Prepare gbrain keyless-mode fix for garrytan/gbrain#94 (--mode claude-code)

work/gbrain-no-api-keys/ISSUE-codex-kimi.md

@@ -0,0 +1,36 @@
+# Ready-to-paste issue for https://github.com/garrytan/gbrain/issues/new
+
+**Title:**
+
+codex-cli and kimi-cli recipes: subscription-native keyless dispatch for Codex and Kimi users (follow-up to #94)
+
+**Body:**
+
+## Context
+
+#94 asked for a Claude Code-native mode with no API keys; PR #3464 implements it: `gbrain init --mode claude-code` routes chat, query expansion, and the subagent loop through the `claude-cli` recipe (Claude Code OAuth session, subscription-billed), and picks up a local Ollama daemon for keyless embeddings when one is running.
+
+The same friction exists for users of other agent CLIs with subscription auth. This issue proposes extending the pattern so **any** agent subscription can power gbrain keylessly:
+
+## Proposal
+
+**1. `codex-cli` recipe** — OpenAI Codex CLI users sign in with their ChatGPT subscription; `codex exec` provides a non-interactive mode suitable for subprocess dispatch. A `CodexCliLanguageModel` modeled on `src/core/ai/providers/claude-cli-language-model.ts` plus a pure-data recipe (`implementation: 'codex-cli'`, `auth_env.required: []`) would give Codex subscribers chat + expansion + subagent dispatch with zero API keys — the exact #334/#94 story for the other big agent population.
+
+**2. `kimi-cli` recipe** — Moonshot's Kimi CLI is the same shape: subscription-authenticated agent CLI with a non-interactive mode. Same provider pattern, same recipe contract. (gbrain already ships a hosted `moonshot` recipe; this would be its keyless sibling, the way `claude-cli` sits alongside `anthropic`.)
+
+**3. Possible generalization** — the two providers above will be ~90% identical to the claude-cli one (spawn binary, isolate cwd, parse JSON/text output, map tool calls). It may be worth extracting a shared `agent-cli` subprocess base where each recipe supplies binary name/env var, argv template, and an output parser — so the third agent CLI is a recipe, not a provider implementation.
+
+**4. `--mode claude-code` naming** — with more than one agent CLI supported, init could grow `--mode codex` / `--mode kimi` (or a generic `--mode agent-cli --agent <bin>`), all sharing the keyless machinery PR #3464 introduced (`claude_code_mode` sentinel, implicit `--no-embed` import, conservative search-mode seeding, Ollama embedding probe).
+
+## Why not in PR #3464
+
+Shipping subprocess providers untested would be worse than proposing them: verifying the `codex exec` / Kimi CLI output protocols needs machines with those binaries logged in. The claude-cli path in #3464 was verified end-to-end; these two deserve the same bar.
+
+## Questions for the maintainer
+
+1. Would PRs for `codex-cli` / `kimi-cli` recipes be welcome, following the claude-cli provider pattern?
+2. Preference between per-provider implementations vs. the shared `agent-cli` base (option 3)?
+3. Any constraints on adding init modes per agent CLI vs. a single generic flag?
+
+---
+_Generated by [Claude Code](https://claude.ai/code)_

