brooktrails/gllm
Watch
2
0
Fork
You've already forked gllm
 0
Code Issues 3 Pull requests 2 Projects Releases Packages Wiki Activity Actions

feat: allow setting system prompt from a file #94

Open
rcsheets wants to merge 2 commits from feat/system-prompt-from-file into main
pull from: feat/system-prompt-from-file
merge into: brooktrails:main
Conversation 15 Commits 2 Files changed 6 +192 -33
rcsheets commented 2026-09-27 21:34:51 +00:00
Owner
Copy link

glchat could only take a system prompt inline, with --system. A prompt of any length lives better in a file, so this adds --system-file <path>:

glchat --system-file ~/prompts/reviewer.md

Behaviour

  • The file is read by glchat.ReadTextFile, the checks /read already applied, now factored out so both share them: regular files only, at most 1 MiB, UTF-8 without NULs, non-empty. Trailing whitespace is trimmed, so the newline an editor leaves is not part of the prompt, and a leading ~ is expanded (the shell does not expand it in --system-file=~/x).

  • --system and --system-file are mutually exclusive. Ranking one above the other would leave the losing flag unread, and a mistyped path would then start a conversation with no mention of the file.

  • An empty name (--system-file "", typically an unset variable in a script) is an error rather than "no system prompt".

  • A file that cannot be used is reported before discovery and before the TUI starts, against the flag that named it:

    glchat: --system-file: /tmp/pipe is not a regular file

The regular-file check matters more here than it does for /read: this read happens before the UI is up, so a plain os.ReadFile on a fifo is a client that prints nothing and never starts.

/read itself is unchanged in behaviour; its error messages are byte-for-byte the same, now produced by wrapping ReadTextFile's error with the /read: prefix.

Testing

  • internal/glchat: ReadTextFile trims only the tail and reports the on-disk size; a device is refused from the stat without being opened. The existing /read rejection table still passes unchanged.
  • cmd/glchat (new test file): the two flags are refused together; a missing file, a directory, a device, an empty file and an empty name each fail with an error led by --system-file:. These cases all return before discovery, so nothing is dialled.
  • gofmt, go build ./..., go vet ./... and go test ./... are clean.
  • By hand against the built binary: a fifo passed to --system-file is refused immediately, where it previously blocked.

README and AGENTS.md are updated; AGENTS.md records that any new path putting a file's text in front of the model goes through ReadTextFile.

🤖 Generated with Claude Code

glchat could only take a system prompt inline, with `--system`. A prompt of any length lives better in a file, so this adds `--system-file <path>`: glchat --system-file ~/prompts/reviewer.md ## Behaviour - The file is read by `glchat.ReadTextFile`, the checks `/read` already applied, now factored out so both share them: regular files only, at most 1 MiB, UTF-8 without NULs, non-empty. Trailing whitespace is trimmed, so the newline an editor leaves is not part of the prompt, and a leading `~` is expanded (the shell does not expand it in `--system-file=~/x`). - `--system` and `--system-file` are mutually exclusive. Ranking one above the other would leave the losing flag unread, and a mistyped path would then start a conversation with no mention of the file. - An empty name (`--system-file ""`, typically an unset variable in a script) is an error rather than "no system prompt". - A file that cannot be used is reported before discovery and before the TUI starts, against the flag that named it: glchat: --system-file: /tmp/pipe is not a regular file The regular-file check matters more here than it does for `/read`: this read happens before the UI is up, so a plain `os.ReadFile` on a fifo is a client that prints nothing and never starts. `/read` itself is unchanged in behaviour; its error messages are byte-for-byte the same, now produced by wrapping `ReadTextFile`'s error with the `/read:` prefix. ## Testing - `internal/glchat`: `ReadTextFile` trims only the tail and reports the on-disk size; a device is refused from the stat without being opened. The existing `/read` rejection table still passes unchanged. - `cmd/glchat` (new test file): the two flags are refused together; a missing file, a directory, a device, an empty file and an empty name each fail with an error led by `--system-file:`. These cases all return before discovery, so nothing is dialled. - `gofmt`, `go build ./...`, `go vet ./...` and `go test ./...` are clean. - By hand against the built binary: a fifo passed to `--system-file` is refused immediately, where it previously blocked. README and AGENTS.md are updated; AGENTS.md records that any new path putting a file's text in front of the model goes through `ReadTextFile`. 🤖 Generated with [Claude Code](https://claude.com/claude-code)
feat: allow setting system prompt from a file
All checks were successful
ci / test_and_build (pull_request) Successful in 26s
Details
9760238de5
pr-reviewer-bot commented 2026-09-27 21:34:52 +00:00
Collaborator
Copy link

The latest review couldn't run.

