Assess

Assess

Host-neutral workflow. Invoke by skill name (assess). Claude Code slash routing, YAML hook loaders, and .claude/chain live in references/claude-code.md.

Comprehensive assessment skill for answering "is this good?" with structured evaluation, scoring, and actionable recommendations.

🎯 Quick Start

assess backend/app/services/auth.py
assess our caching strategy
assess --model=opus the current database schema
assess frontend/src/components/Dashboard

Effort levels (CC 2.1.111+ adds xhigh)

xhigh silently falls back to high on a model that does not implement it: no error, no log line. doctor Category 14 reports this, and only when it can positively prove the active model lacks the tier.

Argument Resolution

Step 0: resolve a conversational reference first

$ARGUMENTS is often not a path. For a bare pronoun or deictic (them, this, that, these, they, same, the above, the last one, what we just did) or an empty target after flags are stripped, the subject is in the conversation. Read back for the NEAREST concrete one (a file just discussed, a diff or PR just opened, a component just investigated) and announce the resolution in one line, so a wrong guess costs a correction rather than a turn: "Reading 'them' as the 3 pretool guards we just probed; say otherwise and I'll switch."

Refusing is the bug, not the safe option. Asking "what does this refer to?" when the previous turn named the subject burns a round-trip re-deriving what is already on screen. Measured 2026-08-28: the operator sent assess them throguhly one message after "bug in orchestkit hooks", mid-investigation of pretool/bash/dangerous-command-blocker, and this skill replied that "them" had "no antecedent anywhere in this conversation". It had two.

Ask only when the conversation is genuinely empty (a fresh session opening with a bare pronoun). Every other case: resolve and announce.

Not unique to this skill: verify, cover, fix-issue, review-pr and implement all read $ARGUMENTS as a literal path or topic, and no skill mentions resolving a reference. Tracked separately; this one fixes its own door.

TARGET = "$ARGUMENTS"  # Full argument string, e.g., "backend/app/services/auth.py"
# $ARGUMENTS[0] is the first token (CC 2.1.59 indexed access)

# Model override detection (CC 2.1.72)
MODEL_OVERRIDE = None
for token in "$ARGUMENTS".split():
    if token.startswith("--model="):
        MODEL_OVERRIDE = token.split("=", 1)[1]  # "opus", "sonnet", "haiku", "fable"
        TARGET = TARGET.replace(token, "").strip()

Pass MODEL_OVERRIDE to all Agent() calls via model=MODEL_OVERRIDE when set. Accepts symbolic names (opus, sonnet, haiku, fable on harnesses whose Agent tool lists it; note fable is premium API spend after 2026-07-12) or full IDs (claude-opus-5-5) per CC 2.1.74.

Switching to Opus via /model (CC 2.1.144+): /model now changes the model for the current session only, so picking Opus for an assess run no longer persists past it. Press d in the picker only to set a default for new sessions.

Effort detection (CC 2.1.120+)

$CLAUDE_EFFORT is the primary signal. CC 2.1.120 sets this env var from /effort or the model picker. --effort= token in $ARGUMENTS is the explicit override fallback (also covers older CC).

# Read env first (CC 2.1.120+), then check explicit override
EFFORT = os.environ.get("CLAUDE_EFFORT")  # "low" | "medium" | "high" | "xhigh" | None
for token in "$ARGUMENTS".split():
    if token.startswith("--effort="):
        EFFORT = token.split("=", 1)[1]   # explicit override wins
        TARGET = TARGET.replace(token, "").strip()
EFFORT = EFFORT or "high"  # default when CC < 2.1.120 and no flag

Use EFFORT to gate dimension count, agent count, and the optional xhigh uncertainty pass — see "Effort levels" table above. On CC < 2.1.120 the env var is unset; the explicit --effort= override is the only path. doctor Category 14 reports a provably unsupported xhigh request.