work/gbrain-no-api-keys/PR-DESCRIPTION.md

@@ -0,0 +1,101 @@
+# Upstream PR description for garrytan/gbrain
+
+Ready to paste when opening the PR from a fork (apply
+`0001-claude-code-keyless-mode.patch` on a branch off `master`).
+
+---
+
+**Title:** feat(init): `--mode claude-code` — keyless mode for Claude Code subscribers (#94)
+
+**Body:**
+
+Closes #94.
+
+## What
+
+`gbrain init --mode claude-code` gives a Claude Code subscriber a fully working
+brain with **zero provider API keys** — no `OPENAI_API_KEY`, no
+`ANTHROPIC_API_KEY`:
+
+```bash
+gbrain init --mode claude-code   # keyless: PGLite, keyword + graph search
+gbrain import ~/notes/
+gbrain search "who works at Acme"
+```
+
+Search runs on the existing keyword + graph + title arms; chat commands route
+through the existing `claude-cli` recipe (Claude Code OAuth session, billed to
+the subscription).
+
+## Why this is a small diff
+
+Most of the machinery already existed — this PR wires the seams:
+
+- the `claude-cli` chat recipe (#334) already runs keyless and supports the
+  subagent loop;
+- hybrid search already degrades gracefully to keyword + graph + title when no
+  embedding provider is available (Codex C3 / D1 paths);
+- `gbrain sync` already honors the `embedding_disabled` sentinel.
+
+What was missing: `init` fail-louded without keys, `import` refused outright on
+a no-embedding brain, and chat defaulted to `anthropic:*`.
+
+## Changes
+
+- **init**: new `--mode claude-code` flag. Implies `--pglite`; writes
+  `claude_code_mode: true` alongside `embedding_disabled: true` (every existing
+  skip-embed callsite keeps working); defaults `chat_model` to
+  `claude-cli:claude-sonnet-4-6`; seeds `search.mode = conservative` (the one
+  bundle with no reranker / LLM-expansion spend); best-effort check that the
+  `claude` binary is on PATH. Re-inits stay keyless without re-passing the flag.
+- **import**: a claude-code brain imports without vectors (implicit
+  `--no-embed`) instead of exiting 1 with the deferred-setup error.
+- **embed**: refusal message tailored to keyless mode, with the vector upgrade
+  recipe.
+- **advisor**: `embedding_disabled` is not a setup smell when keyless by design.
+- **Bug fix**: explicit `--embedding-model` / `--model` now clears a persisted
+  deferred/keyless sentinel. Previously the documented upgrade path
+  (`gbrain init --force --embedding-model …` — the exact command
+  `assertEmbeddingEnabled` prints) was silently ignored because the seeded
+  `noEmbedding` won in `initPGLite`. A resolved model also removes stale
+  sentinels from `config.json` (the "one or the other, never both" invariant).
+- **Bug fix**: the post-init "subagent features require ANTHROPIC_API_KEY"
+  caveat no longer fires for `claude-cli:*` chat models — that recipe declares
+  `supports_subagent_loop` and drives Minions through the OAuth session.
+- Docs: README quick-start subsection, CHANGELOG, `init --help`, and the
+  fail-loud no-provider hint now advertises the keyless path.
+
+## Tests
+
+- New `test/init-claude-code-mode.test.ts` (12 tests): keyless predicate,
+  tailored refusal messages, re-init seeding, and the claude-cli recipe's
+  keyless contract (no required env vars, subagent loop supported).
+- All touched-area suites green (init-env-detection, init-embed-check,
+  init-mode-picker, init-provider-picker, sync-no-embed-sentinel, advisor-core,
+  import-file, embed-preflight, embedding-dim-check, config);
+  `bun run typecheck` clean.
+- End-to-end smoke with all provider keys unset:
+  init → import → keyword search returns the right page → `embed --stale`
+  refuses with the keyless message → `init --force --pglite --embedding-model
+  zeroentropyai:zembed-1 --embedding-dimensions 1280` clears the sentinels and
+  configures the model.
+
+## Out of scope (deliberate)
+
+- Query expansion / reranking via Claude Code skills (the issue's stretch
+  ideas): expansion is already gated behind `isAvailable('expansion')` and off
+  in the conservative bundle; the reranker fails open. Nothing breaks keylessly;
+  wiring them through `claude-cli` can be a follow-up.
+- Postgres/Supabase keyless installs: the flag targets the local PGLite path
+  per the issue; the runtime behavior (import/search) is engine-agnostic, so
+  extending later is trivial.
+
+## Unrelated heads-up found while working on this
+
+`master` (at least since faf5cdb) has an accidentally **committed
+`node_modules` symlink** pointing at `/tmp/fleet/repo/node_modules` — a path
+from someone's build environment. It's tracked despite `node_modules/` being in
+`.gitignore` (gitignore doesn't apply to already-tracked paths), so every fresh
+clone gets a dangling symlink and `bun install` behaves oddly until you
+`rm node_modules && bun install`. Worth a one-line cleanup commit:
+`git rm --cached node_modules`. Not fixed in this PR to keep it scoped to #94.