Jump to the current review

Review history
  • 2026-09-27 21:34 UTC · safety check · No verdict returned · r-b98bfc-fa072f
  • 2026-09-27 23:03 UTC · safety check · couldn't run · r-b9a0d7-40cfdc
  • 2026-09-27 23:06 UTC · safety check · couldn't run · r-b9a188-0ccbc7
  • 2026-09-27 23:08 UTC · safety check · failed · r-b9a205-d05dd0
  • 2026-09-28 01:38 UTC · safety check · failed · r-b9c52a-87c7be
  • 2026-09-28 01:42 UTC · safety check · failed · r-b9c5f1-ef593a
  • 2026-09-28 01:47 UTC · safety check · LGTM · r-b9c715-8031ee
  • 2026-09-28 01:50 UTC · safety check · Flag · r-b9c7fb-7b37cf
  • 2026-10-01 09:06 UTC · safety check · couldn't run · r-be22a4-af2ea1
  • 2026-10-01 22:32 UTC · safety check · couldn't run · r-bedf90-89a512
  • 2026-10-01 22:33 UTC · safety check · couldn't run · r-bedfd1-0924e6 · current
<!-- pr-reviewer:history --> **The latest review couldn't run.** [Jump to the current review](#issuecomment-4716) <details><summary>Review history</summary> - 2026-09-27 21:34 UTC · safety check · No verdict returned · `r-b98bfc-fa072f` - 2026-09-27 23:03 UTC · safety check · couldn't run · `r-b9a0d7-40cfdc` - 2026-09-27 23:06 UTC · safety check · couldn't run · `r-b9a188-0ccbc7` - 2026-09-27 23:08 UTC · safety check · failed · `r-b9a205-d05dd0` - 2026-09-28 01:38 UTC · safety check · failed · `r-b9c52a-87c7be` - 2026-09-28 01:42 UTC · safety check · failed · `r-b9c5f1-ef593a` - 2026-09-28 01:47 UTC · safety check · LGTM · `r-b9c715-8031ee` - 2026-09-28 01:50 UTC · safety check · Flag · `r-b9c7fb-7b37cf` - 2026-10-01 09:06 UTC · safety check · couldn't run · `r-be22a4-af2ea1` - 2026-10-01 22:32 UTC · safety check · couldn't run · `r-bedf90-89a512` - 2026-10-01 22:33 UTC · safety check · couldn't run · `r-bedfd1-0924e6` · **current** </details>
rcsheets commented 2026-09-27 23:03:50 +00:00
Author
Owner
Copy link

@pr-reviewer-bot review

@pr-reviewer-bot review
pr-reviewer-bot commented 2026-09-27 23:03:51 +00:00
Collaborator
Copy link

Superseded — a newer review is further down this thread. This pass is kept on the dashboard as r-b9a188-0ccbc7.

<!-- pr-reviewer:superseded --> *Superseded — a newer review is further down this thread. This pass is kept on the dashboard as `r-b9a188-0ccbc7`.*
rcsheets commented 2026-09-27 23:06:47 +00:00
Author
Owner
Copy link

@pr-reviewer-bot review

@pr-reviewer-bot review
pr-reviewer-bot commented 2026-09-27 23:06:48 +00:00
Collaborator
Copy link

Superseded — a newer review is further down this thread. This pass is kept on the dashboard as r-b9a205-d05dd0.

<!-- pr-reviewer:superseded --> *Superseded — a newer review is further down this thread. This pass is kept on the dashboard as `r-b9a205-d05dd0`.*
rcsheets commented 2026-09-27 23:08:51 +00:00
Author
Owner
Copy link

@pr-reviewer-bot review

@pr-reviewer-bot review
pr-reviewer-bot commented 2026-09-27 23:08:53 +00:00
Collaborator
Copy link

Superseded — a newer review is further down this thread. This pass is kept on the dashboard as r-b9c52a-87c7be.

<!-- pr-reviewer:superseded --> *Superseded — a newer review is further down this thread. This pass is kept on the dashboard as `r-b9c52a-87c7be`.*
rcsheets commented 2026-09-28 01:38:48 +00:00
Author
Owner
Copy link

@pr-reviewer-bot review

@pr-reviewer-bot review
pr-reviewer-bot commented 2026-09-28 01:38:50 +00:00
Collaborator
Copy link

Superseded — a newer review is further down this thread. This pass is kept on the dashboard as r-b9c5f1-ef593a.

<!-- pr-reviewer:superseded --> *Superseded — a newer review is further down this thread. This pass is kept on the dashboard as `r-b9c5f1-ef593a`.*
rcsheets commented 2026-09-28 01:42:07 +00:00
Author
Owner
Copy link

@pr-reviewer-bot review

@pr-reviewer-bot review
pr-reviewer-bot commented 2026-09-28 01:42:09 +00:00
Collaborator
Copy link

Superseded — a newer review is further down this thread. This pass is kept on the dashboard as r-b9c7fb-7b37cf.

