Refine Work Item

!accelerator config context --skill refine-work-item --fail-safe !accelerator config agents --fail-safe

If no "Agent Names" section appears above, use these defaults: accelerator:reviewer, accelerator:codebase-locator, accelerator:codebase-analyser, accelerator:codebase-pattern-finder, accelerator:documents-locator, accelerator:documents-analyser, accelerator:web-search-researcher.

Work items directory: !accelerator config path work --fail-safe Active integration: !accelerator config work integration --fail-safe ID pattern: !accelerator config work id_pattern --fail-safe

When ID pattern is {tracker}, the tracker owns every item's id, so decompose never allocates numbers: it creates each child as a draft through accelerator work create (see "Under {tracker}" in 4a), and /sync-work-items promotes the drafts later.

Work Item Template

The template below defines the sections and frontmatter fields that every work item must contain. Read it now — use it to know valid kinds, statuses, priorities, and section names without re-reading the file at runtime.

!accelerator config template work-item --fail-safe

You are tasked with refining a work item through one or more of five operations: decompose it into child work items, enrich it with codebase context, sharpen vague acceptance criteria, add a t-shirt size indicator, or populate its dependencies. Every operation is interactive and targeted — you propose, the user approves, and the Edit tool makes the minimum change needed.

Step 0 — Parameter Check

If no work item path or number was provided, respond with:

I'll help you refine a work item. Please provide the path or work item number.

Example: `/refine-work-item {work_dir}/0042-user-auth.md`
Or by number: `/refine-work-item 42`

Run `/list-work-items` to see available work items.

Then wait for the user's input.

If a path or number was provided: proceed to Step 1.

Accepted forms:

• A path (e.g. meta/work/0042-user-auth.md)
• A bare work item number (e.g. 0042 or 42, resolved against {work_dir})

If the resolved path does not exist, report:

No work item file at <path> — run `/list-work-items` to see available work items.

and exit without reading any other file, spawning any agent, or writing anything.

If the YAML frontmatter cannot be parsed (missing closing --- or syntax error), report:

Could not parse frontmatter in <path> — the file may be corrupted. Re-open it
and check that the YAML frontmatter is bracketed by two `---` lines and
contains all nine required fields, or run `/update-work-item <path>` which
surfaces the same diagnostic with field-level detail.

and exit without editing the file or spawning agents.

Step 1 — Read Target and Context

Read the target work item fully. If the parent field is non-empty, also read the parent work item. The work item template (loaded above) tells you the valid kinds, statuses, and priorities — do not re-read it at runtime.

Step 2 — Analyse (mandatory parallel agents)

Spawn BOTH agents in the same tool-use turn (parallel, not sequential):

• {codebase locator agent}: find files relevant to the work item's Requirements and Summary
• {codebase analyser agent}: analyse how the relevant components currently work

Wait for both before presenting the menu. Even if the work item seems straightforward, always spawn these agents — the menu previews depend on their findings.

Step 3 — Present Refinement Menu

Present the five operations with one-line descriptions. Reference agent findings in the previews so the user can pick informed:

1. decompose — split into child work items (epic→stories, story→tasks); e.g. "Requirements suggest 3 child stories (R1, R2, R3)"
2. enrich — add Technical Notes from codebase analysis; e.g. "4 relevant files identified (src/auth/session.ts, …)"
3. sharpen — tighten vague acceptance criteria; e.g. "2 vague criteria detected (AC3, AC5)" or "all criteria already testable"
4. size — append a t-shirt size indicator with rationale; e.g. "estimate M based on files in auth/ and session/"
5. link — populate Dependencies from related work items; e.g. "1 potential blocker found (0031)" or "no related work items found"

Operations that have nothing to do should be marked as such (e.g. "sharpen — all criteria already testable").

User can select one, multiple, or "all relevant". Regardless of selection order, always execute in canonical order: decompose → enrich → sharpen → size → link. This ensures Technical Notes content is in place before size's **Size**: line is prepended, and decompose's children exist before any link operation references them.

Step 4 — Execute Operations

4a. Decompose

Propose 2–5 candidate children (2–4 for story decomposing to tasks) with draft titles and one-line Summaries derived from the Requirements section.

Bug/spike challenge: if the work item kind is bug or spike, first use the AskUserQuestion tool with two options:

1. Yes, proceed anyway — continue with decompose
2. No, cancel — exit decompose and return to the menu

Existing children: if the Requirements section already contains a ### Child work items subsection, offer:

append (add new children to the existing list) / skip (do not decompose further) / cancel

Never replace the existing list silently. For the append path, use the anchor described below under "append to existing ### Child work items".

Approval grammar: each proposal MUST include a one-line grammar legend immediately under the numbered child list:

Commands: approve all | edit N: <title> | drop N | add: <title> | regenerate | cancel

The legend appears on every proposal turn — the first proposal and after every grammar iteration.

