AGENT SKILL

challenge

Run a risk-scaled blind-spot discovery pass over a converged, exported frame BETWEEN `think` and `spec-to-plan` (the seventh origin skill, third leg in flow order): pressure-test the spec through structured lenses, route every finding back through the existing deterministic moves as proposed-only content the human adjudicates, and on a clean pass record the examined lenses/surfaces and residual uncertainty — never a claim that there are no unknown unknowns. Use when the user says "challenge this spec", "blind-spot pass", "pressure- test the frame", "what are we missing", "unknown unknowns", or after `think` exports and before `devague plan new`. Authored and maintained in agentculture/devague (origin = devague); guildmaster pulls this skill from here and broadcasts it to the AgentCulture mesh — it is NOT vendored from guildmaster like the inbound skills here.

License 许可
Apache-2.0
最近更新
2026-09-29
Stars
★ 3
Forks
0

01 前言 Front Matter

name
challenge
description
Run a risk-scaled blind-spot discovery pass over a converged, exported frame BETWEEN `think` and `spec-to-plan` (the seventh origin skill, third leg in flow order): pressure-test the spec through structured lenses, route every finding back through the existing deterministic moves as proposed-only content the human adjudicates, and on a clean pass record the examined lenses/surfaces and residual uncertainty — never a claim that there are no unknown unknowns. Use when the user says "challenge this spec", "blind-spot pass", "pressure- test the frame", "what are we missing", "unknown unknowns", or after `think` exports and before `devague plan new`. Authored and maintained in agentculture/devague (origin = devague); guildmaster pulls this skill from here and broadcasts it to the AgentCulture mesh — it is NOT vendored from guildmaster like the inbound skills here.

03 SKILL.md 正文

<!-- lineage: claude (.claude/skills/challenge/SKILL.md) -> colleague (../colleague/.colleague/skills/challenge.md, ff7331e v1.63.1) -> qwen (this file). Surface adaptation: qwen drives the devague CLI directly through its shell tool; there is no curated tool allow-list to navigate. -->

challenge — hunt the blind spots before the plan inherits them

The skill is named challenge; it is the blind-spot discovery leg of the devague method — the seventh origin skill, sitting third in flow order, between the spec leg and the plan leg:

scope -> think -> challenge -> spec-to-plan -> assign-to-workforce -> deviate -> summarize-delivery

Before this leg existed, a frame could converge on precisely stated claims while the original framing was still incomplete: open questions only captured uncertainty someone had already noticed — nothing actively hunted omitted dimensions, hidden dependencies, or assumptions shared by everyone in the frame, and no record existed of which surfaces were ever examined (issue 73's problem statement). Strictly speaking, an unknown unknown cannot be listed directly — once articulated, it becomes a known unknown. The useful capability is therefore to raise the odds of discovering blind spots and lower the cost of the surprises that remain, not to promise their elimination.

That is the surprise-cost rationale: an articulated blind spot becomes a known unknown the method can manage; an unexamined one surfaces later as a mid-run deviate or a production surprise. Discovery before planning is cheaper than either — a proposed claim the human rejects costs minutes; the same gap found mid-fan-out stops a wave, and found in production it costs whatever the blast radius costs.

This doc is written for two readers. The operator — the main agent — runs the pass: sweeps the lenses, drives the deterministic CLI move by move, and proposes findings. The gate-owning human adjudicates: every finding lands proposed, and confirming, rejecting, or resolving it is the human exercising the existing spec gate (gate 1) — challenge adds no fourth gate, mirroring how deviate amends gate 2 rather than adding one.

How to run in qwen

The pass drives the devague CLI — the same agent-agnostic commands for any agent or operator (if devague isn't on PATH: uv tool install devague). Qwen Code has no curated tool allow-list: every move — capture, interrogate, park, converge, scope, question, review, export, and plan risk — runs directly through the shell tool (run_shell_command), and the operator runs the same commands directly in a terminal. confirm and reject are user-only decisions either way.

When it runs

The timing is a recorded decision — quote it, don't re-derive it:

the challenge pass runs after /think exports: challenge the converged, exported frame before devague plan new; findings reopen the frame, reconverge, and re-export the same dated spec file — /think stays self-contained (resolves q1)

— decision c17 in docs/specs/2026-07-15-challenge-skill.md.

