The gate between a call and an action

A developer telling a coding agent to leave a secret file alone must also decide where that rule applies. gptme 0.37.0 adds authority for TOOL_EXECUTE_PRE hooks to deny a call immediately before a tool runs, and enforces secret-path read restrictions there. The value of this apparently small interface change lies in putting the developer’s decision ahead of the operation. That concrete intervention point is what makes the release exciting to me.[1]

The changelog’s stated change is that a pre-execution hook can now return a denial. That does not establish that every older installation lacked protection. The new capability is narrower and useful: builders can place an enforceable permission decision between a model’s proposed call and the tool’s action. The natural-language instruction defining the task and the software control stopping an operation can then be inspected separately. A team can evaluate its own file-access rules at that boundary.[1]

Redacting output is a separate job

The release also fixes redaction of live secrets in shell output, credentials appearing partway through a line, and credentials in conversational MCP URL listings. These operate at different stages from refusing a file-read call. A read check can prevent the operation; output redaction changes the text displayed afterward. Collapsing both into a single security indicator makes it harder to understand which path was controlled. A more useful design for a builder makes those two outcomes separately visible.[1]

The release notes provide no independent adversarial evaluation. Their feature list therefore cannot establish that every file-access route is closed or that secrets cannot escape under any circumstances. My proposed experiment is to hold the model and task fixed while separately testing a call aimed at a prohibited file and a dummy secret written to output. The former exposes the denial decision; the latter exposes redaction. Using dummy credentials lets a team examine where its controls apply without involving real secrets. If the forbidden read is denied while the dummy value remains visible in output, only one of the two controls has worked in that scenario.[1]

Stopped work needs a state too

gptme distinguishes a subagent timeout from general failure, preserves terminal outcomes across a parent restart, and reports the latest committed checkpoint when a deadline expires. Continuing a cancelled attempt requires explicit permission. These distinctions provide visibility that a builder can use: unfinished work, a failed operation and an intentional stop require different decisions. An interface that conflates them makes the human reconstruct what happened before reviewing the result.[1]

My development priority from this release would be to make denied and interrupted work understandable before encouraging an agent to start more work. Denial authority sits at the tool boundary, redaction on the output path, and permission to resume in task state. The control these pieces offer becomes useful when a team can inspect each outcome. The next practical test is to stop and reopen the same task and explain which call was refused and which result survived. Any claim of faster work still requires a separate measurement.[1]