STEP -1: MCP Probe + Resume Check

Load: Read("../chain-patterns/references/mcp-detection.md")

# 1. Probe MCP servers (once at skill start)
# memory is alwaysLoad in .mcp.json (CC 2.1.121+, #1541) — probe below kept as fallback for older CC:
ToolSearch(query="select:mcp__memory__search_nodes")

# 2. Store capabilities
Write(".claude/chain/capabilities.json", {
  "memory": probe_memory.found,
  "skill": "assess",
  "timestamp": now()
})

# 3. Check for resume
state = Read(".claude/chain/state.json")  # may not exist
if state.skill == "assess" and state.status == "in_progress":
    last_handoff = Read(f".claude/chain/{state.last_handoff}")

Phase Handoffs

STEP 0: Verify User Intent with AskUserQuestion

BEFORE creating tasks, clarify assessment dimensions:

AskUserQuestion(
  questions=[{
    "question": "What dimensions to assess?",
    "header": "Dimensions",
    "options": [
      {"label": "Full assessment (Recommended)", "description": "All dimensions: quality, maintainability, security, performance"},
      {"label": "Code quality only", "description": "Readability, complexity, best practices"},
      {"label": "Security focus", "description": "Vulnerabilities, attack surface, compliance"},
      {"label": "Quick score", "description": "Just give me a 0-10 score with brief notes"}
    ],
    "multiSelect": false
  }]
)

Based on answer, adjust workflow (passed to Phase 2 as focus):

• Full assessment (full): All 7 phases, parallel agents, dimension subset scaled by effort
• Code quality only (quality): Skip security and performance phases
• Security focus (security): Prioritize security-auditor agent
• Quick score (quick): Single pass, brief output

STEP 0b: Select Orchestration Mode

Load details: Read("references/orchestration-mode.md") for env var check logic, Agent Teams vs Agent Tool comparison, and mode selection rules.

🚨 Task Management (CC 2.1.16)

# 1. Create main task IMMEDIATELY
TaskCreate(
  subject="Assess: {target}",
  description="Comprehensive evaluation with quality scores and recommendations",
  activeForm="Assessing {target}"
)

# 2. Create subtasks for each assessment phase
TaskCreate(subject="Understand target and gather context", activeForm="Understanding target")   # id=2
TaskCreate(subject="Discover scope and build file list", activeForm="Discovering scope")        # id=3
TaskCreate(subject="Rate quality across 6 dimensions", activeForm="Rating quality")             # id=4
TaskCreate(subject="Analyze pros and cons", activeForm="Analyzing pros/cons")                   # id=5
TaskCreate(subject="Compare alternatives", activeForm="Comparing alternatives")                 # id=6
TaskCreate(subject="Generate improvement suggestions", activeForm="Generating suggestions")     # id=7
TaskCreate(subject="Compile assessment report", activeForm="Compiling report")                  # id=8

# 3. Set dependencies for sequential phases
TaskUpdate(taskId="3", addBlockedBy=["2"])  # Scope needs target understanding
TaskUpdate(taskId="4", addBlockedBy=["3"])  # Rating needs scoped file list
TaskUpdate(taskId="5", addBlockedBy=["4"])  # Pros/cons needs quality scores
TaskUpdate(taskId="6", addBlockedBy=["4"])  # Alternatives need quality scores
TaskUpdate(taskId="7", addBlockedBy=["5", "6"])  # Suggestions need analysis
TaskUpdate(taskId="8", addBlockedBy=["7"])  # Report needs suggestions

# 4. Update status as you progress
TaskUpdate(taskId="2", status="in_progress")  # When starting
TaskUpdate(taskId="2", status="completed")    # When done — repeat for each subtask

🔄 Workflow Overview

Phase 1: Target Understanding

Identify what's being assessed and gather context. TARGET here is the value Step 0 already resolved, which is not necessarily what the user typed.

# PARALLEL - Gather context
Read(file_path=TARGET)                                   # only when TARGET is a path
Grep(pattern=TARGET, output_mode="files_with_matches")   # topic or symbol
mcp__memory__search_nodes(query=TARGET)                  # past decisions

Read failing is NOT a reason to stop. A target resolved from the conversation is usually a subject rather than a filename ("the three pretool guards", "today's hook fixes"), so the Read misses and the Grep plus the conversation carry the context. Treat a failed Read as "this is a topic, not a path" and continue to Phase 1.5, which discovers the real file list anyway.

Phase 1.5: Scope Discovery

Load Read("references/scope-discovery.md") for the full file discovery, limit application (MAX 30 files), and sampling priority logic. Always include the scoped file list in every agent prompt.

Progressive Output (CC 2.1.76)

Output results incrementally as each evaluation phase completes:

The Phase 2 workflow returns once, so show every dimension's score from its result and lead with priorityConcerns (any dimension below 4/10) as a concern needing user attention. Per-agent streaming applies only on the Agent tool fallback.

