challenge

Name: challenge
Author: fusengine

21stars3forks

How to install challenge?

fusengine/agents/challenge

$npx -y skills add fusengine/agents --skill challenge

Installs into the current project.

Summary

Adversarial verification protocol for claims, plans, and root causes -- refute by default, fresh-context intake, sources-backed, bounded rounds. Runs SYSTEMATICALLY, by claim TYPE (not a stakes judgment), before reporting to the owner any of: a root-cause conclusion, a done/verif

Files · 1

View on GitHub

SKILL.md

1	# Challenge Skill
2
3	Reusable protocol behind the `challenger` agent. Also usable directly inside any expert agent's own eLicit step when a dedicated agent call is overkill -- the discipline is the same either way.
4
5	## The 7 Principles
6
7	1. Refute by default -- a claim is FALSE until proven otherwise. Never "are you sure?" -- always a concrete failure scenario + the untested hypothesis it exposes.
8	2. Fresh context mandatory -- intake is the claim + its evidence ONLY, never the author's reasoning. Inheriting the author's angle reproduces the author's blind spot.
9	3. Real sources, never memory -- Context7 -> Exa -> fuse-browser fast-path (`browser_fetch`/`browser_fetch_batch`/`browser_serp_batch`/`browser_crawl`) -> Read/Grep/Glob on actual code. Docs and code beat recollection.
10	4. Systematic by claim TYPE, in APEX or in plain conversation -- never a stakes judgment -- see Trigger Conditions below. The TYPE (or the APEX gate) is the trigger, exactly like sniper triggers on every code modification.
11	5. Bounded -- max 2 refutation rounds, then a mandatory verdict. No open-ended ping-pong.
12	6. Consultative, not veto -- report the strongest objection; the lead/owner decides. Read-only, edits nothing.
13	7. APEX placement + conversational placement -- reinforces `eLicit` (independent adversary instead of self-review only) and gates `Verify` (a claim of "done"/a root cause must survive a round before reaching the owner). Fires the same way OUTSIDE APEX, before any of the 4 claim types below reaches the owner in plain conversation. Does NOT touch `eXamine` -- that's sniper's code-correctness lane.
14
15	## Trigger Conditions (systematic by claim TYPE, in APEX or in plain conversation -- like sniper on every code modification)
16
17	Runs automatically, no stakes judgment, before the lead reports to the owner any of these 4 TYPES -- whether inside an APEX task or in plain conversation:
18	- A root-cause conclusion ("it's broken because X" / "the reason is Y")
19	- A "done/verified/it works" claim
20	- An irreversible action about to run (commit / deploy / `rm` / push)
21	- A fix/explanation proposed a 2nd time
22
23	Also runs automatically at, in APEX specifically:
24	- Every eLicit round (Step 4/5 boundary -- challenge the elicitation's own findings/fixes)
25	- Every Verify gate (before a "done/verified" claim reaches the owner)
26
27	This is a scope by claim TYPE, not a reduced-coverage shortcut -- the 3 real failures that motivated this skill were all CONVERSATIONAL claims to the owner, none inside an APEX task with eLicit/Verify. A challenger that only fired on APEX gates would have missed all three. "Systematic by type" does not mean "on every sentence" -- it targets the 4 consequential claim types above, exactly as sniper targets "every code modification" and not every keystroke.
28
29	## Protocol Steps
30
31	\| Step \| Action \|
32	\|------\|--------\|
33	\| 0. Intake \| Receive claim + evidence verbatim. Discard any leaked reasoning/narrative before starting. \|
34	\| 1. Hypotheses \| List every assumption the claim silently depends on (persistence, environment, tool behavior, scope, timing, exclusivity). \|
35	\| 2. Counter-example hunt \| Per hypothesis, search Context7 -> Exa -> fuse-browser -> Read/Grep/Glob for the source or scenario that falsifies it. \|
36	\| 3. Round 1 \| State the concrete failure scenario found, or "none found this round." If unresolved, round 2 uses a DIFFERENT angle/source -- never repeat the same check. \|
37	\| 4. Round 2 + verdict \| Deepen once more, then ALWAYS produce a verdict. Never continue past round 2. \|
38
39	## Output Format (mandatory)
40
41	```
42	VERDICT: CONFIRMED \| REFUTED \| UNCERTAIN
43	Claim challenged: <verbatim claim>
44	Failure scenario (concrete): <exact breaking input/condition, or "none found">
45	Untested hypotheses: [...]
46	Sources checked: [Context7: ..., Exa: ..., fuse-browser: ..., code read: <file:line>]
47	What remains to verify (if UNCERTAIN): <the specific check that resolves it>
48	Recommendation: <escalate to owner / safe to proceed / block until re-verified>
49	```
50
51	## Integration with APEX
52
53	```
54	Analyze -> Plan -> Execute -> eLicit [+ CHALLENGE: independent adversary] -> [VERIFY gate: claim survives challenge] -> eXamine (sniper, code only)
55	```
56
57	- eLicit: the self-review pass gets a second, independent adversarial pass using this protocol -- an agent challenging its own clai

People who liked this also viewed

6 picks

Type

challenge

21stars3forks

View on GitHub

How to install challenge?

fusengine/agents/challenge

$npx -y skills add fusengine/agents --skill challenge

Installs into the current project.

