce4bb584`).\n- #4452 status-line scratch-root resolution at render time (`d27175a1c3`).\n- #4373 Windows replacement-receipt reconciliation over the release branch's existing managed-session storage fixes (`bc9417f4eb`).\n\n### Providers and presets\n\n- #4304 native TypeScript Kiro OAuth/CodeWhisperer provider (`7913ce221a`).\n- #4294 Muse Spark 1.2 open-weight catalog, profile, reasoning, cache-normalization, and routing closure (`6ff17fe082` through `efd850df86`).\n- #4377 direct xAI Grok 4.6 catalog and role profiles (`cb3eba7a93`, `8c7170a212`).\n- #4462 bundled Grok CLI 4.6 catalog and bounded effort parsing (`0ce9f10436`, `5a931cc529`, `4b15d708bf`).\n\nCHANGELOG conflicts were resolved by retaining the released 0.13.2 history and existing release entries while adding only the new Unreleased provider/preset records. Generated model data was merged with its source generators and must pass reproducibility tests before the release cut.\n\n## Explicit exclusions\n\n- General settings, MCP, or customization migrations beyond the minimum cohesive pet dependencies listed above\n- SDK lifecycle or notification architecture\n- GJC master/supervisor work\n- Autorouting\n- Autoresearch\n- macOS TCC/signing change #4275 pending dedicated release-build and permission-identity validation\n- Version bump, release metadata cut, tag, publish, release execution, PR creation, or `main` mutation\n\n## Integration order\n\n1. Record this scope and boundary decision.\n2. Original hotfixes and bounded TUI utilities: #4437, #4424, supplied #4509 anchor/#4481 commit, #4446, #4453.\n3. Session/storage resilience: #4396, #4411, #4421, #4450.\n4. Crash recovery resilience: #4470, #4495.\n5. Pet: #4468, then #4499.\n6. Run focused tests after each coherent group, followed by branch-wide checks, build, test, and install/package smoke.\n\n## Validation evidence\n\nThis section is updated as groups are integrated. Exact commands, outcomes, skips, and environmental blockers are recorded before final delivery.\n",
 	"render-mermaid.md": "# RenderMermaid\n\n`RenderMermaid` is an optional built-in tool that renders Mermaid source to terminal-friendly text.\n\n## Enable it\n\nDisabled by default. Turn it on in `/settings` under **Tools → Render Mermaid**, or in `~/.gjc/agent/config.yml`:\n\n```yaml\nrenderMermaid:\n  enabled: true\n```\n\n## What it does\n\n- Tool name: `render_mermaid`\n- Input: Mermaid source in the required `mermaid` field\n- Output: rendered ASCII/Unicode text, not SVG or PNG\n- Storage: when artifact storage is available, the full render is also saved as an `artifact://...`\n\nThere are no model-specific or environment-variable prerequisites. Once enabled, any model that can call built-in tools can use it.\n\n## Parameters\n\n```json\n{\n  \"mermaid\": \"graph TD\\n  A[Start] --> B[Stop]\",\n  \"config\": {\n    \"useAscii\": false,\n    \"paddingX\": 2,\n    \"paddingY\": 2,\n    \"boxBorderPadding\": 0\n  }\n}\n```\n\nAvailable `config` fields:\n\n- `useAscii` — `true` for plain ASCII, `false` for Unicode box-drawing characters (default and usually more readable)\n- `paddingX` — horizontal spacing between nodes\n- `paddingY` — vertical spacing between nodes\n- `boxBorderPadding` — inner padding inside node boxes\n\n## Current limitations\n\n`RenderMermaid` uses the `beautiful-mermaid` ASCII renderer. It works best for flowcharts and small diagrams.\n\nComplex sequence diagrams, especially with `alt` / `else` blocks, can become very wide in a terminal. That is current renderer behavior, not a provider or model configuration problem.\n\nIf a sequence diagram is hard to read:\n\n1. Keep Unicode output (`useAscii: false`)\n2. Reduce spacing with a tighter config such as `paddingX: 2`, `paddingY: 2`, `boxBorderPadding: 0`\n3. Prefer smaller sub-diagrams over one large sequence diagram\n4. Open the saved artifact if the inline preview is truncated in the TUI\n\n## Example\n\nInput:\n\n```mermaid\ngraph TD\n  A[Start] --> B{Decision}\n  B -->|Yes| C[Action]\n  B -->|No| D[End]\n```\n\nTypical result:\n\n```text\n┌─────┐\n│Start│\n└─────┘\n   │\n   ▼\n┌────────┐\n│Decision│\n└────────┘\n```\n",