Process user input:

• approve all, yes, lgtm → proceed to the pre-write warning
• edit N: <new title> → update child N's title, re-show updated proposal with legend restated. Do NOT write.
• drop N → remove child N, renumber remaining children 1…M, re-show with legend. Do NOT write.
• add: <title> → append a new child with that title, re-show with legend. Do NOT write.
• regenerate → discard current proposal, generate a fresh set of 2–5 candidates from the same Requirements. Do NOT write.
• cancel, abort → print "decompose cancelled — no children written" and return to the menu. No numbers allocated.
• Any other input → print "unrecognised command", restate the legend, and re-show the unchanged proposal unchanged. Do NOT treat as approval.

Pre-write warning: before writing, use the AskUserQuestion tool with two options. State the allocation count and warn about partial state:

1. Yes, proceed — allocate N numbers and write N files (partial state is possible if aborted mid-write; delete the newly-written child file(s) to recover)
2. No, cancel — cancel without writing

A re-key in progress. accelerator work next-number and accelerator work create refuse before reading or writing anything while an accelerator work rekey run is unfinished. Before any other branch on a non-zero exit from one of them, check its stderr for these codes; in every case nothing was written:

• E_REKEY_IN_PROGRESS (exit 4) — tell the user a re-key is unfinished: run /rekey-work-items to finish or undo it. Stop.
• E_REKEY_PLAN_UNREADABLE (exit 4) — relay the printed remedy as is, never suggesting the re-key be finished. Stop.
• E_IDENTITY_RUN_BUSY (exit 1) — another identity command is still running; retry once it ends.

On approval under {tracker}: do not call accelerator work next-number and do not write children with the Write tool. For each child, write its body (H1 through the remaining template sections, with the NNNN placeholder left literally in place) to a scratch file and run:

accelerator work create "<child title>" <kind> <priority> \
  --status draft --author "<resolved author>" --producer refine-work-item \
  --parent "work-item:<parent id>" --body-file <scratch-file> \
  [--tag <value>]...

with no --push, so each child is saved as a draft. Its stdout is the draft's path; the draft's id is the draft- prefix of its filename, and that is the ID the parent's ### Child work items links and the ledger name. If a call exits non-zero, stop, relay its stderr, and list the children already written. Then continue at step 3 below. The pre-write warning says "create N drafts" rather than "allocate N numbers".

On approval under every other pattern:

1. Call accelerator work next-number --count N exactly once to allocate N consecutive numbers.

2. For each child, write NNNN-kebab-slug.md, populating every unified frontmatter field as specified under Populate frontmatter below.

   Immediately before writing each child, verify the computed filename does not already exist. If it does, abort with:

   Collision: <path> already exists (concurrent session?). Aborting.
   Allocated: NNNN, NNNN, …; Wrote: N-1 files (list them).
   Delete the newly-written child file(s).

   Child body includes: Summary (from proposal), Context (linking to parent with "Child of NNNN — <parent title>"), Requirements (minimal but substantive; no [bracketed placeholder] text), Acceptance Criteria (minimal but substantive; sharpen can tighten these later), Dependencies (blank), and remaining template sections.

3. Append ### Child work items to the parent's Requirements section:

   After all children are written successfully:

   • Read the parent file
   • Locate \n## Requirements\n and \n## Acceptance Criteria\n
   • Extract the last non-empty line of Requirements (the line immediately before the blank-line transition to Acceptance Criteria); call this <req_tail>
   • Build: old_string = <req_tail>\n\n## Acceptance Criteria\n
   • Build: new_string = <req_tail>\n\n### Child work items\n\n- NNNN — title\n…\n\n## Acceptance Criteria\n
   • Edge case — empty Requirements: use ## Requirements\n\n## Acceptance Criteria\n as old_string and insert ### Child work items between them.
   • Pre-Edit, count occurrences of old_string in the parent file. If not exactly 1, abort the parent update with:

     Could not locate a unique '## Acceptance Criteria' anchor in <path>
     (matches found: <N>). Parent not updated.
     Children NNNN, NNNN, NNNN remain on disk; add their links manually,
     or delete the newly-written child file(s) and re-run decompose.

     Children already written remain on disk.

   Append to existing ### Child work items (re-decompose path):

   • Locate ### Child work items subsection, find its last - NNNN — title line; call this <last_link>
   • Build: old_string = <last_link>\n\n## Acceptance Criteria\n
   • Build: new_string = <last_link>\n- NNNN — title\n…\n\n## Acceptance Criteria\n
   • Same uniqueness pre-check; same abort diagnostic if not exactly 1.
4. Print a completion ledger, one line per child:

   Wrote NNNN — title
   Wrote NNNN — title
   Allocated 3 numbers, wrote 3 files.

