dotemacs/modules/ai-config.el, branch load-graph-classify-start

refactor(auth): consolidate the auth-source secret lookup into one helper

2026-05-23T00:32:32+00:00

The auth-source-search + funcall-the-secret block was copied four times: calendar-sync--calendar-url, cj/auth-source-secret (ai-config), cj/--auth-source-password (transcription), and cj/slack--get-credential. Each searched authinfo, pulled :secret, and called it when the netrc backend returned a function. I pulled that into cj/auth-source-secret-value in system-lib (a leaf, so calendar-sync doesn't have to depend on ai-config and drag in the gptel stack). It takes an optional user and returns the secret or nil. The four callers now delegate to it: ai-config layers its required-secret error on top, and the others keep their nil-on-miss behavior. With the direct auth-source-search calls gone, I dropped the now-unused (require 'auth-source) from transcription, slack, and calendar-sync. The helper's autoload covers it. The transcription tests that exercise the delegated path stay green, and the primitive and the error wrapper get their own tests.

fix(ai-config): require gptel backend libs so the fork's constructors load

2026-05-22T20:34:07+00:00

cj/toggle-gptel and gptel chat errored with "Symbol's function definition is void: gptel-make-anthropic". The local gptel fork on :load-path with :ensure nil ships no generated autoloads, so (require 'gptel) loads gptel.el but never gptel-anthropic.el or gptel-openai.el, where the gptel-make-* constructors live. cj/ensure-gptel-backends then reached gptel-make-anthropic before it was defined. cj/ensure-gptel-backends now requires gptel-anthropic and gptel-openai first, through a small cj/--gptel-load-backend-libs helper. Verified end-to-end: with the fork on load-path, the constructors are fbound and both backends build.

refactor(ai-config): switch gptel to local fork, drop tab-width advice

2026-05-18T06:24:30+00:00

I switched the gptel use-package form to `:load-path "~/code/gptel"` with `:ensure nil` so Emacs loads from the fork instead of the MELPA release. The fork now carries the narrow `tab-width' copy in `gptel-org--create-prompt' that karthink redirected the upstream PR to, which replaces the local `:around' advice on `gptel--with-buffer-copy-internal' I'd been carrying. I also dropped the stale test file `tests/test-ai-config-gptel-prompt-tab-width.el' and the matching stub in `tests/testutil-ai-config.el'. Both existed only to test the advice I removed.

fix(ai-config): gptel-model must be a symbol, not a string

2026-05-16T11:07:46+00:00

The default-backend swap to gpt-5.5 (commit 0f029ab5) set `gptel-model' as the string "gpt-5.5". gptel's modeline-display code calls `symbolp' on the model value and signals `wrong-type-argument symbolp "gpt-5.5"' on every render, which manifested as Emacs freezing in the AI-Assistant buffer ("Querying ChatGPT..." → error in process sentinel → repeated redisplay errors). Both default-setting sites now use `'gpt-5.5' (interned symbol). The Anthropic backend tolerated string model names so the original "claude-opus-4-7" string worked, which is why this hadn't surfaced before.

chore(ai-config): switch default gptel backend to ChatGPT / gpt-5.5

2026-05-16T10:58:24+00:00

Two places set the default backend + model on gptel initialization -- `cj/ensure-gptel-backends' (the lazy-init fallback) and the `use-package gptel :config' block (the eager-set after initialization). Both now pick the ChatGPT backend with `gpt-5.5' instead of Claude with `claude-opus-4-7'.

