SuperAPI AI Directory

specs-kit-task-review

Coding & development

View repository

Updated 10-27 10:54

About this skill

Task Review

Verifies that implemented tasks meet specifications and pass code quality standards. This is the bridge between implementation and verification.

Overview

This command reviews a completed task to ensure:

  1. Task Implementation: The task was implemented according to its specifications
  2. Spec Compliance: The implementation aligns with the functional specification
  3. Code Quality: The code passes code review standards
  4. Acceptance Criteria: All acceptance criteria are met
  5. Definition of Done: The documented completion conditions are fully satisfied

Input: docs/specs/[id]/tasks/TASK-XXX.md (from /skill:specs-kit-spec-to-tasks) Output: Review report with pass/fail status and findings

Workflow Position

Idea → Functional Specification → Tasks → Implementation → Review → Code Cleanup → Done
              (brainstorm)           (spec-to-tasks)       (task-implementation)  (task-review)   (code-cleanup)

Usage

# Review a specific task
/skill:specs-kit-task-review docs/specs/001-user-auth/tasks/TASK-001.md

Arguments

Argument Required Description
--task No Task file path
--spec No Path to spec folder

Examples

Basic Usage

/skill:specs-kit-task-review docs/specs/001-user-auth/tasks/TASK-001.md

Using Spec Detection

/skill:specs-kit-task-review --task=TASK-001

Inputs

Determine the task to review from the invocation context:

  • when driven by the loop, the task is provided in the <task> block of the prompt;
  • when invoked directly, the task path (--task=) and optional spec folder (--spec=) arrive as skill arguments.

Resolve, in order:

  • task → the task file path (required)
  • spec → the spec folder path (optional; derived from the task path when absent)
  • --no-confirm → skip confirmation prompts when set

If any required input is missing or ambiguous, ask the user via ask_user_question.

Retry mode

A prompt carrying a <retry_review> block is reviewing a retry: this task was already implemented, reviewed and rejected, and what you are looking at is the second (or later) attempt. The block carries two things, and it changes how you work:

  • the findings the previous review blocked on. Verify each one and state its outcome — closed, still open, or no longer applicable — in your report. They are what to verify, not a conclusion to adopt: the previous verdict is not in the prompt and you are not being asked to agree with it.
  • a <diff>, when the workspace is under git: the patch from the tree the previous review judged to the tree now. Read it first, and open the workspace only for what it does not answer. When the diff is marked truncated="true", the summary above it still names every file: read the rest from the workspace.

Do not re-run the whole review from scratch: Phase 1 and Phase 2 are a re-read of what a previous attempt already established, and the retry exists because of a bounded list. The verdict, however, stays unbounded: the patch narrows your reading, not your judgement. A change can break code it does not touch, so the acceptance criteria and the DoD are still checked in full (Phase 3), and a regression outside the patch is a finding like any other.

Without a <retry_review> block, this is a first review: run every phase in full.

Core Principles

  • Thorough verification: Check every acceptance criterion and every DoD item
  • Spec alignment: Ensure implementation matches functional requirements
  • Code quality: Verify code passes review standards
  • Evidence-based: Base findings on actual code, not assumptions
  • Use a checklist: Track all progress throughout
  • No time estimates: DO NOT provide or request time estimates

Phase 1: Task Analysis

Goal: Read and understand the task and its specifications

Actions:

  1. (Argument parsing completed in Phase 0)
  2. Read the task file (docs/specs/[id]/tasks/TASK-XXX.md)
  3. Extract:
    • Task ID and title
    • Description
    • Acceptance criteria
    • Definition of Ready (DoR) and Definition of Done (DoD) sections
    • Dependencies
    • Reference to specification file
    • imp-requirements and ac-mapping from frontmatter — which spec ACs this task claims to implement
    • If either section is missing, stop the review and require the task document to be updated before continuing
  4. Read the functional specification file (from task's spec reference)
  5. Verify both files exist and are valid
  6. If files not found, ask user for correct path via ask_user_question

