fix(tools): honor recursive list_dir depth

[CoreyRDean]

Non-technical summary

• intent now honors recursive list_dir requests, so the model can inspect one or two nested directory levels without guessing filenames.
• This matters now because the tool catalog already documented depth, but the implementation silently ignored it, weakening read-only context gathering.
• The system is better afterward because nested discovery is now real, bounded, and visible in both the tool result shape and prompt guidance.

Technical summary

• Implemented recursive list_dir traversal in internal/tools/tools.go, with depth defaulting to 1, a global max_entries cap, and no symlink traversal.
• Added a relative path field alongside name/type/size so nested entries stay grounded to the requested directory.
• Added focused tests covering depth semantics and truncation behavior in internal/tools/tools_test.go.
• Updated internal/model/prompt.go to advertise list_dir(path[, depth]) and bumped PromptTemplateVersion so cached prompts refresh against the new capability.
• Updated docs/SPEC.md so the documented return shape matches the recursive tool output.
• Breaking changes: none.

Additional notes

• Trade-off: the tool result now includes an extra path field, which is additive and keeps top-level name behavior intact.
• Follow-up intentionally deferred: broader file-grounding policy work such as enforcing read_file before edit proposals remains out of scope for this bounded increment.
• Remaining gap: this improves nested discovery, but higher-level prompt/engine policies can still evolve separately for richer multi-step editing behavior.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: c371048df0

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".

[chatgpt-codex-connector]

Skip unreadable subdirectories during recursive list_dir

When depth > 1, a single os.ReadDir failure in any child directory aborts the entire list_dir call and returns only an error, even if the parent directory was readable and already had useful entries. This is a regression from the non-recursive behavior and will surface in repos that contain permission-restricted folders (or transiently inaccessible mounts), where callers now lose all listing data instead of getting a partial tree. Consider treating child traversal failures as per-entry errors (or skipping that subtree) rather than failing the whole tool result.

Useful? React with 👍 / 👎.