 	"research-plan-ledger.md": "# Research plan items and evidence ledger\n\nResearch/deep-research workflows need a planning contract that is stronger than an execution-order checklist. A plan item should name the claim under investigation, the uncertainty around it, what evidence is required, what counterexamples would falsify it, and how a verifier should handle source conflicts.\n\nThis document defines the public product-facing spike for issue #932. It intentionally avoids private operator, session, channel, and routing internals.\n\n## Research plan item schema\n\n```ts\ntype ResearchPlanConfidence = \"low\" | \"medium\" | \"high\";\n\ntype ResearchPlanItem = {\n  claim: string;\n  confidence: ResearchPlanConfidence;\n  unknowns: string[];\n  evidenceNeeded: string[];\n  counterexampleQueries: string[];\n  sourceConflictPolicy: string;\n  dropCondition: string;\n  verifierChecks: string[];\n};\n```\n\nField intent:\n\n- `claim`: The smallest claim that can survive or fail verification.\n- `confidence`: Planner's initial confidence before evidence collection.\n- `unknowns`: Known gaps the final answer must resolve or explicitly carry forward.\n- `evidenceNeeded`: Evidence workers must collect before the claim can be accepted.\n- `counterexampleQueries`: Directed search prompts for evidence that would weaken or falsify the claim.\n- `sourceConflictPolicy`: How the verifier treats conflicting sources, stale sources, or mismatched methodology.\n- `dropCondition`: The explicit condition that removes this claim from the final answer.\n- `verifierChecks`: Checklist the verifier applies before accepting the claim.\n\n## Evidence ledger schema\n\n```ts\ntype ResearchEvidenceVerdict = \"support\" | \"contradict\" | \"uncertain\";\n\ntype ResearchEvidenceEntry = {\n  claim: string;\n  source: string;\n  confidence: ResearchPlanConfidence;\n  verdict: ResearchEvidenceVerdict;\n  notes?: string;\n};\n\ntype ResearchLedgerVerdict = {\n  claim: string;\n  finalVerdict: \"accepted\" | \"rejected\" | \"uncertain\";\n  survivingSources: ResearchEvidenceEntry[];\n  rejectReason?: string;\n  unresolvedUnknowns: string[];\n};\n```\n\nThe ledger is claim-centric. Workers add evidence entries against plan-item claims; the verifier reduces those entries into a final verdict. Accepted claims can be cited in the final answer. Rejected claims are named with `rejectReason`. Uncertain claims are either excluded or marked explicitly as unresolved.\n\n## Ralplan/research workflow shape\n\n1. Planner emits `ResearchPlanItem[]` alongside the normal plan narrative when the task is research-heavy.\n2. Workers gather independent evidence for each item, including counterexample-oriented searches.\n3. Verifier checks contradictions, source quality, stale information, and unresolved uncertainty using the item's `verifierChecks`, `sourceConflictPolicy`, and `dropCondition`.\n4. The final answer cites accepted claims, lists rejected claims with reasons, and marks any surviving uncertainty.\n\n## Example\n\n```ts\nconst item: ResearchPlanItem = {\n  claim: \"Model X reduces latency by 30% on production-like workloads\",\n  confidence: \"medium\",\n  unknowns: [\"production workload mix\"],\n  evidenceNeeded: [\"benchmark with production-like fixture\", \"baseline comparison\"],\n  counterexampleQueries: [\"regression on long-context workload\", \"cold-start latency increase\"],\n  sourceConflictPolicy: \"Reject the claim when any credible counterexample contradicts the benchmark.\",\n  dropCondition: \"Drop if a counterexample contradicts the claim or key unknowns remain unresolved.\",\n  verifierChecks: [\"check source freshness\", \"compare benchmark harness\", \"inspect counterexample evidence\"],\n};\n```\n\nIf the ledger contains a supporting benchmark and a credible long-context counterexample, the verifier rejects the broad claim instead of letting a plausible summary survive by vibes.\n\n## Current spike\n\nThe first implementation spike lives in `packages/coding-agent/src/research-plan/ledger.ts` and provides:\n\n- TypeScript interfaces for research plan items, evidence entries, and final verdicts.\n- Validators for product-facing plan/evidence objects.\n- A deterministic verifier helper that rejects plausible claims when counterexample/source-conflict/drop-condition evidence applies.\n- Regression tests in `packages/coding-agent/test/research-plan-ledger.test.ts`.\n\nFuture runtime integration can make `/skill:ralplan` emit these structures as a fenced JSON block or structured sidecar in the persisted ralplan artifact. The spike keeps the schema independent from private session state so it can be exposed in docs and tests safely.\n",