work/gbrain-no-api-keys/README.md

@@ -0,0 +1,87 @@
+# gbrain: Claude Code-native keyless mode (fix for garrytan/gbrain#94)
+
+Prepared fix for [garrytan/gbrain#94](https://github.com/garrytan/gbrain/issues/94) —
+"Plans for a Claude Code-native mode (no API keys, like gstack)?"
+
+The patch adds `gbrain init --mode claude-code`: a Claude Code subscriber gets a
+fully working brain with **zero provider API keys** (no `OPENAI_API_KEY`, no
+`ANTHROPIC_API_KEY`).
+
+## Applying
+
+```bash
+git clone https://github.com/garrytan/gbrain && cd gbrain
+git checkout ddd66e1   # base commit the patch was built against (master, 2026-07-27)
+git am path/to/0001-claude-code-keyless-mode.patch
+```
+
+## What was already there vs. what was missing
+
+gbrain already had most of the building blocks; the blockers were at the seams:
+
+| Piece | Status before the patch |
+|---|---|
+| `claude-cli` chat recipe (OAuth session, no key, subagent loop) | ✅ existed (their #334) |
+| Hybrid search degrades to keyword + graph + title when no embedding provider | ✅ existed (Codex C3 / D1 paths) |
+| `gbrain sync` skips embedding on the `embedding_disabled` sentinel | ✅ existed |
+| `gbrain init` without keys | ❌ exits 1 unless you know `--no-embedding` |
+| `gbrain import` on a no-embedding brain | ❌ refuses outright (deferred-setup error) |
+| Chat/expansion defaults | ❌ point at `anthropic:*`, which needs a key |
+| Upgrade path `init --force --embedding-model …` out of deferred mode | ❌ latent bug: persisted sentinel silently beat the explicit flag |
+
+## What the patch does
+
+- **`gbrain init --mode claude-code`** — implies `--pglite`; writes
+  `claude_code_mode: true` alongside `embedding_disabled: true` (so every
+  existing skip-embed callsite keeps working); defaults `chat_model` to
+  `claude-cli:claude-sonnet-4-6`; seeds `search.mode = conservative` (the only
+  mode bundle with no reranker / LLM-expansion spend); warns if the `claude`
+  binary isn't on PATH.
+- **`gbrain import`** treats a claude-code brain as implicit `--no-embed` —
+  chunks land keyword-searchable instead of the command exiting 1.
+- **`gbrain embed`** refuses with a keyless-specific message including the
+  vector upgrade recipe.
+- **advisor** no longer flags the missing embedding provider as a setup smell
+  on keyless brains.
+- **Bug fix**: explicit `--embedding-model` / `--model` now clears a persisted
+  deferred/keyless sentinel, making the documented upgrade path work; a
+  resolved model also removes stale sentinels from `config.json`.
+- **Bug fix**: the post-init "subagent features require ANTHROPIC_API_KEY"
+  caveat is suppressed for `claude-cli:*` chat models (that recipe drives the
+  subagent loop through the Claude Code OAuth session).
+- Docs: README quick-start section, CHANGELOG entry, `init --help` text, and
+  the fail-loud no-provider hint now advertises the keyless path.
+- Tests: `test/init-claude-code-mode.test.ts` (12 tests) covering the keyless
+  predicate, tailored refusal messages, re-init seeding, and the claude-cli
+  recipe's keyless contract (no required env vars, subagent loop supported).
+
+## Verification performed
+
+- `bun test` on the new suite plus all touched-area suites
+  (init-env-detection, init-embed-check, init-mode-picker, init-provider-picker,
+  sync-no-embed-sentinel, advisor-core, import-file, embed-preflight,
+  embedding-dim-check, config): **all green**.
+- `bun run typecheck` (tsc --noEmit): clean.
+- End-to-end smoke with **all provider keys unset**:
+  `init --mode claude-code` → config sentinels + conservative search mode
+  written → `import` of markdown succeeds without vectors → `search` returns
+  keyword-arm results → `embed --stale` refuses with the keyless message →
+  `init --force --pglite --embedding-model zeroentropyai:zembed-1
+  --embedding-dimensions 1280` clears the sentinels and configures the model.
+
+## Not in scope (deliberate)
+
+- Query expansion / reranking via Claude Code skills (the issue's stretch
+  ideas): expansion is already gracefully gated behind
+  `isAvailable('expansion')` and off in the conservative bundle; the reranker
+  fails open. Nothing breaks keylessly; wiring them through `claude-cli` can be
+  a follow-up.
+- Postgres/Supabase keyless installs: `--mode claude-code` targets the local
+  PGLite path per the issue; the runtime behavior (import/search) is
+  engine-agnostic, so extending the flag later is trivial.
+
+## Submitting upstream
+
+This session couldn't open a PR against `garrytan/gbrain` (repo outside the
+session's GitHub scope), so the change ships here as a `git am`-ready patch.
+Apply on a fork branch and open the PR referencing issue #94.