Phase 2: Quality Rating (6 Dimensions)

Rate each dimension 0-10 with weighted composite score. Load Read("../quality-gates/references/unified-scoring-framework.md") for dimensions, weights, grade interpretation, and per-dimension criteria. Load Read("references/quality-model.md") for assess-specific overrides.

Do NOT hand-roll the assessors. Run the executor, which owns Phases 2 and 2.5:

result = Workflow(
  scriptPath="${CLAUDE_SKILL_DIR}/workflows/assess-fanout.js",
  args={"target": TARGET, "effort": EFFORT, "focus": FOCUS,   # FOCUS from STEP 0
        "mode": "comparison" if COMPARING else "default",     # quality-model.md
        "domain": "frontend" or "backend",                     # picks the performance engineer
        "scopeFiles": SCOPE_FILES,                             # Phase 1.5 list
        "projectContext": MEMORY_CONTEXT,                      # Phase 1 memory search
        "rubric": Read("rubric.json"), "modelOverride": MODEL_OVERRIDE,
        "feature": FEATURE})
Write(".claude/chain/02-evaluation.json", result)

The script owns the mechanics, not the prose. It picks the assessors from focus and effort (security first), gives each a score schema that demands file:line evidence, sends every decision-bearing score to blind refuters (Phase 2.5), and computes the weighted composite, grade, rubric verdict and blockers. A score with no file:line evidence counts as unscored, a repeated dimension keeps only its first entry, and a selected dimension with a min_blocker that nobody scored is a blocker. It returns composite, grade, verdict, blockers (producer basis), postRefutation, chainVerdict, chainVerdictIfConfirmed, revisions, confirmationNeeded, manualReview, advisory, priorityConcerns, quickWins, unscored, rejectedDimensions, unassessed, dimensions, ledger and reasons. It never asks and never writes: those stay in this shell.

Fallback (no Workflow tool, ORCHESTKIT_FORCE_TASK_TOOL=1, or the cross-model lane below): Read("references/agent-spawn-definitions.md") for Agent tool and Agent Teams spawns, then run Phase 2.5 by hand.

Composite Score: Weighted average of the scored dimensions (see quality-model.md).

Phase 2.5: Adversarial Refutation (effort-gated)

The assessor that scores a dimension is also its only judge, a self-preferential bias. A separate blind refuter forms its own band for each decision-bearing score. Effort gate: low/medium skip it; high runs up to 4 single advisory refuters (no auto-swing); xhigh runs a 3-refuter majority that revises to the near band edge. On the Workflow path the script already ran it; the shell finishes it:

1. Write the returned ledger to .claude/chain/02b-refutation.json (engine section 10).
2. Re-open every cited file:line in revisions (engine section 3). A citation that does not hold reverts that dimension to its producer score; recompute the composite with the returned weights.
3. If confirmationNeeded is non-empty, AskUserQuestion before using chainVerdictIfConfirmed; otherwise use chainVerdict. Refutation alone never raises a score or flips fail to pass (engine section 7).
4. List manualReview dimensions as "not independently refuted", and surface every advisory overturn at high effort.

Protocol and assess bindings: Read("references/adversarial-refutation.md") (loads the shared engine ../../shared/rules/adversarial-refutation.md). Producer findings must first pass the evidence-replay gate before entering any score or verdict: Read("../../shared/rules/evidence-replay.md").

Cross-model refuter (optional, provenance-labeled, cost-gated)

When ORK_ALT_MODEL_CMD is configured and effort is high/xhigh, one quorum slot per high-weight or boundary-adjacent dimension score can route to a non-Claude model (Codex/GPT) for diverse failure modes. Off by default; substitutes one same-model slot, stamps refuter_model for provenance, cannot silently raise the grade (engine §7), owns no credentials/egress (shells out via ORK_ALT_MODEL_CMD, matches the egress guard #2533), and degrades to same-model on an absent command. Shares the review-pr operational doc: Read("../review-pr/references/cross-model-refuter.md").

The workflow does not run this lane (a script cannot shell out). When the user wants it, choose the Agent tool fallback before Phase 2 and run Phases 2 and 2.5 there.

Refuters are ALWAYS isolated spawns with no team_name, fed only the dimension and the scoped files: no producer score, identity, or prose. Keep the producer-basis score AND a labeled post-refutation score.