Prior work check

Prior Work Check

Overview

Before writing new code, confirm the codebase doesn't already do this. The single most common cause of waste in mature codebases is re-implementing a capability that already exists under a different name. Three deterministic pre-task queries run against MTK's file-based artifacts (lessons, specs, code index) before a spec is approved or a planner opens a batch.

Three deterministic queries run against the repo. Each must complete before the spec is handed to the approval gate or the planner begins task breakdown.

When To Use

• During spec-driven-development step 8c, before the implement Phase 2.5 approval gate
• Before planning-and-task-breakdown opens a batch
• During implement Phase 0.7 when a pre-written plan is supplied as input — run the existing-plan reconciliation mode below so already-implemented batches are not re-run
• When the engineer says "let's add an X" and X sounds generic (cache, retry, validator, formatter, normalizer)
• When the change manifest names files in an area you haven't touched recently

When NOT To Use

• Typo fixes, single-line config tweaks
• Bug fixes where the file in scope is already named
• Changes inside a clearly-isolated feature module with no shared surface

Workflow

Query 1 — search_prior_work

Look for an existing implementation of the capability under any name.

1. Extract 2-4 capability keywords from the spec summary and success_criteria. Verbs + nouns ("issue token", "retry policy", "currency formatter").
2. Run, in parallel:
   • grep -rn -i "<keyword>" --include="*.cs" --include="*.py" --include="*.ts" --include="*.tsx" --include="*.js" src/ (or the active stack's source roots — read .claude/tech-stack).
   • If CODE_INDEX.md exists at the repo root, query it with bash scripts/query-code-index.sh find "<keyword>" (case-insensitive row search that prints each match prefixed with its ## <Domain> section) rather than a raw grep. If the script is absent (repo bootstrapped before v7.17) or the keyword may live outside table rows (domain headings, prose), fall back to grep -i "<keyword>" CODE_INDEX.md.
   • If an MCP code-index tool is available (e.g. mcp__gitnexus__query, mcp__ast-index__*), prefer it over raw grep.
3. For each hit, report:
   • file:line
   • one-line summary of what it does
   • whether it can be reused, extended, or is genuinely different

Query 2 — get_constraints

Surface the rules and lessons that govern this area.

1. Read tasks/lessons.md and .claude/lessons/personal.md (if present). Grep for the capability keywords and the touched module names.
2. Read .claude/references/architecture-principles.md if present and grep for tags matching the touched slice. Note any [EXTRACTED] principles that constrain the design — these block contradicting changes. If the file is absent (it is an optional generated artifact), note constraints-unavailable in the report instead.
3. Read CLAUDE.md and any .claude/rules/*.md file whose name matches the slice (e.g. git-workflow.md for git-touching work).
4. Report every constraint discovered, with severity:
   • block — EXTRACTED principle or explicit lesson "never do X"
   • flag — INFERRED ≥0.7 principle or "prefer X"
   • note — soft preference or general guidance

Query 3 — get_risk_profile

Identify the risk surface this change crosses.

1. Map the change_manifest[].path entries to risk categories:
   • regulated — auth, payments, audit, PII, financial-state
   • boundary — public API, exported contracts, migrations, message bus
   • shared — utilities used by 3+ slices
   • isolated — single-slice internal change
2. For each non-isolated category, list the reviewers/agents that must run in Phase 4 (compliance-reviewer always for regulated; architecture-reviewer for boundary; etc.).
3. Confirm security_impact in the spec matches the highest risk category. If security_impact: none but a regulated path is touched, emit a BLOCKING finding — the spec is mis-classified.

