skill-pipeline（技能管道）

Skill Pipeline

The conductor, not a player. This skill classifies tasks, selects the pipeline variant, calibrates depth, and orchestrates handoffs between skills. It produces no artifacts of its own — its output is routing decisions that activate other skills.

Task Classification

On every coding task, classify before acting. Evaluate scope signals and map to a task class.

Input signals: file count, task description, existing plan/handoff files, batch indicators, CI environment.

Task received
  │
  ├─ Trivial (typo, rename, version bump)
  │  → No skills. Just do it.
  │
  ├─ Small (isolated fix, single-file, <10 logic lines)
  │  → verify-gate + simplify-and-harden
  │
  ├─ Medium (feature in known area, 2-5 files)
  │  → intent-framed-agent + verify-gate + simplify-and-harden
  │
  ├─ Large (complex refactor, new architecture, unfamiliar codebase, high-risk logic)
  │  → Full standard pipeline
  │  → Recommend /plan-interview before starting
  │
  ├─ Long-running (multi-session, high context pressure, prior handoff exists)
  │  → Full standard pipeline with context-surfing as critical skill
  │
  └─ Batch (multiple features from spec, 5+ discrete tasks, issue triage)
     → Team-based pipeline (agent-teams-simplify-and-harden)

When uncertain, start with Medium. Add skills if drift or quality issues appear mid-task.

For detailed heuristics, edge cases, and examples: read references/classification-rules.md.

Pipeline Selection

Route task class to the right variant:

Heuristic: Standard pipeline for depth (single complex feature). Team-based pipeline for breadth (batch of tasks). CI pipeline when CI=true or GITHUB_ACTIONS=true.

Activation Sequences

Standard Pipeline (Inner Loop)

pre-flight-check (SessionStart hook — surfaces prior learnings)
  → classify
  → (recommend /plan-interview if Large or Long-running)
  → intent-framed-agent (at planning-to-execution transition)
  → context-surfing (auto-activates when intent frame + plan exist; concurrent with intent monitoring)
  → [IMPLEMENTATION]
  → verify-gate (compile + test + lint; fix loop if red)
  → simplify-and-harden (post-completion, if non-trivial diff)
  → self-improvement (on errors, corrections, or S&H learning candidates)

Skill-by-class activation:

Team-Based Pipeline

classify (Batch)
  → (recommend /plan-interview if no spec exists)
  → agent-teams-simplify-and-harden
    ├─ Team lead emits Intent Frame #1
    ├─ Phase 1: parallel implementation agents
    ├─ verify-gate (compile + test + lint)
    ├─ Phase 2: parallel audit agents (simplify, harden, spec)
    ├─ Fix loop (up to 3 audit rounds)
    └─ Learning loop output
  → self-improvement

CI Pipeline

classify (CI detected)
  → simplify-and-harden-ci (headless scan, PR changed files only)
  → self-improvement-ci (pattern aggregation, promotion recommendations)

Outer Loop Pipeline

The outer loop runs across sessions, not within them. Trigger on cadence (weekly, sprint boundary) or when pre-flight-check surfaces promotion-ready patterns.

learning-aggregator (read .learnings/, find patterns, rank promotion candidates)
  → harness-updater agent (apply promotions to CLAUDE.md, AGENTS.md, copilot-instructions.md)
  → eval-creator (create permanent test cases from promoted patterns)
  → eval-creator run (regression check on all existing evals)

When to trigger the outer loop:

• Weekly: recommended minimum cadence
• Sprint boundary: after a burst of sessions
• When pre-flight-check reports promotion-ready count > 3
• After a significant incident or recurring failure
• Manually: user invokes /learning-aggregator

Outer loop is always human-gated. learning-aggregator produces a gap report. harness-updater shows diffs for approval. No automatic writes to instruction files without human review.

Depth Calibration

Not just which skills — how deep each goes:

Handoff Rules

Artifacts flow between skills. The orchestrator ensures each skill receives what it needs.