Concretely: think finishes its own arc (converge, export) untouched. Then challenge pressure-tests the exported spec before devague plan new seeds a plan from it. Findings land as proposed claims, honesty conditions, questions, or parks — which reopens the frame — the human adjudicates, the frame reconverges, and devague export re-exports the same dated file (docs/specs/<created-date>-<slug>.md; exports are prefixed with the frame's creation date, so a re-export overwrites in place rather than spawning a duplicate). That reconverge-and-re-export loop repeats until the pass's findings are all adjudicated and the spec artifact carries them.

Proportionality — scale the pass to the risk

The pass is mandatory but proportional: lightweight for ordinary work, rigorous for high-risk work (integration option 1 from issue 73, per the issue author — a distinct operator skill, no new CLI engine until the workflow proves it needs one). Which work is high-risk is likewise a recorded decision:

the named escalation signals that deepen the pass from lightweight to rigorous: migrations, security-sensitive work, distributed state, hardware, destructive operations, other hard-to-reverse changes, concurrency hazards, and any surface that can lose user data (resolves q3)

— decision c19 in docs/specs/2026-07-15-challenge-skill.md.

If any escalation signal applies, run the rigorous form: every lens, deliberate counter-evidence hunting, and cheap probes where they would settle a real question. If none applies, a lightweight sweep — one pass over the lenses against the exported spec, minutes not hours — satisfies the method. Lightweight never means skipped: even the lightest pass leaves durable records (see the hard rules).

The lenses

Sweep the exported spec, the live frame, and the surfaces the idea touches through these structured lenses (from issue 73):

  • adjacent systems and hidden dependencies — what else reads, writes, or assumes the thing being changed;
  • unstated assumptions and missing counter-evidence — what everyone in the frame believes without a claim saying so, and what was never checked because nobody argued the other side;
  • overlooked actors, lifecycle stages, data flows, and failure modes — who and what the frame forgot: other users, upgrade/downgrade paths, half-completed operations;
  • security, migration, concurrency, operations, and reversibility — the classic hard surfaces, each also an escalation signal when present;
  • missing observability, containment, rollback, and recovery paths — when the surprise happens anyway, how it is seen, bounded, and undone;
  • cheap probes or experiments — small, scratch-space checks that could expose a surprise now instead of mid-run (probes never mutate the repo or .devague/ state).

The method

  1. Confirm the entry condition. A converged frame that think has already exported, and no plan seeded from it yet (devague status shows where the frame stands; the exported spec-md is the artifact under challenge).
  2. Set the depth. Check the idea against the c19 escalation signals above. Any hit → rigorous; none → lightweight. Say which you chose and why — the depth decision is part of the pass's record.
  3. Sweep the lenses read-only. Read the exported spec claim by claim, the frame's parked vagueness and resolved questions, and the actual surfaces the idea touches. Nothing mutates during the sweep; the only mutations are the deterministic moves that record findings.
  4. Route every finding through an existing move. Use the routing table below. Everything the agent proposes carries --origin llm and lands proposed — the pass cannot silently convert speculation into confirmed requirements.
  5. Let the human adjudicate. devague review lists every proposal with ids; devague confirm / devague reject / devague question --resolve are user-only decisions. This is the existing spec gate doing its job.
  6. Reconverge and re-export. devague converge, then devague export — the same dated spec file now carries the adjudicated findings and the pass's provenance (the exported spec renders scope entries in its Scope exploration section).
  7. On a clean pass, record the pass itself. A sweep that finds nothing still records which lenses and surfaces were examined (devague scope entries, one per lens/surface, e.g. challenge pass / concurrency lens: devague/store.py) and what residual uncertainty remains (park). A bare "no issues found" is not an outcome this skill produces.

Where findings land

Challenge keeps no parallel prose-only artifact — the frame is the record, and every output category from issue 73 has an existing deterministic move to land in:

Output category (issue 73) What it is Landing move
known facts something the pass established, with provenance capture --kind requirement / --kind decision / --kind boundary (--origin llm → lands proposed)
assumptions beliefs the frame leaned on unstated capture --kind assumption --origin llm, then pressure-test with interrogate --honesty / --hard-question / --contradicts
known unknowns / open questions articulated uncertainty question "<text>" when it needs a user decision; park --kind unknown_nonblocking|unknown_blocking when not decidable now
unexamined surfaces what this pass did not (or could not) look at devague scope "<surface>" --finding "<what was and wasn't examined, and why>"
residual surprise risk uncertainty that survives the pass park on the frame while speccing; devague plan risk --kind <kind> once the plan exists
resilience measures containment, rollback, recovery the surprise cost demands spec-side capture --kind requirement / --kind boundary; plan-side devague plan risk (see below)

Every finding names the lens and surface it came from (the challenge pass / <lens>: <surface> convention in scope entries; provenance citations inside claim text). That provenance bar is how the method hunts blind spots without encouraging speculative issue generation — a finding you cannot trace to something you actually read is speculation, not a finding.