work/gbrain-no-api-keys/open-upstream-pr.sh

@@ -0,0 +1,23 @@
+#!/usr/bin/env bash
+# Opens the upstream PR garrytan/gbrain <- mnemonik-dev:claude-code-keyless-mode.
+#
+# Run from any machine with a GitHub token that can see garrytan/gbrain
+# (a classic PAT with `public_repo`, or a fine-grained token; the branch
+# already sits on the mnemonik-dev/gbrain fork):
+#
+#   GITHUB_TOKEN=ghp_... ./open-upstream-pr.sh
+#
+# Why this script exists: the Claude Code session that prepared the fix is
+# repo-scoped to mnemonik-dev/* — its outbound proxy 403s every GitHub API
+# call to garrytan/gbrain, and cross-owner add_repo is a v1 limitation.
+set -euo pipefail
+cd "$(dirname "$0")"
+
+: "${GITHUB_TOKEN:?set GITHUB_TOKEN to a token that can access garrytan/gbrain}"
+
+curl -sS -X POST \
+  -H "Authorization: Bearer ${GITHUB_TOKEN}" \
+  -H "Accept: application/vnd.github+json" \
+  -H "X-GitHub-Api-Version: 2022-11-28" \
+  https://api.github.com/repos/garrytan/gbrain/pulls \
+  -d @pr-payload.json | python3 -c 'import json,sys; d=json.load(sys.stdin); print(d.get("html_url") or json.dumps(d, indent=2))'