feat(gptel-tools): wire web_fetch as a local tool

2026-05-16T10:17:21+00:00

Fourth ADOPT entry from `docs/design/gptel-tools-shortlist.org'. Lets gptel pull a URL into the conversation so the model can read docs / current API shapes / etc. without me copy-pasting. Shape: - URL must be `http://' or `https://' (file://, ftp://, javascript:, scheme-less, etc. are rejected at the validator). - HTML responses go through `pandoc -f html -t plain' so the model gets a reading shape that isn't full of markup; falls back to `w3m -dump -T text/html' if pandoc isn't on PATH; signals `user-error' if neither is. Pass `raw=t' to skip stripping. - Output capped at 200KB by default, hard cap 1MB; `max_bytes' argument lets the caller pick a lower cap. Truncation reported inline. - 4xx / 5xx response codes signal `error' with the code -- the alternative is returning an error page body, which the model would treat as content. `:confirm t' on the tool because every call is a real outbound network request. The tool's description warns that URLs go wherever the user-agent points, including internal networks if that's what the URL names. `tests/test-gptel-tools-web-fetch.el' -- 20 tests across Normal / Boundary / Error. URL validator covers http / https / non-string / empty / non-http schemes. `--effective-max-bytes' covers default / low-clamp / hard-cap / passthrough. Truncate helper covers under-cap / at-cap / over-cap with the marker. HTML stripper runs against real pandoc / w3m (both installed in dev env, neither should mangle simple markup). Orchestrator stubs `cj/gptel-web-fetch--retrieve' via `cl-letf' to cover normal / raw / 4xx / 5xx / oversize / bad-scheme paths. Wired into `cj/gptel-local-tool-features' so gptel exposes the tool on next restart. Note: `call-process-region' invocation flattened to a single `with-temp-buffer' with DELETE=t -- the initial draft nested a second temp buffer and routed output to the inner one, which got killed before `buffer-string' on the outer ran. Test caught it.

feat(gptel-tools): wire git_status / git_log / git_diff as local tools

2026-05-16T09:26:20+00:00

Three read-only git context tools so gptel can see what's changed without me pasting `git status` / `git log` / `git diff` output into every chat turn. Builds the first batch from the ADOPT bucket in `docs/design/gptel-tools-shortlist.org`. Shape per tool: - `gptel-tools/git_status.el` — `git status --short --branch` for a directory inside a git working tree under HOME. Returns the porcelain output, or a "Clean working tree" marker when only the branch line is present. - `gptel-tools/git_log.el` — `git log --oneline -nN` with an optional `--since` filter. N defaults to 20, capped at 100; nil / non- integer / out-of-range N falls back to the default. - `gptel-tools/git_diff.el` — `git diff [REF1 [REF2]] [-- FILE]`. Output capped at ~500KB so a runaway diff can't blow up context; truncation is reported inline. Validation is uniform: path must resolve under HOME, must be a directory, must be inside a git working tree (verified via `git rev-parse --is-inside-work-tree`). Color is disabled via `-c color.ui=false` at the git level (`git status` doesn't accept `--no-color` directly). Tests run against real temp git repos created via `process-file`, not mocked — there's nothing in gptel-tools/git_*.el that's process-mockable in a meaningful way, and a real `git init` + a couple of commits is cheaper than building a fake. 31 tests total: 7 for git_status, 11 for git_log, 13 for git_diff. Wired into `cj/gptel-local-tool-features` so gptel exposes the three tools on next restart.

feat(ai-conversations-browser): dired-style browser for saved GPTel conversations