Resilience placement — spec-side or plan-side

Where a resilience measure lands is a recorded decision:

resilience measures land in both spec and plan by nature: spec-side as requirement/boundary claims when they change what to build, plan-side as plan risks or tasks when they change how to build it — the skill coaches which is which (resolves q2)

— decision c18 in docs/specs/2026-07-15-challenge-skill.md.

The coaching: ask "does this change what ships, or how it gets built?" A rollback path the user needs, a fail-closed version check, a containment boundary — those change the product: capture them spec-side as requirement / boundary claims so the re-exported spec carries them. A staging sequence, a merge-order constraint, an uncertainty the workforce must build around — those change the build: land them plan-side via devague plan risk --kind <kind> (blocking or nonblocking, honestly chosen) once spec-to-plan seeds the plan, where the plan's convergence gate keeps blocking risks visible until resolved.

Hard rules (do not violate)

  • Never conclude there are no unknown unknowns. Not after a rigorous pass, not after a clean one. The only honest clean-pass output is a record of which lenses and surfaces were examined — devague scope entries — plus the residual uncertainty that remains — park — so the pass leaves durable provenance instead of a comforting absolute (issue 73 success criteria; the anti-fabrication contract in docs/llm-guidance.md).
  • LLM-origin findings stay proposed until the user confirms. Every finding the agent proposes carries --origin llm and lands proposed; only the user's confirm makes it real. The pass must not be able to silently convert speculation into confirmed requirements.
  • Findings route through existing deterministic moves only. capture, interrogate, question, park, devague scope, devague plan risk — nothing else. No parallel prose artifact, no new CLI verb, engine, or state model (issue 20; issue 73's stated preference). If it didn't land in a move, it didn't land.
  • Provenance on every finding. Name the lens and the surface it came from. If you didn't read it, don't claim it — same bar as scope.
  • Proportional, never skipped. Lightweight is the floor, not an exemption; any c19 escalation signal makes the rigorous form mandatory. Skipping the pass entirely is not a depth setting.
  • Not a fourth standing gate. The three human gates stay: exported spec, implementation split plan, final PR. Challenge output is adjudicated inside the existing spec gate — mirroring how deviate amends gate 2 rather than adding one.
  • The sweep is read-only until it routes. Reading spec, frame, and surfaces never edits files or state; probes run in scratch space. The only mutations are the recording moves themselves.

Worked example

Challenging the exported spec for a store-schema migration (illustrative slug store-schema-v3) — "migrations" and "any surface that can lose user data" are both c19 escalation signals, so the pass runs rigorous:

# Entry condition: think exported docs/specs/2026-07-15-store-schema-v3.md
# and no plan exists yet. Depth: rigorous (migration + data-loss signals).

# adjacent-systems lens: an older installed devague reads the same store
devague capture --origin llm --kind assumption "older installed devague binaries refuse a v3 store via the fail-closed schema_version check in devague/store.py"
devague interrogate c9 --origin llm --honesty "a v2-reading binary pointed at a v3 store exits with the version hint, not a traceback"
devague scope "challenge pass / adjacent-systems lens: devague/store.py schema_version gate" --finding "older binaries fail closed on v3; seeded the compat assumption" --seeds c9

# failure-mode lens: the migration can die halfway
devague capture --origin llm --kind requirement "migration writes to a temp file and renames — a killed run never leaves a half-written store"

# overlooked-actors lens: needs a user decision, not a guess
devague question "do mesh agents share one store, or does each checkout own its own?"

# reversibility lens: genuinely unknown, not decidable now
devague park "whether a v3->v2 downgrade path is ever needed" --kind unknown_nonblocking

# concurrency lens found nothing — record the clean pass, not a conclusion
devague scope "challenge pass / concurrency lens: devague/store.py + delivery_store.py" --finding "single-writer CLI, no locking today; clean pass — residual risk only if two agents ever share a checkout"

# --- HUMAN adjudicates: the existing spec gate at work ---
devague review
devague confirm c9 h4 c10
devague question --resolve q1 --decision "each checkout owns its own store"

# Reconverge and re-export — the SAME dated file, now carrying the pass
devague converge
devague export

# Residual risk that changes HOW to build lands plan-side once
# spec-to-plan seeds the plan:
devague plan new --frame store-schema-v3
devague plan risk "two agents sharing a checkout could interleave store writes mid-migration" --kind unknown_nonblocking

Every finding above is traceable to a lens and a surface; the clean lens is recorded as examined rather than silently dropped; and nothing the agent proposed became confirmed without the human's confirm.

