work-gaps

name	work-gaps
description	After one or more discovery or analysis rounds, explicitly asks "What is still missing?" Modalities not swept, claims not verified, sources not read, edge cases not considered, downstream effects not traced, assumptions not checked. The output becomes the seed for the next round of work (more angles, more verification, more research). Use near the end of Research, Review, or Understand phases, or before declaring a task complete. Invoke via /work-gaps.
metadata	{"short-description":"Explicit 'what did we miss?' critic after discovery work"}

/work-gaps — Gap and Omission Finder

Implements the Completeness critic pattern — the final gate of the quality system: after find → dedup → verify, it asks "what did the LIST / SCOPE itself never cover?" and emits a machine-checkable clearedToClose verdict the orchestrator can branch on.

When to use

• After a sweep, review round, or design phase, before you are ready to say "we're done".
• You want a deliberate, last-mile check for the things that systematic processes tend to miss (the modality you forgot, the claim that was assumed true, the downstream consumer you didn't think about).

It is cheap insurance against "we covered everything important" when you actually covered everything we thought of.

Workflow shape — PARALLEL / BARRIER gate (default single-pass)

This skill is the closing gate of the canonical Review harness (find → dedup → work-refute → work-gaps; see workflow-mode.md §2).

• Default: single-pass critic. One critic run over the reconstructed scope is enough for most work, and it is cheap relative to the work it gates.
• Scale-up variant: PARALLEL / BARRIER fan-out. For large or high-stakes scopes, fan out one blind sub-agent per omission category (modality-missing, unverified-claim, downstream, assumption, cross-cutting, …), each unaware of the others' angle so they don't collapse onto the same gaps. Then barrier-collect all results and dedup the union against SEEN (everything ever discovered, including verifier-rejected items — L4) before emitting gaps. The barrier is justified here by a real cross-item dependency: the union must be deduped before the close decision.

Why a barrier and not a pipeline: the clearedToClose gate is a single decision computed over all category results at once — you cannot close on a partial union. The dedup-against-SEEN step is plain set logic, never an agent call.

This skill is not a LOOP itself — but its gaps[] output is the seed that feeds a LOOP: each gap becomes input for another work-sweep, work-refute, or research round (see Integration).

Usage

/work-gaps "context of what was done so far" [focus="optional extra areas"]

The orchestrator calls this at natural checkpoints and at the very end of complex work.

How it works

Step 0 — Scope-coverage diff (mandatory; this is the move that catches the real bug)

Before hunting category gaps, reconstruct the set you actually swept and diff it against the territory that should have been swept (L5). This is the sharp version of completeness — finding what was never on the list, not re-checking what was. Concretely:

• Reconstruct, from todo/journal/handoff, the actually-swept set: files read, angles run, claims verified, consumers traced.
• Run at least one independent global oracle that does NOT depend on the fan-out list — e.g. a repo-wide grep for the core entity/identifier of the task, a list of all callers/consumers, a git diff --name-only against the change set. The point is to surface items outside the fan-out list (this is exactly what caught the 5 orphan files in the grounding case — see anti-error-lessons.md).
• coverageDiff = global_oracle_set − actually_swept_set. Every item in the diff is a candidate gap with category scope-gap and an oracle attached.

Record the oracle command and its result in the output (step0.globalOracle). If you cannot run any independent global oracle, say so explicitly — that is itself a coverage gap, and it blocks clearedToClose (you have no evidence the scope was complete).

Step 1 — Reconstruct the work done

• The critic gets (or reconstructs via tools + todo/journal) what has already been explored, verified, or decided.
• It receives the list of known modalities/angles used, claims that were checked, sources read, etc. This is the actually-swept set Step 0 diffs against.

Step 2 — Systematic gap hunt (by categories)

• The prompt forces the critic to go through a checklist of common omission categories (the category enum below):
  • modality-missing — modalities / angles not used
  • unverified-claim — claims or findings never adversarially verified
  • downstream — downstream effects / consumers not traced
  • assumption — assumptions never turned into direct checks
  • cross-cutting — security, observability, multi-tenancy, error budgets, performance under load
  • time-based — historical / future-state / migration / ordering considerations
  • docs-ops — documentation / operational impact
  • scope-gap — items surfaced by the Step 0 global oracle (outside the fan-out list)
• Plus the hard-to-see cases (e.g., "what happens when this is called from the public-facing portal?").
• In the PARALLEL / BARRIER variant each category is a blind sub-agent; here you walk them sequentially.