work/gbrain-no-api-keys/pr-payload.json

@@ -0,0 +1,7 @@
+{
+  "title": "feat(init): --mode claude-code — keyless mode for Claude Code subscribers (#94)",
+  "head": "mnemonik-dev:claude-code-keyless-mode",
+  "base": "master",
+  "maintainer_can_modify": true,
+  "body": "Closes #94.\n\n## What\n\n`gbrain init --mode claude-code` gives a Claude Code subscriber a fully working\nbrain with **zero provider API keys** — no `OPENAI_API_KEY`, no\n`ANTHROPIC_API_KEY`:\n\n```bash\ngbrain init --mode claude-code   # keyless: PGLite, keyword + graph search\ngbrain import ~/notes/\ngbrain search \"who works at Acme\"\n```\n\nSearch runs on the existing keyword + graph + title arms; chat commands route\nthrough the existing `claude-cli` recipe (Claude Code OAuth session, billed to\nthe subscription).\n\n## Why this is a small diff\n\nMost of the machinery already existed — this PR wires the seams:\n\n- the `claude-cli` chat recipe (#334) already runs keyless and supports the\n  subagent loop;\n- hybrid search already degrades gracefully to keyword + graph + title when no\n  embedding provider is available (Codex C3 / D1 paths);\n- `gbrain sync` already honors the `embedding_disabled` sentinel.\n\nWhat was missing: `init` fail-louded without keys, `import` refused outright on\na no-embedding brain, and chat defaulted to `anthropic:*`.\n\n## Changes\n\n- **init**: new `--mode claude-code` flag. Implies `--pglite`; writes\n  `claude_code_mode: true` alongside `embedding_disabled: true` (every existing\n  skip-embed callsite keeps working); defaults `chat_model` to\n  `claude-cli:claude-sonnet-4-6`; seeds `search.mode = conservative` (the one\n  bundle with no reranker / LLM-expansion spend); best-effort check that the\n  `claude` binary is on PATH. Re-inits stay keyless without re-passing the flag.\n- **import**: a claude-code brain imports without vectors (implicit\n  `--no-embed`) instead of exiting 1 with the deferred-setup error.\n- **embed**: refusal message tailored to keyless mode, with the vector upgrade\n  recipe.\n- **advisor**: `embedding_disabled` is not a setup smell when keyless by design.\n- **Bug fix**: explicit `--embedding-model` / `--model` now clears a persisted\n  deferred/keyless sentinel. Previously the documented upgrade path\n  (`gbrain init --force --embedding-model …` — the exact command\n  `assertEmbeddingEnabled` prints) was silently ignored because the seeded\n  `noEmbedding` won in `initPGLite`. A resolved model also removes stale\n  sentinels from `config.json` (the \"one or the other, never both\" invariant).\n- **Bug fix**: the post-init \"subagent features require ANTHROPIC_API_KEY\"\n  caveat no longer fires for `claude-cli:*` chat models — that recipe declares\n  `supports_subagent_loop` and drives Minions through the OAuth session.\n- Docs: README quick-start subsection, CHANGELOG, `init --help`, and the\n  fail-loud no-provider hint now advertises the keyless path.\n\n## Tests\n\n- New `test/init-claude-code-mode.test.ts` (12 tests): keyless predicate,\n  tailored refusal messages, re-init seeding, and the claude-cli recipe's\n  keyless contract (no required env vars, subagent loop supported).\n- All touched-area suites green (init-env-detection, init-embed-check,\n  init-mode-picker, init-provider-picker, sync-no-embed-sentinel, advisor-core,\n  import-file, embed-preflight, embedding-dim-check, config);\n  `bun run typecheck` clean.\n- End-to-end smoke with all provider keys unset:\n  init → import → keyword search returns the right page → `embed --stale`\n  refuses with the keyless message → `init --force --pglite --embedding-model\n  zeroentropyai:zembed-1 --embedding-dimensions 1280` clears the sentinels and\n  configures the model.\n\n## Out of scope (deliberate)\n\n- Query expansion / reranking via Claude Code skills (the issue's stretch\n  ideas): expansion is already gated behind `isAvailable('expansion')` and off\n  in the conservative bundle; the reranker fails open. Nothing breaks keylessly;\n  wiring them through `claude-cli` can be a follow-up.\n- Postgres/Supabase keyless installs: the flag targets the local PGLite path\n  per the issue; the runtime behavior (import/search) is engine-agnostic, so\n  extending later is trivial.\n\n## Unrelated heads-up found while working on this\n\n`master` (at least since faf5cdb) has an accidentally **committed\n`node_modules` symlink** pointing at `/tmp/fleet/repo/node_modules` — a path\nfrom someone's build environment. It's tracked despite `node_modules/` being in\n`.gitignore` (gitignore doesn't apply to already-tracked paths), so every fresh\nclone gets a dangling symlink and `bun install` behaves oddly until you\n`rm node_modules && bun install`. Worth a one-line cleanup commit:\n`git rm --cached node_modules`. Not fixed in this PR to keep it scoped to #94.\n\n🤖 Generated with [Claude Code](https://claude.com/claude-code)\n\nhttps://claude.ai/code/session_01WQ7za53ZWxWqL2XAfSgTux"
+}