After the pass — hand off to spec-to-plan

Once the frame reconverges and the same dated spec file is re-exported, the pass is done — the examined surfaces, residual uncertainty, and adjudicated findings all live in frame state and render into the spec artifact. Continue with spec-to-plan as usual: the plan seeds from the challenged frame, and any residual surprise risk you routed plan-side lands via devague plan risk as first-class plan state. If a surprise still gets through mid-fan-out, that is deviate's job — and every approved dN deviation record is evidence for what the next challenge pass's lenses should look harder at.

Provenance

This is a first-party skill — its origin is agentculture/devague, the seventh in the outbound family after scope, think, spec-to-plan, assign-to-workforce, deviate, and summarize-delivery, sitting third in flow order as the blind-spot discovery leg between think and spec-to-plan. guildmaster pulls it from here and broadcasts it to the AgentCulture mesh; because devague is upstream, it is never re-vendored back from guildmaster's re-broadcast copy. The cite, don't import policy still holds: downstream repos copy it, they don't symlink or depend on it. See docs/skill-sources.md.

04 同仓库更多 Skills

agentculture
agentculture/culture-nodes

sample

Fixture skill proving the .qwen/skills leg carries at least one skill.

★ 3
agentculture
agentculture/culture-nodes

sample

Fixture skill proving the .qwen/skills leg carries at least one skill.

★ 3
agentculture
agentculture/culture-nodes

sample

Fixture skill proving the .qwen/skills leg carries at least one skill.

★ 3
agentculture
agentculture/culture-nodes

think

Think a vague feature idea into a buildable spec by working backwards (the idea→spec leg; drives the `devague` CLI). Start from the announcement ("pretend it shipped"), capture and classify claims, interrogate them with honesty conditions and hard questions, park open vagueness as a first-class object, and export a spec only once the frame *converges*. Use when the user says "think this through", "spec this", "work backwards", "turn this idea into a spec", "announcement frame", or "devague", or when a feature request is too vague to build yet. Explore scope first when the idea touches an existing codebase — the sibling `scope` skill is the optional opening leg ahead of `think`. Once a spec exports, hand off to the sibling `spec-to-plan` skill to turn it into a plan. Authored and maintained in agentculture/devague (origin = devague); guildmaster pulls this skill from here and broadcasts it to the AgentCulture mesh — it is NOT vendored from guildmaster like the other skills here.

★ 3
agentculture
agentculture/culture-nodes

summarize-delivery

Close the loop after an `assign-to-workforce` run by turning what actually happened into an accountability artifact — planned versus actual delivery, mid-work decisions, plan drift, evidence-backed delivery claims, and remaining work. The plan the user confirmed is the contract; this skill records where execution obeyed it, where it changed, and what is genuinely safe to claim as delivered. Runs on complete, partial, AND failed runs — failure is reported faithfully, never smoothed over. Use when the user says "summarize delivery", "delivery summary", "wrap up", "close the loop", "what did we actually ship", or "plan versus actual", or after `assign-to-workforce` merges (or fails to merge) a plan's waves. Authored and maintained in agentculture/devague (origin = devague); guildmaster pulls this skill from here and broadcasts it to the AgentCulture mesh — it is NOT vendored from guildmaster like the inbound skills here.

★ 3
agentculture
agentculture/culture-nodes

spec-to-plan

Turn a converged devague spec into a buildable plan by working forwards (the spec→plan leg; drives the `devague plan` CLI group). Seed a plan from a converged frame, add tasks that collectively cover every coverage target (the frame's confirmed claims + honesty conditions), give each task acceptance criteria and an honest dependency order, park genuine unknowns as first-class risks, and export a plan only once it *converges*. Use when the user says "spec to plan", "stp", "turn this spec into a plan", "plan this spec", "make a build plan", or after the `think` skill exports a spec. Authored and maintained in agentculture/devague (origin = devague); guildmaster pulls this skill from here and broadcasts it to the AgentCulture mesh — it is NOT vendored from guildmaster like the other skills here.

★ 3