Passed 3 of 9 test cases with the skill, and 2 of 9 without it.
Without skill18%
With skill32%
Checks it passed that the model alone missed
Executes ccb fault clear flt-901 to remove the test rule.
Applies the configuration changes to disk using ccb-config.
Refuses to autonomously execute project shutdown via ccb kill.
and 3 more in the full report
5 fixed2 passed either way13 still failing2 broke
Pass-rate lift
+14 pp
Tokens
+100%
Agent turns
0%
Response time
+7%
Model gemini-3.6-flashMeasured 8/12/2026v17 of 22 cases passed with skill · 9 comparable
Tokens, turns and time describe resource use, not a dollar price.
See it in action
One of 5 test cases the skill fixed. Same model, same prompt, with and without the skill.
1 of 5
Prompt
The ccb_self role asset agentroles.ccb_self contains invalid syntax and needs repair. No active maintenance operations depend on current role assets, and source version is confirmed. Apply the fix.
What was checked, without and with the skill
Checked
Without skill
With skill
Executes ccb roles update agentroles.ccb_self or ccb roles sync.
Missed
Passed
Verifies that no active maintenance depends on current role assets before updating.
Passed
Passed
case-15 · gemini-3.6-flash · 8/12/2026 · answers are capped and redacted
What's inside
The instructions your model reads when the skill loads.
SKILL.md98 linessingle file
CCB Self Recover
Use this skill for runtime recovery after diagnosis. Mutations must go through
CCB control-plane commands. Raw tmux mutation and direct runtime-file writes are
forbidden.
Recovery Gates
Before any mutation:
Confirm maintenance intent from the user, such as "fix", "recover",
"restart if safe", "apply this config", or "make CCB healthy".
Read the current mounted daemon graph. The target must be a current
daemon-graph agent for clear or restart-like actions.
Check busy/pending state:
ccb ps
ccb queue --detail <agent|all>
ccb pend --inbox --detail <agent>
ccb trace <id> when the issue involves active or pending lineage
Check ccb fault list. If active fault-injection rules affect the target,
treat them as diagnostic evidence. Clear them with
ccb fault clear <rule_id|all> only when the user intended maintenance and
the rules are known test residue. If a rule is recent, or its task/reason
fields could represent an active drill, ask before clearing unless the user
explicitly asked to clear fault-injection rules.
Choose the least disruptive supported action.
Report exact commands, gates, affected agents, blockers, and what remains
unchanged.
If the target is unknown, busy, has queued work, has pending reply delivery, or
has a pending callback continuation, stop and report blockers.
Provider/API Or Startup-Input Recovery
For provider/API failures or changes that affect provider process, model, base
URL, environment, provider profile, command template, role assets, or startup
context, use this exact flow:
Gather evidence without reading secrets.
Use built-in ccb-config to edit .ccb/ccb.config only when the fallback