AgentSkillsCN

evidence-discipline

教授正确的会话生命周期:declare_intent → map_context → compact_session。涵盖会话的启动、更新与关闭。

SKILL.md
--- frontmatter
name: "evidence-discipline"
description: "Enforces evidence-based claims. Prove with command output before concluding. Run tests before claiming success."
triggers:
  - "Before making completion claims"
  - "When verification is needed"
  - "After running tests or builds"
version: "2.6.0"

Evidence Discipline

Core principle: Never claim, always prove. Never accept, always validate.

When to Use

  • Claiming work is "done" or "fixed"
  • Accepting user instructions that conflict with existing architecture
  • Validating subagent results
  • Deciding if a test, build, or deployment succeeded
  • User says "it works" but no evidence shown

The Evidence Chain

Every claim needs a verification path. HiveMind gives you concrete tools:

Before Claiming Completion

bash
# 1. Does the code compile?
npx tsc --noEmit

# 2. Do tests pass?
npm test

# 3. Is the hierarchy consistent?
node bin/hivemind-tools.cjs validate chain

# 4. Are all source files accounted for?
node bin/hivemind-tools.cjs source-audit

# 5. Does the full ecosystem check pass?
node bin/hivemind-tools.cjs ecosystem-check

Before Accepting Conflicting Instructions

bash
# What does the trajectory say we're doing?
node bin/hivemind-tools.cjs state hierarchy

# What decisions did we already make?
recall_mems({ shelf: "decisions" })

# What does the plan say?
# Read the active plan file from hierarchy cursor

# What does git history show?
git log --oneline -10
git log --grep="<stamp>" --oneline

After Subagent Returns

code
Subagent says "Done"
    │
    ├── Does the result text contain failure signals?
    │   (failed, error, couldn't, unable, blocked, partially, skipped)
    │   YES → export_cycle({ outcome: "failure" | "partial", findings: "..." })
    │   NO  → continue
    │
    ├── Does the result describe what was ACTUALLY done?
    │   VAGUE → ask for specifics before proceeding
    │   SPECIFIC → verify independently
    │
    ├── Can you verify independently?
    │   RUN: npm test / tsc / grep for expected changes
    │   PASS → export_cycle({ outcome: "success", findings: "..." })
    │   FAIL → export_cycle({ outcome: "failure", findings: "..." })
    │
    └── Update hierarchy
        map_context({ level: "action", status: "complete" | "blocked" })

Reward — What Evidence Gets You

ActionIntelligence Gained
export_cycle after every subagentFull decision trail in hierarchy tree + mems
save_mem({ shelf: "decisions" })Any future agent can recall WHY you chose X over Y
git log --grep="<stamp>"Commits linked to hierarchy nodes by timestamp
grep -r "<stamp>" .hivemind/Full evidence chain: tree + mems + anchors + archives
node bin/hivemind-tools.cjs session trace <stamp>One command, complete picture

Consequence — What Skipping Costs

SkipCost
Claim "done" without npm testBroken code deployed, trust lost
Accept instruction without checking trajectoryArchitecture contradicted, rework needed
Ignore subagent failure signalsFalse completion, bug surfaces later at 10x cost
No save_mem for key decisionNext session re-debates the same choice

Red Flags — You're About to Skip Evidence

ThoughtReality
"The subagent said it works"Subagents hallucinate success. Verify independently.
"The user confirmed it"Users can be wrong. Run the verification commands.
"I tested it mentally"Mental models miss edge cases. Run actual commands.
"It's obvious this is correct"Obvious things break. One npm test takes 3 seconds.
"I'll verify at the end"Compaction may erase the context. Verify NOW.
"The error is unrelated"Prove it's unrelated. git diff + test output = evidence.

The Minimum Evidence Bar

Nothing is "done" without ALL of these:

  1. Verification command ran (test/build/lint — at least one)
  2. Output inspected (not just "0 exit code" — read the output)
  3. Hierarchy updated (map_context status: "complete")
  4. If subagent was involved: export_cycle called with accurate outcome