GBrain stores internal cross-page references in slug form (e.g.
`[Alice](./alice)`) because the slug is the canonical identifier in the
DB. That works inside GBrain's own resolution layer.

But when those pages are exported as `.md` files on disk and opened in
standard markdown viewers (Obsidian, VS Code preview, GitHub web view,
typical mkdocs/jekyll renderers), the viewers look for a literal file
at `./alice` — which doesn't exist. The actual file is `./alice.md`.

Result: every internal link in an exported brain is silently broken on
disk. The user clicks `[小龙]` in `龙虾群.md`, sees a 404 / empty page,
and cannot navigate the brain outside of GBrain itself. This defeats
half the value of having the brain stored as portable markdown.

Fix:

Add `normalizeInternalLinks(content)` that runs over each page's
serialized markdown right before `writeFileSync` and rewrites slug-form
internal links to filename-form by appending `.md`:

  [Alice](./alice)            -> [Alice](./alice.md)
  [Alice](alice)              -> [Alice](alice.md)
  [Alice](../people/alice)    -> [Alice](../people/alice.md)
  [小龙](../people/小龙)        -> [小龙](../people/小龙.md)

Conservative: leaves untouched anything that looks external or already
extended:

- URL schemes (http:, https:, mailto:, ftp:, file:, tel:, ...) — skip
- Anchors (#section)                                            — skip
- Empty targets                                                 — skip
- Trailing slash (directory references)                         — skip
- Already has any extension (.md, .png, .pdf, .MD, ...)         — skip
- Preserves query strings and anchors when appending:
  [Section](./alice#bio) -> [Section](./alice.md#bio)
  [Search](./alice?q=t)  -> [Search](./alice.md?q=t)

The DB content stays slug-form (GBrain's internal convention is
unchanged). Only the on-disk export gets the `.md` annotation, so the
exported markdown is viewable as-is by any standard renderer.

Real-world reproduction this fix addresses:

  $ gbrain put 龙虾群 < <(echo '[小龙](./小龙)')
  $ gbrain export --dir /tmp/out
  $ cat /tmp/out/龙虾群.md
  # before this PR: contains [小龙](./小龙)  — clicking 404s
  # after this PR:  contains [小龙](./小龙.md) — clicking opens the file

Impact:
- 2 files changed, +149 / -1 lines (1 line of helper invocation +
  ~40 lines of helper + comment + 26 tests)
- Zero behavior change for external URLs, anchors, or already-extended
  links
- DB content unchanged — only the on-disk export representation gains
  the `.md` annotation
- Existing exports remain valid (re-running export on an already-exported
  brain is idempotent because already-extended links are skipped)

Tests:
- 26 new tests covering: same-dir slug, parent-dir slug, deep nesting,
  CJK slugs, multiple links per line, multi-line markdown, all 6
  external schemes (http/https/mailto/file/ftp/tel), all 4 extension
  cases (md/png/pdf/uppercase), anchor preservation, query preservation,
  empty/trailing-slash/no-link edge cases.
- All 26 tests pass.
- Full suite: 612 pass / no new regressions (4 pre-existing PGLiteEngine
  failures are unrelated and exist on master).

Fifth in a series of practical PRs from a real Chinese-speaking deploy.
Companion to:
- garrytan#114 (chunker CJK)
- garrytan#115 (slugify CJK)
- garrytan#119 (sync git quotepath CJK)
- garrytan#121 (self-contained API keys)

Same theme: GBrain is meaningfully more useful when the markdown export
is a first-class deliverable, not a half-broken side-effect.

The "no keys when neither env nor file provides them" case became
impossible once garrytan#121 (self-contained API keys) ships — loadConfig now
reads keys from config.json if env is empty. Test now asserts the
stronger invariant: after env deletion, the previous env sentinel value
must not leak back via the returned config. File-level keys may
legitimately persist and are no longer asserted undefined.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

GBrain stores internal cross-page references in slug form (e.g.
`[Alice](./alice)`) because the slug is the canonical identifier in the
DB. That works inside GBrain's own resolution layer.

But when those pages are exported as `.md` files on disk and opened in
standard markdown viewers (Obsidian, VS Code preview, GitHub web view,
typical mkdocs/jekyll renderers), the viewers look for a literal file
at `./alice` — which doesn't exist. The actual file is `./alice.md`.

Result: every internal link in an exported brain is silently broken on
disk. The user clicks `[小龙]` in `龙虾群.md`, sees a 404 / empty page,
and cannot navigate the brain outside of GBrain itself. This defeats
half the value of having the brain stored as portable markdown.