   On aborted or partial write, show allocated vs. written counts and list any written filenames.

After a successful decompose, proceed to Step 5 (hierarchy display) before running any remaining operations.

Populate frontmatter

When writing each child (step 2 above), substitute every field below with the indicated value:

• type: ← work-item
• id: ← from the script, zero-padded four-digit string (own-identity; legacy files use work_item_id)
• title: ← per-child proposal title; body H1 matches exactly
• date: ← current UTC timestamp via date -u +%Y-%m-%dT%H:%M:%S+00:00
• author: ← first match in chain: parent work item's author field → configured author value (from context config) → the session's VCS user identity (the user-identity idiom in the SessionStart VCS Command Reference, which falls back to the git identity when the jj identity is unset, so git-only and colocated repositories both resolve) → ask the user once and apply to all children
• producer: ← refine-work-item
• kind: ← derived: epic → story, story → task, bug/spike → ask user to confirm before proceeding (already done in the challenge step), any other kind → story with a one-line notice
• status: ← literal draft
• priority: ← inherit from parent; if parent has none, ask once and apply to every child written in this session
• tags: ← verbatim copy of the parent's tags array (empty array [] if the parent has none)
• last_updated: ← same UTC timestamp as date
• last_updated_by: ← same author value resolved above
• schema_version: ← 1

Optional linkage/foreign-ref keys are omit-by-default: the template shows each as ""/[], but write a key into the artifact only when it has a value, and omit it entirely otherwise (do not carry the empty placeholder through).

• parent: ← the parent work item's ID as a typed-linkage ref "work-item:NNNN", where NNNN is the parent's id field — from id on unified files, or work_item_id on legacy files — canonicalised to the configured pattern's full-ID shape, e.g. "1" → "work-item:0001". Always fill — every decomposed child has a parent.
• blocks: ← list of typed-linkage refs to work items this child blocks (["work-item:NNNN", ...]). Fill when blocking edges are explicit; otherwise omit the key.
• blocked_by: ← list of typed-linkage refs to work items that block this child. Prefer writing the canonical blocks: on the other side; emit blocked_by: only when the canonical side cannot be written, and omit it otherwise.
• derived_from: ← list of typed-linkage refs to artifacts this child is derived from (["plan:NNNN", ...]). Fill when derivation is explicit; otherwise omit the key.
• relates_to: ← list of typed-linkage refs to related artifacts. Fill when relationships are explicit; otherwise omit the key.
• source: ← typed-linkage ref to the originating source artifact ("issue-research:NNNN"). Fill when the source is explicit; otherwise omit the key.
• external_id: ← cross-system pointer (e.g. a Jira/Linear key). Fill when propagated from the parent or linked to an external tracker; otherwise omit the key.

4b. Enrich

Read the target work item's existing Technical Notes content.

If the codebase agents returned nothing concrete (no specific files or components identified), report:

no enrichment could be grounded in code — skipping enrich

and make no Edit.

Otherwise propose Technical Notes content with specific path:line references drawn from agent results. Do not invent references.

If non-trivial Technical Notes content already exists (anything beyond a **Size**: line), ask:

replace (deletes existing Technical Notes) / append (add after existing content) / skip?

• replace → show a unified diff (old struck, new added) and require an explicit second y/n confirmation before invoking Edit
• append → add new content after existing content. Preserve any leading **Size**: line placed by a prior size operation — never overwrite it
• skip → make no Edit

On approval via Edit: modify only the Technical Notes section. Do not touch Requirements, Acceptance Criteria, Summary, or any frontmatter field.

4c. Sharpen

Read the target work item's Acceptance Criteria. Identify criteria that are vague or untestable (non-measurable phrases like "should be fast", "handles errors gracefully", "works correctly").

If every criterion is already specific and testable, report:

all acceptance criteria already testable — nothing to sharpen

and make no Edit.

Otherwise, for each vague criterion propose a specific, measurable rewrite (e.g. "p95 latency under 200ms under default benchmark dataset"). Skip criteria that are already testable — only propose rewrites for vague ones. Iterate with the user until each proposed rewrite is agreed.

On approval via Edit: modify only the Acceptance Criteria section with the agreed rewrites. Preserve criteria that were not sharpened byte-for-byte.

4d. Size

Read the target work item's Technical Notes. Check whether a **Size**: line already exists as the first line.

Propose a t-shirt size (XS, S, M, L, XL) with a rationale referencing specific files or subsystems from agent results. Iterate with the user.