Key handoffs:

1. Plan file (docs/plans/plan-NNN-<slug>.md) — produced by plan-interview, consumed by intent-framed-agent (context), context-surfing (wave anchor), agent-teams (task extraction).
2. Intent Frame — produced by intent-framed-agent, consumed by context-surfing (wave anchor strengthening). Copied into handoff files on drift exit.
3. Handoff file (.context-surfing/handoff-[slug]-[timestamp].md) — produced by context-surfing on drift exit, consumed by next session for resume.
4. Verify-gate signal — produced by verify-gate (pass/fail + diagnostics), consumed by simplify-and-harden (only activates after green gate) and fix loop (on failure).
5. Learning candidates (learning_loop.candidates) — produced by simplify-and-harden and agent-teams, consumed by self-improvement for pattern tracking.
6. Learning entries (.learnings/*.md) — produced by self-improvement, consumed by learning-aggregator for cross-session analysis and by pre-flight-check at session start.
7. Gap report — produced by learning-aggregator, consumed by harness-updater agent for promotion and eval-creator for test case generation.
8. Eval cases (.evals/cases/*.md) — produced by eval-creator, consumed by regression runs and surfaced by pre-flight-check.

Precedence: If context-surfing and intent-framed-agent both fire simultaneously, context-surfing's exit takes precedence. Degraded context makes scope checks unreliable.

For the full artifact/signal/budget table: read references/handoff-matrix.md.

Decision Points

The orchestrator intervenes at these moments:

Task Arrival

Classify the task. Select pipeline variant and depth. Emit routing decision. If Large/Long-running, recommend /plan-interview. If Batch, recommend team-based variant.

Plan Approval

When user approves a plan from plan-interview, flow directly into the execution stage — no separate "should I proceed?" prompt. This means activating intent-framed-agent to emit an Intent Frame. The intent frame itself still requires user confirmation before coding begins (that confirmation is part of intent-framed-agent, not an extra gate). Populate task tracking with checklist items.

Planning-to-Execution Transition

When no plan-interview was used and the user signals readiness ("go ahead", "implement this", "let's start"), activate intent-framed-agent. Emit Intent Frame. Wait for user confirmation of the frame before coding.

Implementation Complete

Activate verify-gate to run compile, test, and lint checks. If any fail, enter the fix loop (up to 3 attempts per phase). Once all checks pass and the diff meets the non-trivial threshold (see references/classification-rules.md), activate simplify-and-harden. If the diff is trivial, signal completion directly after verify-gate passes.

Drift Detected

If context-surfing fires a drift exit, stop execution. Write handoff file. If the task was classified below Large, consider re-classifying upward for the next session.

Session Resume

Check for handoff files in .context-surfing/. If found, read completely. Re-establish context from handoff. Re-classify if needed. Resume from recommended re-entry point.

Overrides

Users can override any routing decision:

• Force depth: depth=small / depth=large — override classification
• Force variant: variant=teams / variant=standard — override pipeline selection
• Skip review: --no-review — skip simplify-and-harden
• Force planning: invoke /plan-interview on any task regardless of classification
• Skip all skills: user says "just do it" on a non-trivial task — respect the override

Re-classification

Tasks can change class mid-execution. Watch for:

• Escalation signals: scope expanded beyond original estimate, many more files affected than expected, context-surfing drift exit, intent-framed-agent detects significant scope change
• De-escalation signals: task turns out simpler than planned, plan reveals minimal changes needed

When re-classification is warranted:

1. Note the signal that triggered re-classification
2. Adjust active skills (add or remove pipeline stages)
3. If escalating to Large: recommend /plan-interview if no plan exists
4. If de-escalating: drop unnecessary stages, proceed with lighter pipeline

Anti-Patterns

• Do NOT re-implement skill logic. This skill classifies and dispatches. Each individual skill owns its own procedure.
• Do NOT auto-invoke plan-interview. It is a human gate. Recommend it; let the user decide.
• Do NOT create a third monitoring layer. During execution, intent-framed-agent monitors scope and context-surfing monitors quality. The orchestrator dispatches at decision points, then gets out of the way.
• Do NOT override individual skill guardrails. If simplify-and-harden has a 20% budget cap, the orchestrator respects it.
• Do NOT force skills on trivial tasks. The pipeline exists to help, not to slow down simple work.

Pipeline Variant Details

For complete step-by-step walkthroughs of each variant including hybrid scenarios and session resume: read references/pipeline-variants.md.