 	"resolve-tool-runtime.md": "# Resolve tool runtime internals\n\nThis document explains how preview/apply workflows are modeled in coding-agent and how built-in or custom tools can participate via the tool-choice queue and `pushPendingAction`.\n\n## Scope and key files\n\n- [`src/tools/resolve.ts`](../packages/coding-agent/src/tools/resolve.ts)\n- [`src/tools/ast-edit.ts`](../packages/coding-agent/src/tools/ast-edit.ts)\n- [`src/extensibility/custom-tools/types.ts`](../packages/coding-agent/src/extensibility/custom-tools/types.ts)\n- [`src/extensibility/custom-tools/loader.ts`](../packages/coding-agent/src/extensibility/custom-tools/loader.ts)\n- [`src/sdk/session.ts`](../packages/coding-agent/src/sdk/session.ts)\n\n## What `resolve` does\n\n`resolve` is a hidden tool that finalizes a pending preview action.\n\n- `action: \"apply\"` executes the queued action's `apply(reason)` callback and returns that result with resolve metadata.\n- `action: \"discard\"` invokes `reject(reason)` if provided; otherwise returns `Discarded: <label>. Reason: <reason>`.\n\nIf no pending action exists, `resolve` fails with:\n\n- `No pending action to resolve. Nothing to apply or discard.`\n\n## Pending actions use the tool-choice queue\n\nPreview producers call `queueResolveHandler(...)`, which pushes a one-shot forced `resolve` directive onto the session tool-choice queue and adds a `resolve-reminder` steering message.\n\nRuntime behavior:\n\n- the queued handler owns the pending `apply`/`reject` callbacks,\n- `resolve` looks up the current queue invoker with `session.peekQueueInvoker()`,\n- if the model rejects the forced tool choice, the queue directive is requeued,\n- `resolve` does not maintain a separate pending-action stack.\n\nMultiple pending previews therefore follow the active tool-choice queue ordering, not an independent pending-action store.\n\n## Built-in producer example (`ast_edit`)\n\n`ast_edit` previews structural replacements first. When the preview has replacements and is not applied yet, it queues a resolve handler that contains:\n\n- label (human-readable summary)\n- `sourceToolName` (`ast_edit`)\n- `apply(reason: string)` callback that reruns AST edit with `dryRun: false`\n\n`resolve(action=\"apply\", reason=\"...\")` passes `reason` into this callback.\n\n## Custom tools: `pushPendingAction`\n\nCustom tools can register resolve-compatible pending actions through `CustomToolAPI.pushPendingAction(...)`. The custom tool loader forwards these actions to `queueResolveHandler(...)` when that hook is available.\n\n`CustomToolPendingAction`:\n\n- `label: string` (required)\n- `apply(reason: string): Promise<AgentToolResult<unknown>>` (required) — invoked on apply; `reason` is the string passed to `resolve`\n- `reject?(reason: string): Promise<AgentToolResult<unknown> | undefined>` (optional) — invoked on discard; return value replaces the default \"Discarded\" message if provided\n- `details?: unknown` exists on the public custom-tool type but is not currently forwarded by the loader into resolve metadata\n- `sourceToolName?: string` (optional, defaults to `\"custom_tool\"`)\n\n### Minimal usage example\n\n```ts\nimport type { CustomToolFactory } from \"@gajae-code/coding-agent\";\n\nconst factory: CustomToolFactory = (pi) => ({\n  name: \"batch_rename_preview\",\n  label: \"Batch Rename Preview\",\n  description: \"Previews renames and defers commit to resolve\",\n  parameters: pi.zod.object({\n    files: pi.zod.array(pi.zod.string()),\n  }),\n\n  async