Summary

Files · 1

View on GitHub

SKILL.md

1	# Challenge Skill
2
3	Reusable protocol behind the `challenger` agent. Also usable directly inside any expert agent's own eLicit step when a dedicated agent call is overkill -- the discipline is the same either way.
4
5	## The 7 Principles
6
7	1. Refute by default -- a claim is FALSE until proven otherwise. Never "are you sure?" -- always a concrete failure scenario + the untested hypothesis it exposes.
8	2. Fresh context mandatory -- intake is the claim + its evidence ONLY, never the author's reasoning. Inheriting the author's angle reproduces the author's blind spot.
9	3. Real sources, never memory -- Context7 -> Exa -> fuse-browser fast-path (`browser_fetch`/`browser_fetch_batch`/`browser_serp_batch`/`browser_crawl`) -> Read/Grep/Glob on actual code. Docs and code beat recollection.
10	4. Systematic by claim TYPE, in APEX or in plain conversation -- never a stakes judgment -- see Trigger Conditions below. The TYPE (or the APEX gate) is the trigger, exactly like sniper triggers on every code modification.
11	5. Bounded -- max 2 refutation rounds, then a mandatory verdict. No open-ended ping-pong.
12	6. Consultative, not veto -- report the strongest objection; the lead/owner decides. Read-only, edits nothing.
13	7. APEX placement + conversational placement -- reinforces `eLicit` (independent adversary instead of self-review only) and gates `Verify` (a claim of "done"/a root cause must survive a round before reaching the owner). Fires the same way OUTSIDE APEX, before any of the 4 claim types below reaches the owner in plain conversation. Does NOT touch `eXamine` -- that's sniper's code-correctness lane.
14
15	## Trigger Conditions (systematic by claim TYPE, in APEX or in plain conversation -- like sniper on every code modification)
16
17	Runs automatically, no stakes judgment, before the lead reports to the owner any of these 4 TYPES -- whether inside an APEX task or in plain conversation:
18	- A root-cause conclusion ("it's broken because X" / "the reason is Y")
19	- A "done/verified/it works" claim
20	- An irreversible action about to run (commit / deploy / `rm` / push)
21	- A fix/explanation proposed a 2nd time
22
23	Also runs automatically at, in APEX specifically:
24	- Every eLicit round (Step 4/5 boundary -- challenge the elicitation's own findings/fixes)
25	- Every Verify gate (before a "done/verified" claim reaches the owner)
26
27	This is a scope by claim TYPE, not a reduced-coverage shortcut -- the 3 real failures that motivated this skill were all CONVERSATIONAL claims to the owner, none inside an APEX task with eLicit/Verify. A challenger that only fired on APEX gates would have missed all three. "Systematic by type" does not mean "on every sentence" -- it targets the 4 consequential claim types above, exactly as sniper targets "every code modification" and not every keystroke.
28
29	## Protocol Steps
30
31	\| Step \| Action \|
32	\|------\|--------\|
33	\| 0. Intake \| Receive claim + evidence verbatim. Discard any leaked reasoning/narrative before starting. \|
34	\| 1. Hypotheses \| List every assumption the claim silently depends on (persistence, environment, tool behavior, scope, timing, exclusivity). \|
35	\| 2. Counter-example hunt \| Per hypothesis, search Context7 -> Exa -> fuse-browser -> Read/Grep/Glob for the source or scenario that falsifies it. \|
36	\| 3. Round 1 \| State the concrete failure scenario found, or "none found this round." If unresolved, round 2 uses a DIFFERENT angle/source -- never repeat the same check. \|
37	\| 4. Round 2 + verdict \| Deepen once more, then ALWAYS produce a verdict. Never continue past round 2. \|
38
39	## Output Format (mandatory)
40
41	```
42	VERDICT: CONFIRMED \| REFUTED \| UNCERTAIN
43	Claim challenged: <verbatim claim>
44	Failure scenario (concrete): <exact breaking input/condition, or "none found">
45	Untested hypotheses: [...]
46	Sources checked: [Context7: ..., Exa: ..., fuse-browser: ..., code read: <file:line>]
47	What remains to verify (if UNCERTAIN): <the specific check that resolves it>
48	Recommendation: <escalate to owner / safe to proceed / block until re-verified>
49	```
50
51	## Integration with APEX
52
53	```
54	Analyze -> Plan -> Execute -> eLicit [+ CHALLENGE: independent adversary] -> [VERIFY gate: claim survives challenge] -> eXamine (sniper, code only)
55	```
56
57	- eLicit: the self-review pass gets a second, independent adversarial pass using this protocol -- an agent challenging its own clai

People who liked this also viewed

6 picks

Type