Step 3 — File each gap responsibly (L1 + L2 — do NOT just assert it)

Filing a gap is a claim; treat it like one. Before promoting a gap — especially downstream, assumption, or unverified-claim:

• Re-open the cited file:line independently and confirm the gap from the current source (L2). Never file "X is unhandled" off remembered or quoted evidence — re-read it. Record what you re-read in the gap.
• Try the cheapest oracle first (L1): a test, an exact grep + manual confirm, a type-check, running the app. Put the command + result in oracleCheck. If the oracle closes the gap (it turns out handled), set oracleCheck.closedGap: true and drop the gap (do not file a phantom).
• A gap that survives its oracle, or for which no oracle exists, is filed with the oracle attempt recorded (closedGap: false).

Step 4 — Prioritize with the two-sided rubric (L6)

Anchor priority in verified real-world impact, guarding both errors — over-rating a non-gap and under-rating a real one:

priority	meaning
high	concretely reachable + damaging + currently uncovered. Requires a reachabilityArgument (who reaches it, by what path). No high without it.
medium	plausible but unverified, or no oracle available to settle it
low	cosmetic / already-mitigated / dev-time hygiene

Never assign high to a speculative gap; never bury an exploitable one at low.

Step 5 — Compute the close gate (L8 evidence gate)

clearedToClose: true only if every coveredCategory carries ≥1 piece of evidence — an oracle result, a re-read file:line, or a clean Step 0 scope-diff — and no high gap is open. A category you claim to have covered but cannot back with evidence does not count as covered: list it in droppedCoverage (L7) and it blocks the close. See Null-handling & the close gate for how a failed critic forces clearedToClose: false.

Structured Output (STRICT contract)

Emit exactly one fenced ```json block matching this schema. The orchestrator parses and validates it and branches on clearedToClose; it re-requests once on malformed output before treating the run as failed (per workflow-mode.md §3–4).

{
  "step0": {
    "actuallySweptSet": ["src/routes/upload-handler.ts", "auth-middleware.ts", "..."],
    "globalOracle": { "command": "grep -rn 'storageTenant' src/", "result": "7 hits across 5 files; 2 not in swept set" },
    "coverageDiff": ["src/storage/handlers.ts", "migration-runner.ts"]
  },
  "gaps": [
    {
      "category": "downstream",
      "description": "storageTenant override is read in storage/handlers.ts, never traced by the review",
      "suggestedAction": "Run work-refute on the ?hostname storage path with lens 'alternative-code-path'",
      "priority": "high",
      "reachabilityArgument": "the public-facing portal hits the upload handler with attacker-controlled ?hostname, post-auth, no session binding",
      "reReadSources": ["src/storage/handlers.ts:212"],
      "oracleCheck": { "command": "grep -n 'req.query.hostname' src/storage/handlers.ts", "result": "handlers.ts:212 uses it for tenant key", "closedGap": false }
    }
  ],
  "coveredCategories": [
    { "category": "unverified-claim", "evidence": "all 6 findings carry workDone:true from work-refute" },
    { "category": "modality-missing", "evidence": "Step 0 scope-diff clean for code paths; grep returned no un-swept callers" }
  ],
  "droppedCoverage": [
    { "category": "time-based", "reason": "no migration/ordering oracle available this round; not investigated" }
  ],
  "overallAssessment": "Coverage solid on code paths but one high-priority downstream gap is open",
  "recommendation": "Run one more work-refute round on the storage path before closing",
  "clearedToClose": false
}

Field rules:

• gaps[].category ∈ modality-missing | unverified-claim | downstream | assumption | cross-cutting | time-based | docs-ops | scope-gap.
• gaps[].priority ∈ high | medium | low. reachabilityArgument is required when priority is high.
• gaps[].reReadSources (optional) string[]: the file:line sources re-opened and independently confirmed for this gap (Step 3 / L2). Strongly expected for downstream, assumption, and unverified-claim gaps, since those must be confirmed from current source rather than asserted from memory.
• gaps[].oracleCheck records the L1 attempt: { command, result, closedGap: boolean }. A gap with closedGap: true should have been dropped, not filed.
• coveredCategories[] each carry an evidence string (oracle result, re-read file:line, or clean scope-diff). A category with no evidence belongs in droppedCoverage, not here (L8).
• droppedCoverage[] lists every category not investigated and why (L7 — no silent caps).
• clearedToClose is a boolean the orchestrator branches on: true only when every coveredCategory has evidence and no high gap is open and Step 0 ran a real global oracle.