<!-- pr-reviewer:superseded --> *Superseded — a newer review is further down this thread. This pass is kept on the dashboard as `r-b9c7fb-7b37cf`.*
rcsheets force-pushed feat/system-prompt-from-file from 9760238de5
All checks were successful
ci / test_and_build (pull_request) Successful in 26s
Details
to 2ac7166bc0
All checks were successful
ci / test_and_build (pull_request) Successful in 28s
Details
2026-09-28 01:46:59 +00:00 Compare
rcsheets commented 2026-09-28 01:50:49 +00:00
Author
Owner
Copy link

@pr-reviewer-bot review again

@pr-reviewer-bot review again
pr-reviewer-bot commented 2026-09-28 01:50:52 +00:00
Collaborator
Copy link

Superseded — a newer review is further down this thread. This pass is kept on the dashboard as r-bedfd1-0924e6.

<!-- pr-reviewer:superseded --> *Superseded — a newer review is further down this thread. This pass is kept on the dashboard as `r-bedfd1-0924e6`.*
feat(glchat): read --system-file the way /read reads a file
All checks were successful
ci / test_and_build (pull_request) Successful in 30s
Details
5f425ac359 
--system-file read its file with a bare os.ReadFile, and only when
--system was empty. That left several ways for it to go wrong quietly:

- A fifo or device blocked the read, before the TUI was up, so the client
  printed nothing and never started.
- With both flags given, --system won and the file was never opened, so a
  mistyped path went unreported.
- The file's trailing newline became part of the system prompt, and an
  empty file meant no system prompt at all, without a word.

/read already had the checks this needs. They move into an exported
glchat.ReadTextFile -- regular files only, at most 1 MiB, UTF-8 without
NULs, non-empty, trailing whitespace trimmed, a leading ~ expanded -- which
/read and --system-file now both call. /read's messages are unchanged.

--system and --system-file are now mutually exclusive rather than ranked,
and an empty --system-file name is an error rather than "no prompt".
Errors lead with the flag, as the other commands' do:

    glchat: --system-file: /tmp/pipe is not a regular file

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
rcsheets force-pushed feat/system-prompt-from-file from 5f425ac359
All checks were successful
ci / test_and_build (pull_request) Successful in 30s
Details
to b17afdafe3
All checks were successful
ci / test_and_build (pull_request) Successful in 30s
Details
2026-10-01 22:32:46 +00:00 Compare
rcsheets commented 2026-10-01 22:33:52 +00:00
Author
Owner
Copy link

@pr-reviewer-bot review

@pr-reviewer-bot review
pr-reviewer-bot commented 2026-10-01 22:33:54 +00:00
Collaborator
Copy link

Automated review by pr-reviewer v0.52.3 | Safety Check | Ministral 3 Instruct | tracking id r-bedfd1-0924e6
This is an AI-generated review and may contain mistakes.

Status: ❌ Failed

This review couldn't be completed: that model isn't loaded on the inference service right now, and no alternate model was able to review it either. Consider splitting this PR into smaller changes. Tracking id r-bedfd1-0924e6.

Comment @pr-reviewer-bot retry to try again.

<!-- pr-reviewer:review --> *Automated review by [pr-reviewer](https://git.brooktrails.org/brooktrails/pr-reviewer) v0.52.3 | Safety Check | Ministral 3 Instruct | tracking id `r-bedfd1-0924e6`* *This is an AI-generated review and may contain mistakes.* **Status:** ❌ Failed --- This review couldn't be completed: that model isn't loaded on the inference service right now, and no alternate model was able to review it either. Consider splitting this PR into smaller changes. Tracking id `r-bedfd1-0924e6`. Comment `@pr-reviewer-bot retry` to try again.
All checks were successful
ci / test_and_build (pull_request) Successful in 30s
Details
This pull request doesn't have enough approvals yet. 0 of 1 approvals granted.
This branch is out-of-date with the base branch
You are not authorized to merge this pull request.
View command line instructions

Checkout

From your project repository, check out a new branch and test the changes.
git fetch -u origin feat/system-prompt-from-file:feat/system-prompt-from-file
git switch feat/system-prompt-from-file
Sign in to join this conversation.
Reviewers
No reviewers
Labels
Clear labels
No items
No labels
Milestone
Clear milestone
No items
No milestone
Projects
Clear projects
No items
No project
Assignees
Clear assignees
No assignees
2 participants
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference
brooktrails/gllm!94
No description provided.