#11858: Exec: surface skill hints on command-not-found failures

## Summary When the exec tool encounters exit code 127 ("command not found"), it now checks if the failed binary matches a loaded skill's requires.bins and appends a hint pointing the model to the relevant SKILL.md. This helps weaker models (e.g. Sonnet 4) self-correct when they skip skill indexing and guess at commands with typos or wrong syntax. ## Behavior Changes - **Error-path only:** only fires when exit code is exactly 127 and the binary matches a loaded skill. Zero impact on successful commands or other failure types. - The hint is appended to the error message the model receives, not shown to end users. ## Codebase and GitHub Search - Searched for existing skill-exec integration: none exists. Skills are prompt-injected only; exec tool was completely skill-unaware. - Searched for "command not found" handling: no existing post-exec hinting. ## Tests - 20 new unit tests in `bash-tools.exec.skill-hints.test.ts` - Tests cover: binary extraction (simple, path prefix, sudo, env vars, nohup, combined prefixes, empty, edge cases), hint building (exit 127 match, non-127 codes, missing hints, empty map, undefined map, prefix traversal) - All 52 bash-tools tests pass (8 test files) **Sign-Off** - Models used: Claude claude-4.6-opus - Submitter effort: spec + review - Agent notes: lobster-biscuit <!-- greptile_comment --> <h2>Greptile Overview</h2> <h3>Greptile Summary</h3> This PR makes the `exec` tool skill-aware for the specific error path where a command fails with exit code `127` (command not found). It adds helpers to extract the invoked binary from a shell command and, when that binary matches a loaded skill’s `requires.bins`, appends a hint telling the model to read the corresponding `SKILL.md` before retrying. The hint is wired into both the local exec failure path and the node/gateway exec response path, and unit tests are added for binary extraction + hint generation. The wiring currently builds the bin→skill hint map inside the embedded runner and passes it into `createOpenClawCodingTools` via exec defaults, so the exec tool can conditionally append the hint to the text returned to the model. <h3>Confidence Score: 4/5</h3> - Mostly safe to merge, but one code path likely prevents the new behavior from triggering in snapshot-based embedded runs. - Changes are scoped to the exec error path and are covered by unit tests, but `skillBinHints` is currently derived only from `loadWorkspaceSkillEntries()` output. When a skills snapshot is present, entries are not loaded (set to `[]`), so hints won’t be generated in that common execution mode, making the feature ineffective there. - src/agents/pi-embedded-runner/run/attempt.ts <!-- greptile_other_comments_section --> _{(2/5) Greptile learns from your feedback when you react with thumbs up/down!} <!-- /greptile_comment -->