Existing-Plan Reconciliation Mode

Invoked by implement Phase 0.7 when the input is a pre-written plan (supplied by the engineer) rather than a fresh request. A supplied plan may be stale — some batches were implemented since it was written. Re-implementing done work is waste and a drift risk, so reconcile the plan against current code before the Phase 2.5 gate:

1. Run Query 1 (search_prior_work) as usual to catch capability duplication.
2. Per-batch satisfied-check (read-only). For each batch in the supplied plan, take its acceptance / verification / linked success_criteria and probe current code for whether it is already satisfied:
   • grep for the batch's target symbols and change_manifest paths (and any public-contract signatures) and read the relevant lines;
   • where a batch has a runnable, non-mutating verification, note whether it already passes. Never mutate state to find out.
   • Classify each batch: already-satisfied | partially-done | not-started, each with file:line evidence.
3. Emit a reconciliation table (batch id → classification → evidence). The orchestrator then, before Phase 2.5: marks already-satisfied batches complete/skipped in the JSON sidecar and ticks them in tasks/todo.md with a one-line note naming what was pruned; surfaces partially-done batches for the engineer to re-scope. Never silently re-run a done batch, and never edit source here — this mode only classifies and reports.

Output

Emit one fenced block summarizing all three queries:

### Prior Work Check

**Query 1 — search_prior_work**
- match: src/Auth/TokenService.cs:42 — already issues JWTs. **Reuse, do not
  reimplement.** Spec should be revised to extend, not create.
- (no other matches)

**Query 2 — get_constraints**
- block: tasks/lessons.md L88 — "never store raw tokens; always Argon2id hash"
- flag: architecture-principles.md [INFERRED:0.85] — "auth changes go via
  AuthSlice/Handlers"

**Query 3 — get_risk_profile**
- regulated: src/Auth/*.cs — compliance-reviewer required in Phase 4
- spec declares `security_impact: low` → **mis-classified, should be `high`**

Verdict

• Any BLOCK finding (existing implementation, EXTRACTED contradiction, mis-classified security impact) → BLOCKED. Spec must be revised before the approval gate runs.
• Any FLAG → PROCEED_WITH_CHANGES — surface to the engineer; they decide whether to amend the spec or justify the deviation in writing.
• Only NOTE findings (or none) → PASS.

Integration

• spec-driven-development runs this skill at step 8c, before the implement Phase 2.5 approval gate. A BLOCKED verdict means the spec is not eligible for approval.
• planning-and-task-breakdown runs this skill again if the spec was approved more than a session ago — risk profile and prior work may have shifted.
• implement Phase 0.7 runs this skill in existing-plan reconciliation mode when a pre-written plan is the input, so stale (already-implemented) batches are pruned before the approval gate instead of being re-run.
• code-simplification --audit-duplicates uses Query 1's grep recipe to find duplicate capabilities across the codebase.

Rules

• Three queries always run — no skipping "because the area is obvious".
• A grep miss is not the same as a confirmed absence. If grep finds nothing but the capability is generic ("validator", "formatter"), widen the search with synonyms before declaring PASS.
• This skill does not edit files. It only reports.
• Findings are deterministic where possible. Don't downgrade severity based on convenience.

Common Rationalizations

See .claude/skills/context-engineering/SKILL.md for the shared table. Prior-work-specific traps: "I already grepped during planning" (you grepped for the spec's chosen name — re-grep for the capability under other names), "there's nothing in lessons.md about this" (re-read with synonyms — lessons are written in human prose, not your chosen keywords), and "security_impact is conservative enough" (regulated paths drive the floor; never the engineer's gut feel).

Red Flags

• Skipped Query 1 because "I already know there's no implementation"
• Query 2 read lessons.md but didn't grep architecture-principles.md
• Query 3 produced regulated matches but the spec's security_impact wasn't reclassified
• Verdict PASS with one or more BLOCK findings hidden in the body

Verification

• All three queries ran with output captured
• At least one synonym/alternate keyword was tried in Query 1
• Lessons AND architecture principles AND CLAUDE.md were consulted in Query 2
• Every path in change_manifest was classified in Query 3
• security_impact was compared against Query 3's highest category
• Verdict matches the highest-severity finding