2026-05-16T07:00:19+00:00

`cj/gptel-load-conversation` prompts via `completing-read`. A dedicated browser shows what each conversation is about at a glance and supports single-key load / delete / rename without having to scroll a minibuffer list. New module `modules/ai-conversations-browser.el` + `cj/gptel-browse-conversations` entry point bound to `C-; a b` ("browse conversations"). Opens `*GPTel-Conversations*` in `cj/gptel-browser-mode` (a `special-mode` derivative). Each row shows date, time, topic slug, and a preview of the most recent message (length configurable via `cj/gptel-browser-preview-length`, default 60 chars). Rows sort newest first. In the browser: - `RET` / `l`: load the conversation (delegates to `cj/gptel-load-conversation` with the file pre-selected via a `cl-letf` stub on `completing-read` so the user isn't prompted twice), then bury the window. - `d`: delete the file under point after `y-or-n-p` confirmation, re-render. - `r`: rename the file under point. Preserves the timestamp, slugifies the new topic, refuses unchanged input and existing targets. - `g`: refresh. - `n` / `p`: next / previous row. - `q`: quit-window. 21 tests cover the helpers (topic parsing, header stripping, preview shaping for truncate / short / empty cases, row-for-file with conversation + non-conversation filenames, rows enumeration, render output for empty + populated cases, newest-first sort, rename-target preservation of timestamp + slug, rename-target error on missing timestamp) and the file-touching actions (delete with y, cancel with n, rename, rename-on-empty-line error).

feat(ai-rewrite): add directive-picker wrappers around gptel-rewrite

2026-05-16T06:55:16+00:00

`gptel-rewrite` is the killer feature for the keep-gptel decision, and it now lives behind two commands instead of the bare call: - `cj/gptel-rewrite-with-directive` (`C-; a r`, replacing the former bare `gptel-rewrite` binding): completing-read on a directive name from `cj/gptel-rewrite-directives`, then rewrite the active region. - `cj/gptel-rewrite-redo-with-different-directive` (`C-; a R`): replay the prior region with a different directive. The region is preserved via markers stored buffer-local on the first call so it survives accept/reject of the prior rewrite. I picked the hook injection approach over an `:after`-advice + state-capture pattern. `gptel-rewrite-directives-hook` is an abnormal hook gptel-rewrite already consults for a per-call system message. Wrapping the call in a one-shot `let`-binding on that hook gives the directive exactly the lifetime of the rewrite and leaves nothing to clean up. Mutating `gptel-directives` globally would mean either restoring it afterward or living with the change -- both worse than the hook. Directives ship inline as a `defcustom` alist with the six names called out in the task -- `terse`, `fix-grammar`, `refactor-readability`, `add-docstring`, `explain-as-comment`, `shorten`. Customization is a `customize-variable` or `setq` away. 9 tests cover the defcustom shape (default names present, bodies non-empty strings), the wrapper (normal path, no-region error, unknown-directive error, last-state recording), and the redo (replays the prior region, errors when no previous, excludes the current directive from the re-pick prompt). `gptel-rewrite` stubbed in tests so no rewrite UI fires.

feat(ai-quick-ask): add cj/gptel-quick-ask one-shot command

2026-05-16T06:48:59+00:00

New module `modules/ai-quick-ask.el`. Bound to `C-; a q` via `cj/ai-keymap` ("quick ask"). `cj/gptel-quick-ask` reads a prompt in the minibuffer, creates a transient `*GPTel-Quick*` buffer in `cj/gptel-quick-mode` (a special-mode derivative with `q` / `escape` / `c` bindings), inserts "Q: " plus a response marker, then calls `gptel-request` with `:stream t` so the answer streams into the buffer. Doesn't touch `*AI-Assistant*`, doesn't autosave. Two follow-up commands work in the buffer: - `cj/gptel-quick-dismiss` (`q` / `escape`): delete the window and kill the buffer. Idempotent when the buffer is absent. - `cj/gptel-quick-continue` (`c`): extract the prompt + response, seed them into `*AI-Assistant*` under proper org headings (matching the `cj/gptel--fresh-org-prefix` shape), display the side window, then dismiss the quick buffer. 13 tests cover the pure helpers (initial-text shape, response extraction across normal / multi-line / no-marker / empty inputs, seed-text shape), the ask path (buffer created in right mode, prompt recorded, gptel-request called, empty-prompt error), the dismiss path (kills buffer / no-op when absent), and the continue path (seeds `*AI-Assistant*`, dismisses quick buffer, errors outside a quick buffer). `gptel-request` is stubbed in tests so nothing hits the network.