Fix:

Add `normalizeInternalLinks(content)` that runs over each page's
serialized markdown right before `writeFileSync` and rewrites slug-form
internal links to filename-form by appending `.md`:

  [Alice](./alice)            -> [Alice](./alice.md)
  [Alice](alice)              -> [Alice](alice.md)
  [Alice](../people/alice)    -> [Alice](../people/alice.md)
  [小龙](../people/小龙)        -> [小龙](../people/小龙.md)

Conservative: leaves untouched anything that looks external or already
extended:

- URL schemes (http:, https:, mailto:, ftp:, file:, tel:, ...) — skip
- Anchors (#section)                                            — skip
- Empty targets                                                 — skip
- Trailing slash (directory references)                         — skip
- Already has any extension (.md, .png, .pdf, .MD, ...)         — skip
- Preserves query strings and anchors when appending:
  [Section](./alice#bio) -> [Section](./alice.md#bio)
  [Search](./alice?q=t)  -> [Search](./alice.md?q=t)

The DB content stays slug-form (GBrain's internal convention is
unchanged). Only the on-disk export gets the `.md` annotation, so the
exported markdown is viewable as-is by any standard renderer.

Real-world reproduction this fix addresses:

  $ gbrain put 龙虾群 < <(echo '[小龙](./小龙)')
  $ gbrain export --dir /tmp/out
  $ cat /tmp/out/龙虾群.md
  # before this PR: contains [小龙](./小龙)  — clicking 404s
  # after this PR:  contains [小龙](./小龙.md) — clicking opens the file

Impact:
- 2 files changed, +149 / -1 lines (1 line of helper invocation +
  ~40 lines of helper + comment + 26 tests)
- Zero behavior change for external URLs, anchors, or already-extended
  links
- DB content unchanged — only the on-disk export representation gains
  the `.md` annotation
- Existing exports remain valid (re-running export on an already-exported
  brain is idempotent because already-extended links are skipped)

Tests:
- 26 new tests covering: same-dir slug, parent-dir slug, deep nesting,
  CJK slugs, multiple links per line, multi-line markdown, all 6
  external schemes (http/https/mailto/file/ftp/tel), all 4 extension
  cases (md/png/pdf/uppercase), anchor preservation, query preservation,
  empty/trailing-slash/no-link edge cases.
- All 26 tests pass.
- Full suite: 612 pass / no new regressions (4 pre-existing PGLiteEngine
  failures are unrelated and exist on master).

Fifth in a series of practical PRs from a real Chinese-speaking deploy.
Companion to:
- garrytan#114 (chunker CJK)
- garrytan#115 (slugify CJK)
- garrytan#119 (sync git quotepath CJK)
- garrytan#121 (self-contained API keys)

Same theme: GBrain is meaningfully more useful when the markdown export
is a first-class deliverable, not a half-broken side-effect.

…#121)

Two small ergonomics fixes folded together (#765 deferred — see TODOS.md
follow-up; the CJK PGLite extraction was bigger than the plan estimated).

#779 reworked (alexandreroumieu-codeapprentice): silence the
missing-max_batch_tokens startup warning for recipes with genuinely
dynamic batch capacity. New `EmbeddingTouchpoint.no_batch_cap?: true`
field. Set on ollama (capacity depends on locally loaded model +
OLLAMA_NUM_PARALLEL), litellm-proxy (depends on backend), llama-server
(set by --ctx-size at server launch). Three less stderr warnings on
every gateway configure; google still warns (it's a real fixed-cap
provider that ought to ship a max_batch_tokens declaration).

Bonus: litellm-proxy now declares `user_provided_models: true`, removing
the last consumer of the legacy `recipe.id === 'litellm'` hardcode in
gateway.ts:223 (D8=A wire-through completion).

#121 reworked (vinsew): self-contained API keys. Two parts:

  1. config.ts: ANTHROPIC_API_KEY env merge was silently missing.
     loadConfig() merged OPENAI_API_KEY but not ANTHROPIC_API_KEY into
     the file-config-shape result. One-line addition.

  2. cli.ts:buildGatewayConfig: when ~/.gbrain/config.json declares
     openai_api_key / anthropic_api_key but the process env doesn't
     have those env vars set (common for launchd-spawned daemons,
     agent subprocess tools, containers that don't propagate
     ~/.zshrc), fold the config-file values into the gateway env
     snapshot. Process env still wins (loaded last) so per-process
     overrides keep working.

Tests (4 cases in test/ai/no-batch-cap-suppression.test.ts):
- Ollama / LiteLLM / llama-server all declare no_batch_cap: true
- configureGateway does NOT warn for those three
- configureGateway STILL warns for google (regression guard)
- Cross-cutting invariant: empty-models recipes declare user_provided_models

Tests: bun test test/ai/ — 128/128 (4 new + 124 prior).

Plan: ~/.claude/plans/ok-lets-turn-this-enumerated-sonnet.md (commit 9 of 11).
#765 (Hunyuan PGLite + CJK keyword fallback) deferred to TODOS.md
follow-up; the CJK extraction (~150 lines + scoring logic + tests) is
larger than the wave's adjacent-fix lane should carry. Closes that PR
with a deferral note.

Co-Authored-By: alexandreroumieu-codeapprentice <noreply@github.com>
Co-Authored-By: vinsew <noreply@github.com>
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>