On approval:

• No existing **Size**: line: insert **Size**: <value> — <rationale> as the FIRST line of Technical Notes, followed by a blank line separating it from any existing content. Do NOT add as a frontmatter key.
• Existing **Size**: line, proposed value AND rationale match byte-for-byte (ignoring leading/trailing whitespace): report "size unchanged — <value>" and make no Edit.
• Existing **Size**: line with a different value or rationale: show a unified diff of the proposed change (existing line struck, new line added) then use the AskUserQuestion tool with two options:
  1. Yes, apply size change — replace the line in place
  2. No, cancel — make no Edit

4e. Link

Count NNNN-*.md files in {work_dir} using a single Glob invocation.

• Count ≤ 30: read them directly via batched Read
• Count > 30: spawn {documents locator agent} scoped to {work_dir}

Propose Blocked by: and/or Blocks: entries in Dependencies, referencing only real work item numbers (verify each proposed number exists before including it). If no related work items are found, print:

no related work items found — link skipped

and make no Edit.

If Dependencies already has non-empty content, ask:

replace (overwrites existing entries) / append (add new entries after existing) / skip?

• replace → show a unified diff and require a second y/n confirmation
• append → add only net-new entries, skipping any that duplicate existing entries
• skip → make no Edit

On approval via Edit: modify only the Dependencies section.

Validate the frontmatter: after editing the work item, run

accelerator corpus frontmatter validate --file <the refined work item path>

If it exits non-zero, the document violates the canonical frontmatter standard; report the emitted violation and fix the frontmatter before completing.

Step 5 — Display Hierarchy

Run immediately after decompose writes at least one child (skip if decompose was cancelled, declined by user, or the parent Edit failed).

Render the parent → children tree using the format pinned in /list-work-items: Unicode box-drawing characters, two-space indent per depth level, ├── for all children except the last, └── for the last child. The canonical fence below MUST appear verbatim in this step's prose so the hierarchy-format guard can verify byte-equality with the matching fence in list-work-items/SKILL.md:

NNNN — parent title (kind: <kind>, status: <status>) ├── NNNN — child 1 title (kind: <kind>, status: <status>) ├── NNNN — child 2 title (kind: <kind>, status: <status>) └── NNNN — last child title (kind: <kind>, status: <status>)

Concrete work item IDs and titles replace the placeholders in the actual output (e.g. 0042 — User Auth Rework (kind: epic)). The status field of the parent is omitted only if blank; all present fields are shown.

Step 5 runs inline after decompose completes, before enrich, sharpen, size, and link operate on the parent. The hierarchy is not re-rendered after subsequent operations.

Step 6 — Offer Review

After the entire selected operation set completes, use the AskUserQuestion tool with two options:

1. Yes, run review — run /review-work-item on this work item now (and each child if decompose was in the selection)
2. No, done — exit without running review

Do NOT invoke /review-work-item automatically — wait for the user's choice.

Important Guidelines

• Mandatory agents at Step 2: always spawn both codebase agents in parallel before presenting the menu, even for simple work items.
• Canonical operation order: decompose → enrich → sharpen → size → link, regardless of the order the user selected them.
• Each operation owns its sections only:
  • decompose: writes new child files and appends ### Child work items to the parent's Requirements (never touching existing Requirements prose)
  • enrich: owns Technical Notes prose content
  • sharpen: owns Acceptance Criteria
  • size: owns the single **Size**: <value> — <rationale> line, always as the FIRST line of Technical Notes; replace in place on re-run
  • link: owns Dependencies
• Never modify any frontmatter field of the target work item (type, id (or work_item_id on legacy files), title, date, author, producer, kind, status, priority, parent, external_id, tags, last_updated, last_updated_by, schema_version) — those transitions are /update-work-item's concern. Children of decompose are new work items getting their initial frontmatter.
• Destructive paths require two-step confirmation: replace mode for enrich, replace mode for link, clobbering an existing **Size**: line — each must show a unified diff and require a second y/n confirmation.
• Edit failure path: if an Edit's target string cannot be matched (file changed between read and edit), abort that specific edit with a clear diagnostic and continue with remaining approved edits. For decompose: write all children first, update the parent last; if the parent Edit fails, children stay on disk and work item numbers are printed for manual linking.
• Numbering is consumed eagerly on decompose with no rollback; the skill assumes a single-session invocation.
• Canonicalise parent field to zero-padded four-digit string (same as /update-work-item), e.g. "1" → "0001".
• Idempotent on re-run: every operation checks existing content before proposing additions.

Relationship to Other Commands

1. /create-work-item or /extract-work-items — create the work item
2. /refine-work-item — decompose and enrich (this command)
3. /review-work-item — automated multi-lens quality review
4. /stress-test-work-item — interactive adversarial examination
5. /update-work-item — status/metadata transitions (not this skill's concern)
6. /create-plan — plan implementation from an approved work item

!accelerator config instructions refine-work-item --fail-safe