Execution Mechanics

See workflow-mode.md §3 for the full table; the mechanics that bind this skill:

Null-handling & the close gate

A failed/empty/malformed critic does NOT mean "no gaps found". That was a silent false-negative — a dead critic that read as "all clear" would close a task with zero coverage evidence. The correct rule:

• A critic run that errors, times out, or returns un-parseable output (after the one allowed re-request) yields result UNKNOWN, not "clean".
• UNKNOWN forces clearedToClose: false, and the failure is logged as a named line in the decision record + surfaced via your harness's todo/plan tool. The task may not be declared complete on an UNKNOWN critic.
• In the PARALLEL / BARRIER variant, a dead category sub-agent is filtered out (.filter(Boolean)) — it is not counted as "that category is clean". Its category goes into droppedCoverage with reason critic-failed, the surviving categories are re-evaluated, and the drop is logged. One dead refuter must never read as coverage (L7, L8).

Other mechanics

• Completeness is itself a quality pattern: cheap relative to the work it gates. Run the default single-pass critic before declaring done; reserve the fan-out for large/high-stakes scopes.
• Direct oracle ≫ panel (L1): Step 0 and every gap's oracleCheck prefer a deterministic check over a probabilistic opinion. A gap an oracle can settle is settled, not argued.
• Structured output: the strict JSON contract above is enforced; validate + re-request once before treating as failed.
• Concurrency cap (fan-out variant only): launch category sub-agents in controlled waves of ~min(16, cores−2); excess queues.
• Budget awareness: if budget runs low, reduce the number of fan-out categories explicitly and list the un-run ones in droppedCoverage — never silently skip a category (L7). Budget pressure never flips clearedToClose to true.
• No silent caps (L7): every un-run category, un-run angle, or dropped gap is a named line in droppedCoverage + your harness's todo/plan tool. A reader must be able to tell "clean" from "truncated".
• log/phase: announce phases via your harness's todo/plan tool — "Step 0 scope-diff: global grep, 2 items outside swept set", "Completeness critic: raised 4 gaps, 1 high-priority open, clearedToClose=false".
• Feeds the loop: the gaps[] it produces are the seed for the next iteration of work-exhaust, another work-sweep, or targeted work-refute. Dedup that seed against SEEN, not survivors (L4), so rejected items don't re-enter.

Worked example (single-pass, from the grounding case)

A reference-cleanup fan-out edited every file on its task list and reported "done". The work-gaps critic ran as the close gate:

• Step 0 reconstructed the swept set (the task-list files) and ran an independent global oracle: grep -rln '<deleted-entity-id>' .. The grep returned 5 files that were never on the task list — coverageDiff = those 5 orphans.
• Each orphan was filed as a scope-gap gap, priority: high with a reachabilityArgument (they still import the deleted entity → build breaks), oracleCheck = the grep hit, closedGap: false.
• Because a high gap was open, clearedToClose: false — the close was blocked and the 5 orphans fed back into a cleanup round. A naive "categories all look covered" critic with no Step 0 global oracle would have closed the task with the bug shipped.

Integration

• Called by work after major phases and as the final gate before "task complete"; the orchestrator branches on clearedToClose.
• Works extremely well after work-refute or work-judge (the things the adversarial process didn't kill may still have gaps around them).
• Its gaps[] feed work-sweep (new angles), work-refute (newly-surfaced claims), and the round structure of work-exhaust — always deduped against SEEN (L4).
• The shared model lives in workflow-mode.md and anti-error-lessons.md; this skill operationalizes L1, L2, L4, L5, L6, L7, L8.
• Can be run on the output of a human + agent collaboration too.

This is one of the highest-leverage cheap patterns: after you've done the hard work of finding and verifying, spend a little more to ask "what did our process systematically fail to look at?"