agent-gauntlet 1.2.0 → 1.2.2

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "agent-gauntlet",
-  "version": "1.2.0",
+  "version": "1.2.2",
   "description": "A CLI tool for testing AI coding agents",
   "license": "MIT",
   "author": "Paul Caplan",

package/skills/gauntlet-commit/SKILL.md

@@ -16,11 +16,15 @@ Commit with optional gauntlet validation. Runs `agent-gauntlet detect` first, va
 Run `agent-gauntlet detect` using `Bash`:
 
 ```bash
-agent-gauntlet detect 2>&1
+agent-gauntlet detect 2>&1; echo "DETECT_EXIT:$?"
 ```
 
-- If no changed files are reported → **skip to Step 4** (commit directly, no validation needed)
-- If changed files are reported → continue to Step 2
+Check the exit code from the `DETECT_EXIT:` line:
+
+- **Exit 0** → gates would run, continue to Step 2
+- **Exit 2** → no gates would run (no changes or no applicable gates), **skip to Step 4** (commit directly)
+- **Exit 1** → error, report the error to the user and stop
+- **Any other exit code** → treat as error, report output to the user, and stop
 
 ## Step 2 - Determine Validation Intent
 

package/skills/gauntlet-run/SKILL.md

@@ -13,26 +13,22 @@ Fix issues you reasonably agree with or believe the human wants to be fixed. Ski
 
 ## Procedure
 
-### Step 1 - Clean Logs
-
-Run `agent-gauntlet clean` to archive any previous log files.
-
-### Step 2 - Run Gauntlet
+### Step 1 - Run Gauntlet
 
 If the caller requests a specific review to be enabled, append `--enable-review <name>` to the run command for each requested review.
 
 Run `agent-gauntlet run` using `Bash` with `timeout: 300000`. **ALWAYS wait for and read the full command output** before proceeding — the command typically takes 1-2 minutes. **Verify you can see a `Status:` line in the output before continuing.**
 
-### Step 3 - Check Status
+### Step 2 - Check Status
 
 **NEVER assume success** — you must see an explicit `Status:` line before continuing. Check it and route accordingly:
-- `Status: Passed` → Go to Step 9.
-- `Status: Passed with warnings` → Go to Step 9.
-- `Status: Failed` → Continue to Step 4. **You MUST continue — do not stop here.**
-- `Status: Retry limit exceeded` → Run `agent-gauntlet clean` to archive logs. Go to Step 9.
+- `Status: Passed` → Go to Step 8.
+- `Status: Passed with warnings` → Go to Step 8.
+- `Status: Failed` → Continue to Step 3. **You MUST continue — do not stop here.**
+- `Status: Retry limit exceeded` → Go to Step 8.
 - No status line visible → **Known issue:** Bun can drop all stdout/stderr when LLM review subprocesses run. Read the console log file to get the status: find the latest `console.*.log` in the gauntlet log directory (e.g., `gauntlet_logs/console.1.log`) and look for the `Status:` line there. If no console log is found there, also check `gauntlet_logs/previous/` for logs from the most recent archived run. If no console log exists in either location, the command may have timed out or failed to run — re-run with a longer timeout or investigate the error. Do NOT proceed as if it passed.
 
-### Step 4 - Extract Failures
+### Step 3 - Extract Failures
 
 Required when status is Failed:
 - Infer the log directory from the file paths in the console output (e.g., if output references `gauntlet_logs/check_._lint.1.log`, the log directory is `gauntlet_logs/`)
@@ -42,31 +38,31 @@ Required when status is Failed:
   b. **Subagent delegation**: If your environment supports delegating work to a subagent but not the Task tool, delegate the extract-prompt instructions with the log directory to a subagent for processing.
   c. **Inline fallback**: If no subagent capability is available, follow the extract-prompt instructions yourself to read the log files and produce the compact failure summary.
 
-### Step 5 - Report Failures
+### Step 4 - Report Failures
 
-Print the compact failure summary returned from Step 4.
+Print the compact failure summary returned from Step 3.
 
-### Step 6 - Fix
+### Step 5 - Fix
 
 Apply the review guidance above to each failure and fix accordingly:
 - CHECK failures with Fix Skill: invoke the named skill
 - CHECK failures with Fix Instructions: follow the instructions
 - REVIEW violations: fix or skip per the review guidance above
 
-### Step 7 - Update Review Decisions
+### Step 6 - Update Review Decisions
 
 For REVIEW violations you addressed:
 - Read `update-prompt.md` from this skill's directory
-- **Update review decisions** using the first available strategy (same as Step 4):
+- **Update review decisions** using the first available strategy (same as Step 3):
   a. **Task tool** (Claude Code): `Task` with `subagent_type="general-purpose"`, `model="haiku"`, `prompt=` update-prompt content + log directory + decisions list. **Task calls MUST be synchronous** — NEVER use `run_in_background: true`.
   b. **Subagent delegation**: Delegate the update-prompt instructions with the log directory and decisions to a subagent.
   c. **Inline fallback**: Follow the update-prompt instructions yourself to update the review JSON files.
 
-### Step 8 - Re-run Verification
+### Step 7 - Re-run Verification
 
-**NEVER skip this step** — if the run failed, you MUST fix and re-run. Run the same command from Step 2 (including any `--enable-review` flags) again with `Bash` and `timeout: 300000`. Do NOT run `agent-gauntlet clean` between retries. The tool detects existing logs and automatically switches to verification mode. **Go back to Step 3** to check the status line and repeat.
+**NEVER skip this step** — if the run failed, you MUST fix and re-run. Run the same command from Step 1 (including any `--enable-review` flags) again with `Bash` and `timeout: 300000`. The tool detects existing logs and automatically switches to verification mode. **Go back to Step 2** to check the status line and repeat.
 
-### Step 9 - Summarize Session
+### Step 8 - Summarize Session
 
 Provide a summary of the session:
 - Final Status: (Passed / Passed with warnings / Retry limit exceeded)