execute(_toolCallId, params) {\n    const previewSummary = `Prepared rename plan for ${params.files.length} files`;\n\n    pi.pushPendingAction({\n      label: `Batch rename: ${params.files.length} files`,\n      sourceToolName: \"batch_rename_preview\",\n      apply: async (reason) => {\n        // apply writes here\n        return {\n          content: [\n            { type: \"text\", text: `Applied batch rename. Reason: ${reason}` },\n          ],\n        };\n      },\n      reject: async (reason) => {\n        // optional: cleanup or notify on discard\n        return {\n          content: [\n            { type: \"text\", text: `Discarded batch rename. Reason: ${reason}` },\n          ],\n        };\n      },\n    });\n\n    return {\n      content: [\n        {\n          type: \"text\",\n          text: `${previewSummary}. Call resolve to apply or discard.`,\n        },\n      ],\n    };\n  },\n});\n\nexport default factory;\n```\n\n## Runtime availability and failures\n\n`pushPendingAction` is wired by the custom tool loader through the active session's resolve queue hook.\n\nIf the runtime did not provide the resolve queue hook, `pushPendingAction` throws:\n\n- `Pending action store unavailable for custom tools in this runtime.`\n\n## Tool-choice behavior\n\nWhen `queueResolveHandler(...)` registers a preview, the agent runtime forces a one-shot `resolve` tool choice so pending previews are explicitly finalized before normal tool flow continues.\n\n## Developer guidance\n\n- Use pending actions only for destructive or high-impact operations that should support explicit apply/discard.\n- Keep `label` concise and specific; it is shown in resolve renderer output.\n- Ensure `apply(reason)` is deterministic and idempotent enough for one-shot execution; `reason` is informational and should not change behavior.\n- Implement `reject(reason)` when the discard needs cleanup (temp state, locks, notifications); omit it for stateless previews where the default message suffices.\n- If your tool can stage multiple previews, remember they are mediated by the tool-choice queue rather than a separate pending-action stack.\n",

package/src/main.ts

@@ -1255,6 +1255,7 @@ export interface RunRootCommandDependencies {
 	getChangelogForDisplay?: typeof getChangelogForDisplay;
 	createInteractiveMode?: CreateInteractiveMode;
 	runPrintMode?: RunPrintMode;
+	quit?: (code: number) => Promise<void>;
 	isResumePickerTerminal?: ResumePickerTerminalCheck;
 	listForResumePickerReadOnly?: ListForResumePickerReadOnly;
 	listManagedForResumePickerReadOnly?: ListManagedForResumePickerReadOnly;
@@ -1961,6 +1962,9 @@ export async function runRootCommand(
 				authStorage.close();
 				if (!deps.suppressProcessExit) await postmortem.cleanup();
 			}
+			if (!deps.suppressProcessExit) {
+				await (deps.quit ?? postmortem.quit)(typeof process.exitCode === "number" ? process.exitCode : 0);
+			}
 		}
 	}
 }

package/src/modes/components/chrome.ts

@@ -1,6 +1,7 @@
-import { Container, type SelectItem, SelectList, Text } from "@gajae-code/tui";
+import { Container, type SelectItem, SelectList } from "@gajae-code/tui";
 import { getSelectListTheme, theme } from "../../modes/theme/theme";
 import { DynamicBorder } from "./dynamic-border";
+import { DynamicThemeText } from "./dynamic-theme-text";
 
 export interface FramedSelectOptions {
 	maxVisible?: number;
@@ -24,10 +25,10 @@ export function FramedSelect(
 	const container = new Container();
 	container.addChild(new DynamicBorder());
 	if (title) {
-		container.addChild(new Text(theme.bold(theme.fg("accent", title)), 0, 0));
+		container.addChild(new DynamicThemeText(() => theme.bold(theme.fg("accent", title))));
 	}
 
-	const selectList = new SelectList(items, opts.maxVisible ?? Math.min(items.length, 10), getSelectListTheme());
+	const selectList = new SelectList(items, opts.maxVisible ?? Math.min(items.length, 10), () => getSelectListTheme());
 	const selectedIndex =
 		opts.selectedValue === undefined ? -1 : items.findIndex(item => item.value === opts.selectedValue);
 	if (selectedIndex !== -1) selectList.setSelectedIndex(selectedIndex);