Keep machine-specific agent config out of the repository

[00PrabalK00]

Running continuum install in this repository produced three untracked config files, and all of them embed the absolute path of whoever ran it:

# .codex/config.toml
args = ["mcp", "serve", "--project", "C:/Users/Prabal/continuum"]

The same path appears in .gemini/settings.json and in the hook commands in .claude/settings.json. Committing any of them breaks the next person who clones, and puts a home directory into a public repository.

They are ignored now, along with .claude/settings.local.json and the scheduled-tasks lock file.

.claude/skills/continuum/SKILL.md is committed instead of ignored: it contains no paths and describes the project rather than the machine, so it is useful to anyone who clones.

Worth considering separately

This is a gitignore fix, not a fix to the underlying behaviour. continuum install could write a relative path, since every one of those agents runs with the project as its working directory, and the plugin manifest already uses ${CLAUDE_PROJECT_DIR} rather than an absolute path. That would make the generated config committable and shareable across a team. It is a behaviour change to install, so it belongs in its own PR rather than being folded into an ignore rule.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: f396a193e0

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[00PrabalK00]

Both right, and the second one showed the ignore was covering for the real problem rather than fixing it.

Shared Claude settings. .claude/settings.json is configuration Continuum merges into rather than owns, exactly as you say, so hiding the whole file to contain one machine-specific string was the wrong trade. The hooks now reference ${CLAUDE_PROJECT_DIR}, which Claude Code resolves itself:

continuum hook session-start --project "${CLAUDE_PROJECT_DIR}"

The file is machine-independent, so it is committed rather than ignored, and anyone cloning gets the hooks. Verified by reinstalling and regenerating rather than by reasoning about it: the written file contains no path at all.

Directory-wide ignores. Narrowed to .codex/config.toml and .gemini/settings.json, so anything else a contributor puts in those directories stays trackable.

Those two still carry an absolute path, because neither Codex nor Gemini documents a project-directory variable I can rely on the way Claude Code does. So they stay ignored, but that is now a statement about two generated files rather than a blanket over three directories. If either grows a portable form later, the same fix applies and the ignore can go.

One test added: the hook commands must contain ${CLAUDE_PROJECT_DIR} and must not contain the project path. 751 in the suite.