Phase 2: Implementation Verification

Goal: Verify the task was implemented according to specifications

Actions:

  1. Identify what files/components were created for this task:

    • In retry mode the <diff> block already answers this: use it instead of exploring
    • Check git diff to see what changed since task was started
    • Look for new files matching the task scope
    • Review implementation details
  2. Verify implementation matches task description:

    • Compare implemented functionality with task description
    • Check if all described features are present
    • Identify any deviations or missing parts
  3. Document findings:

    • What was implemented vs. what was specified
    • Any deviations from the original plan
    • Additional changes that were made
  4. Read decision-log.md if exists:

    • Check for decision-log.md in the spec folder
    • If file exists, read any DEC entries related to this task (TASK-XXX)
    • Use decision context to understand WHY deviations were made
    • Reference specific decision IDs when explaining deviations in findings
  5. Verify test count (do not estimate):

    • Count the actual @Test (or language-equivalent) annotations in the task's test files: run grep -r "@Test" src/test/ --include="*.java" | wc -l or the equivalent for the project language
    • Report the exact count in the review; never guess or round
    • If the count cannot be determined programmatically (non-standard test framework), state only "tests pass" without a number, or use the count reported by the build output (e.g. Tests run: N)

Phase 3: Acceptance Criteria and DoD Validation

Goal: Verify all acceptance criteria and DoD items are met

Actions:

  1. List all acceptance criteria from the task file

  2. List all DoD items from the task file

  3. For each acceptance criterion and DoD item:

    • Identify code/tests/review evidence that validate it
    • Check if tests exist and pass when relevant
    • Verify the requirement is actually met
  4. Mark each item as:

    • ✅ Met (with evidence)
    • ❌ Not met (with explanation)
    • ⚠️ Partially met (with details) — allowed only with a blocking issue that names exactly what is missing. Without that issue the criterion is met: "partially" is not a verdict, it is a finding waiting to be written. What is left is a warning or a suggestion, the criterion counts as met, and the remainder is routed to the task that owns it (Phase 7).
  5. Update traceability-matrix.md:

    • Read docs/specs/[id]/traceability-matrix.md
    • For this task (TASK-XXX), update the matrix:
      • Fill in "Test Files" column with test file names created for this task, written as path/to/file::name of the test when a specific case proves the row. The loop checks these citations when the range closes: a file that is not there, or a name that is not in it, is reported as a claim the matrix cannot back
      • Fill in "Code Files" column with source files created for this task
      • Update "Status" to "Implemented" only for rows whose citation you have actually read in the cited file; leave the rest "Pending"
    • Save updated matrix back to docs/specs/[id]/traceability-matrix.md
  6. **BOUNDED CONTEXT ADHERENCE CHECK **:

    • Read docs/specs/ontology.md for bounded context definitions
    • Determine the primary bounded context of the feature
    • For each file modified/created in the implementation:
      • Determine its bounded context from path conventions or ontology
      • If DIFFERENT from the feature's primary context:
        • Check if the task file has a "Cross-Boundary Warning" section
        • If YES and justification is valid: note in review as "acknowledged cross-boundary"
        • If YES but justification is weak: add warning issue
        • If NO warning section: add blocking issue
    • Why this matters: Tasks that silently cross bounded context boundaries are the #1 cause of architectural drift.

Phase 4: Specification Compliance Check

Goal: Ensure implementation aligns with functional specification AND verify task necessity

Actions:

  1. Review the functional specification to verify compliance

  2. Compare implementation against:

    • User stories and use cases
    • Business rules
    • Integration requirements
    • Data requirements
  3. Identify any gaps or misalignments

  4. Check if implementation introduces any out-of-scope changes

  5. **SPEC FIDELITY CHECK **:

    • Read the task's imp-requirements and ac-mapping from frontmatter
    • For each AC-ID in ac-mapping:
      • Verify the implementation actually satisfies the acceptance criterion
      • Check the criterion's taxonomy in the spec: [IMP], [SEF], or [EXT]
      • If the task claims to implement [SEF] or [EXT] criteria:
        • Flag as "Task Over-Specification"
    • If the task has NO ac-mapping or imp-requirements:
      • Flag as "Legacy Task — no traceability metadata"
  6. Verify task necessity:

    • Ask: "Is this task implementing a criterion that requires new code?"
    • If ALL the task's ACs are [SEF] or [EXT]:
      • Flag as "Unnecessary Task — no implementation needed"
    • If the task creates entities/structs NOT mentioned in the functional spec:
      • Check data-model.md for (derived) marking
      • If NOT marked (derived): Flag as "Invented Entity — not in spec"
  7. Check for spec contradictions:

    • If the implementation does something DIFFERENT from the spec (a requirement, an acceptance criterion or an interface contract):
      • Record it in the spec_conflicts frontmatter list, one line per conflict, naming the requirement and what the code does instead
      • Check decision-log.md for a DEC entry justifying the deviation. A DEC whose Decided By is the implementation agent of this same task does not justify anything: the session that deviated wrote it. Only a DEC decided outside the task (operator, earlier plan, change-spec) closes the conflict
      • Never resolve a contradiction by rewording the requirement, the contract or the criterion: those documents are what the work is measured against, and the loop refuses attempts that rewrite them
    • A conflict is never a "clarification item", a "wording tension" or a note in the body. Either the code changes, or the requirement changes through a decision taken outside this session — and until then the conflict stays in spec_conflicts, where the loop reads it

Phase 5: Convention Provenance Check

Goal: Verify that every convention the implementation relied on was already documented in the project before this task started, and that the agent did not silently invent new patterns that will be picked up as "pre-existing" by downstream tasks.

The motivation is that decision-log.md entries whose Decided By field is the implementation agent itself are not actually authoritative — the reviewer cannot use them as evidence that a convention was followed if the convention was created in the same pass. This phase rebuilds the authority check from the source documents (architecture.md, ontology.md, data-model.md, prior task files, and DEC entries with a pre-existing Decided By).

Runs: Always, on every review (including --no-confirm reviews and reviews driven by the loop). It is the cheapest check in the pipeline and the one most often skipped by mistake.

Actions:

  1. Build the set of pre-existing conventions:

    • Read architecture.md (the project architecture document, not the spec's technical plan) and collect every architectural rule that appears as a bullet, list item, or numbered item. Record the file's Last Updated header — anything added later is not pre-existing.
    • Read ontology.md and collect every bounded-context definition, glossary entry, and naming convention.
    • Read data-model.md and collect every invariant listed under "Business invariants" / "Regole di business" / equivalent headings.
    • Walk every DEC entry in decision-log.md whose Decided By is not the implementation agent for the current task. Mark them pre-existing.
    • Walk every DEC entry whose Decided By is the implementation agent. Mark them as "agent-claimed" — they are candidates to be flagged unless the corresponding convention can be sourced from the architecture / ontology / data-model docs above.
  2. Enumerate the conventions the implementation introduced or relied on:

    • For each file the implementation created or modified, read it and list every non-trivial pattern: factory methods, builder APIs, annotations used as markers, package layouts, naming conventions, visibility choices, constructor signatures, JSON property names, log message shapes, etc.
    • For each pattern, note the file and line where it appears.
  3. Cross-check each implementation pattern against the pre-existing set:

    For every pattern from step 2, run these three questions in order:

    a. Is the pattern already in architecture.md / ontology.md / data-model.md as written before this task started? If yes: ✅ pre-existing, no issue.

    b. Is the pattern backed by a DEC entry with Decided By ≠ this implementation? If yes: ✅ pre-existing (the DEC predates the task even though the implementation honours it), no issue.

    c. Is the pattern only documented by a DEC entry whose Decided By is the implementation itself (or by nothing at all)? If yes: ❌ invented-convention issue. Severity depends on impact:

    • blocking if the convention is observable from outside the task's own files (package name, public API shape, JSON property, schema annotation that the OpenAPI contract picks up, marker annotation other modules will import). Downstream tasks will read this convention as "the way things are done" without knowing it was invented this morning.
    • warning if the convention is purely internal to the task's own classes (private helper, package-private overload, internal validation order). Document it but do not block.
    • suggestion if the convention is cosmetic (comment style, blank line placement, import ordering).
  4. For every invented-convention finding, recommend a remediation:

    • Promote to architecture: add a bullet to architecture.md §3.5 (or the equivalent Architectural Rules section) describing the convention in the same style as the surrounding bullets. Cite the DEC entry that introduced it. Until the bullet exists, downstream tasks cannot rely on the convention.
    • Or ratify via DEC: if the convention is too narrow to belong in architecture.md, add a new DEC entry with Decided By: User or Decided By: Architecture review rather than leaving the agent-claimed DEC in place. The reviewer should not sign off on patterns that are only justified by an agent-claimed DEC.
    • Or revert: if the convention is not actually necessary, change the implementation to match an existing documented pattern and drop the agent-claimed DEC.
  5. Add the findings to issues with the same severity vocabulary already used in this skill (blocking / warning / suggestion). A blocking invented-convention is enough to flip review_status to FAILED — the implementation passes its ACs but leaves the codebase without an authority for a convention downstream tasks will assume is canonical.

  6. Cross-reference with Phase 4 (Spec Compliance): if the implementation already added a DEC entry in decision-log.md to justify the deviation, do NOT treat that DEC as automatic authority. Re-apply the same three-question check to the DEC itself. A DEC written by the implementation to justify its own deviation is exactly the case this phase exists to catch.

Why this phase is mandatory: every task downstream of TASK-N depends on TASK-N's output being consistent with project conventions. If TASK-N invented a convention without documenting it in architecture.md, TASK-(N+1) will adopt it without knowing it is one task old. Three tasks downstream the convention looks "established", and removing it requires touching every consumer. Catching it at the review of the task that introduced it is the cheapest place to intervene.


Phase 6: Code Review

Goal: Verify code passes quality standards

Actions:

  1. Perform the code review focusing on:
    • Architectural alignment
    • Coding standards and patterns
    • Security and performance
    • Error handling and edge cases
    • Maintainability and readability
  2. Document specific code findings (file, line, issue, recommendation)

Phase 7: Review Report Generation

Goal: Generate a summary of review findings and set status

Actions:

  1. Calculate overall status — review_status takes one of two values and no others, and one rule decides between them. The verdict is FAILED when, and only when, at least one of these is true:

    • an issues entry is blocking
    • an acceptance criterion or a DoD item is not met
    • spec_conflicts is not empty
    • escalation is not empty

    Everything else is PASSED. There is no other reason to fail a task, and no count of non-blocking findings adds up to one: warning and suggestion never flip the verdict, however many there are. A report that carries no blocking issue, no unmet criterion, no conflict and no escalation is PASSED even when its findings tables are long.

    A criterion you were tempted to call "partially met" is the case this rule exists for. Either what is missing is worth a blocking issue — then write that issue, mark the criterion not met and fail — or it is not, and the criterion is met with a warning or a suggestion recorded beside it. The report cannot fail a task on "partially" alone.

    The non-blocking findings are not lost, they are routed: send each one to the final cleanup task or to the next task that owns the area, as a routed entry (below). That is the channel for everything a passing review still wants fixed.

    A finding too large for the next implementation pass is still FAILED: put the escalation in the first issues entry rather than inventing a status for it. Automation reads this field and understands nothing else.

    One kind of finding has its own list. When what blocks the task is an action only a person can perform — an account, a credential written into a vault, a signature, a purchase, physical or console access — the verdict stays FAILED and the finding goes into escalation, one line each. No implementation pass can close it, so the loop ends that task on the first entry and continues with the next one instead of spending its remaining attempts reaching the same verdict. Use it only for that: a fix the next implementation could make is an issues entry, not an escalation.

