lsm / neokai / 25257928187

82%

push

github

Prevent coder agents from merging PRs (#1764)

* Prevent coder agents from merging PRs

Make the coding workflow explicit that coders stop after implementation and add a runtime Bash hook guard so coder-role sessions cannot run gh pr merge.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

* Address coder merge guard review nits

Remove redundant coder prompt wording and cover backtick subshell attempts in the PR merge guard.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

* Harden merge-guard regex and cover legacy Room coder sessions

- Add session.type === 'coder' check so Room coder sessions are also blocked
- Extend regex to handle env-var prefixes (GH_TOKEN=... gh pr merge), command
  builtin, and leading whitespace
- Add tests for legacy coder sessions and prefixed command forms

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

* Make merge guard data-driven via DeclarativeToolGuard on workflow node

Move the coder no-merge hook from hardcoded functions in
query-options-builder.ts to a declarative tool guard on the WorkflowNodeAgent
definition. The builder now compiles any declarative guards from the context
without hardcoded knowledge of specific rules.

- Add DeclarativeToolGuard type to shared space types
- Add toolGuards field to WorkflowNodeAgent
- Wire toolGuards through SlotOverrides → AgentSessionInit → QueryOptionsBuilderContext
- Replace hardcoded isCoderRoleSession/createBlockCoderPrMergeHook with
  generic compileToolGuard that works for any declarative guard
- Add CODER_NO_MERGE_GUARD to both coder agent slots in built-in workflows
- Cover env wrapper (env GH_TOKEN=... gh pr merge) in the regex pattern
- Remove Room coder session.type === 'coder' check (per owner direction)
- Update tests to use data-driven approach with toolGuards on context

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

* Fix agent-session test constructor calls for new toolGuards parameter

Co-Authored-By: Claude Opus 4.7 <no... (continued)

7802 of 10201 branches covered (76.48%)

Branch coverage included in aggregate %.

99 of 99 new or added lines in 8 files covered. (100.0%)

60576 of 84617 relevant lines covered (71.59%)

256.01 hits per line