[Dcastroro]

After write_file, patch_file, or run_shell, the next model turn currently receives the original workspace snapshot. This can leave git status, branch, recent commits, and project documentation stale.

Keep the rules/tools/examples prefix stable and render workspace context separately. Mark that context dirty only after a mutating tool actually executes, and rebuild it once before the next prompt. Invalidation also happens if execution raises after a partial mutation. Argument validation, approval checks, and path restrictions are unchanged.

Fixes #15.

Validation:

• New regression cases on base 717cae4: 5 failed, 3 passed.
• Full suite on Windows (Python 3.10 and 3.13) and Linux (Python 3.12): 26 passed, 1 skipped in each environment; fake model responses and temporary local repositories only.
• Untouched base suite on Windows: 18 passed, 1 skipped.
• ruff check . with Ruff 0.17.1 reports 13 existing diagnostics on both base and this branch, with identical rule/message counts and no new diagnostics. The full lint check therefore remains failing; no rules or workflow checks are suppressed. These unrelated baseline diagnostics are outside this fix.
• git diff --check: passed.

The tests cover all three mutating tools, branch/commit changes, partial failure, stable prompt prefix, and no refresh after reads, denied writes, or invalid arguments. No runtime dependencies, credentials, or workflow changes are introduced.

AI assistance: repository research, implementation, and validation were performed with AI assistance. The regression results and the pre-existing lint limitation are provided for maintainer review.

[Copilot]

🟢 Approval recommended

The change is minimal and correctly scoped—dirty-marking only runs for executed mutating tools while preserving prompt ordering and the stable prefix—and it is comprehensively covered by the new regression tests.

No open issues 🎉

What changed in this PR

This PR fixes stale workspace context (git status, branch, recent commits, project docs) being presented to the model after a tool mutates the workspace within a session. Previously, WorkspaceContext.text() was embedded in the once-built prompt prefix, so it never reflected changes made by write_file, patch_file, or run_shell. The fix keeps the rules/tools/examples prefix stable and renders the workspace context as a separate, refreshable section, rebuilding it only when a mutating tool has actually executed (tracked via a workspace_dirty flag set in a finally block, so partial mutations before a failure are still captured). This addresses issue #15 and preserves the existing prompt ordering and prefix-cache behavior.

Changes:

• Added a workspace_dirty flag that run_tool sets after executing write_file, patch_file, or run_shell (including on partial-failure), and prompt() rebuilds WorkspaceContext and clears the flag before composing the next prompt.
• Removed workspace.text() from build_prefix and moved it into prompt() as a distinct section, keeping the overall section order unchanged.
• Added regression tests covering refresh after each mutating tool, git branch/commit changes, partial failure, the stable prefix, and no-refresh after reads, denied writes, or invalid arguments.

File	Description
mini_coding_agent.py	Introduces workspace_dirty, removes workspace text from the stable prefix, refreshes WorkspaceContext in prompt(), and marks dirty in run_tool's finally for mutating tools.
tests/​test_mini_coding_agent.py	Adds a git fixture and parametrized regression tests validating workspace refresh behavior and stable-prefix invariants.

_{Review effort: Balanced}