    A non-blocking suggestion you route to a later task does not by itself flip the verdict, but a suggestion a prior review routed to this task that the implementation left unactioned is a blocking issue (FAILED) — unless a task still in the range will cover it. A deferred fix with no later owner is exactly the gap the review exists to catch.

    Two things the loop decides, not you:

    • a report with a non-empty spec_conflicts is treated as FAILED whatever review_status says, so raising a contradiction and passing it in the same breath is not available
    • a routed entry aimed at a task that is not still pending inside the range is refused and comes back as a rejection, so route only to a task that will actually run
  2. Generate review report (docs/specs/[id]/tasks/TASK-XXX--review.md): Read the review template using this lookup order:

    1. templates/task-review.md
    2. templates/task-review.md inside the installed skill folder for non-Claude agents. Fill in the gathered findings and save to the tasks directory.

    The file must open with the YAML frontmatter block of the template, before the first heading, carrying review_status, a one-line summary, an issues list (issues: [] when the review passes), a spec_conflicts list ([] when nothing contradicts the spec), an escalation list ([] unless the task needs an operator action) and an optional routed list. Every required fix listed in the body has a matching issues entry: that list is what the next implementation pass is given to work from.

    Quote every value with double quotes. The values here are prose and prose contains colons; an unquoted summary, issues entry or routed text carrying a colon followed by a space is not valid YAML, and the loop then has to salvage the verdict line by line.

    routed carries fixes you defer to a later task, not this one: each entry is { to: "<task-id>", text: "<one-line fix>" }, and [] when there is nothing to route. The loop feeds routed entries to the target task's prompt automatically, so they survive between reviews instead of living only as prose a later task would have to grep for.

    The review template defines these sections:

    Section Purpose
    Review Summary High-level status table (AC, DoD, Code Quality, Spec Compliance, Architecture)
    Acceptance Criteria & DoD Results Per-criterion and per-item status with evidence
    Code Review Findings Table of issues with severity, file, category, recommendation
    Spec Compliance & Architectural Alignment Fidelity check, cross-boundary adherence, decision log, traceability update
    Required Fixes Critical / Warnings / Suggestions tables
    Next Steps Action by review status
  3. Update task status:

    • Set status: reviewed and reviewed_date: YYYY-MM-DD in the task frontmatter if PASSED
    • Leave the task status untouched if FAILED: the report is the record of what has to change

    Skip this step entirely when an automated loop drives the review — it owns the task frontmatter and writes it once the whole task is closed.

  4. Synchronization:

    • Run /skill:specs-kit-sync [spec-folder] to synchronize all components
    • Skip this step under an automated loop: synchronization is a phase of its own there, and running it from inside the review doubles the work
  5. Inform user:

    • Display review summary and status
    • Provide link to full review report
    • If PASSED, suggest running Phase T-7 cleanup in task-implementation

Phase 8: Report the convention provenance verdict explicitly

Goal: Make the Convention Provenance Check from Phase 5 visible in the review summary, not buried in the issues table.

Runs: Always.

Actions:

  1. In the review report's summary section, add a row to the standard summary table:

    | Convention provenance | ✅ All conventions pre-existing / ⚠️ N invented-convention findings / ❌ N blocking |

  2. List every invented-convention finding by DEC ID and file in a dedicated section after "Code Review Findings", titled "Convention Provenance Findings". Each entry must include:

    • The pattern (one sentence)
    • Where it appears (file + line)
    • Whether it is blocking / warning / suggestion
    • The proposed remediation (promote to architecture, ratify via DEC, or revert)
  3. The verdict participates in review_status:

    • any blocking invented-convention → FAILED
    • only warning / suggestion → still PASSED, but the report must enumerate them and route them

    This is the single verdict rule of Phase 7 applied to this phase's findings, not a second rule: the motivation is the same, the implementation may have shipped working code, but the loop cannot move on to the next task while the codebase carries undocumented conventions that future tasks will treat as canonical.