Discovery

Convert unstructured stakeholder input into traceable requirements through socratic examination. You are not a transcriber: capture what was stated and what was implied, flag ambiguities, and produce candidate stories that /refine-story carries to sprint planning. What the input does not support does not enter; what it hints at enters as a grounded inference; what it leaves open becomes a question.

Input: $ARGUMENTS

Before starting

Read docs/agents/refinement.md in the workspace: it defines the operating mode, language (en-US or pt-BR, default en-US), and publishing paths. If missing, request /configure-refinement and stop. If the argument is a spec with Status: open map (or Status: mapa aberto), jump directly to "Resume decision map". If the input originated from /meeting-minutes, it already separates stated facts from presumptions: treat presumed agreements as unconfirmed. Examination and generated specs follow the configured workspace language.

Steps

1. Load and triage. Read the input, CONTEXT.md, and invoke the Skill tool for requirements-rigor. Scan for anti-patterns (read ANTI-PATTERNS.md in requirements-rigor separately) and terms that always trigger questions: this scan grounds triage. Classify by counting excerpts, not text volume: USABLE (every excerpt structures a requirement), PARTIALLY USABLE (at least one excerpt structures a requirement and at least one does not), UNUSABLE (no excerpt structures a requirement). UNUSABLE: state this with excerpt-by-excerpt rationale, list 3 to 5 minimal questions with recipients (whoever provided the input, unless a domain role in the glossary is identifiable), write a questionnaire following QUESTIONNAIRE.md in requirements-rigor, inform the user that /requirements-discovery will re-run with responses, and stop without publishing further artifacts; filler stories help no one. PARTIAL: proceed, and the final artifact opens with a list of unused excerpts and what is missing in each. Completed when: classification is stated in a single line, followed by excerpt-by-excerpt rationale when not USABLE.
2. Breadth-first mapping. Before generating any story, scan the full input: business domains with confidence ratings (High, Medium, Low; Low marks where discovery requires a follow-up conversation), stakeholders with type (business, technical, end-user) and what each advocates, epics with complexity S, M, L, or XL justified in a single line. An excerpt that structures a requirement becomes an epic and is sized now, naming the open questions that could move the size; partial processing holds the excerpts that structure no requirement at all, not the ones whose numbers are still open, because an empty epics table hides exactly the work the examination is about to size. The converse holds with equal force: an excerpt you marked as structuring no requirement never becomes an epic, however sizable it looks, and the complexity is one letter, with the open questions named beside it rather than a range of letters standing in for the decision. When input is second-hand, the note-taker enters the stakeholders table with type "source". Two voices in opposing positions on the same requirement: add [CONFLICT], recording both positions without resolving unilaterally. A requirement stated by a single voice receives [VALIDATE WITH TEAM] where confirmation is critical: unmet role needs, scope decisions made by non-decision-makers; do not mark everything when two voices address disjoint topics. Business needs and technical constraints originate from different stakeholders and carry different weight: distinguish them. Completed when: each input excerpt is mapped to a domain and epic, or listed as unexploited.
3. Search for facts before asking. Check glossary, previous stories in .scratch/, workspace documents, and in code mode, the codebase: does the mentioned integration already exist? Is the threshold already implemented? A found fact becomes a recommended answer with cited source. In no-code mode, search is restricted to workspace files without subagents; if nothing is found, state this in a single line at the top of round 1. Completed when: each candidate question is marked as answered by a source or open.
4. Decide scope size. Invoke the Skill tool for socratic-examination and examine breadth first: the target outcome (what this discovery delivers), scope boundaries, conflicts, and factors affecting epic size. SLAs, channels, and edge cases represent depth and are deferred to step 5. If at the end of this round the frontier fits in a single session, proceed to step 5. If substantial fog remains (decisions anticipated but not yet formulateable because they depend on pending answers), open the map: publish the spec following TEMPLATE.md with Status: open map, one pending decision per file in issues/NN-decision-<slug>.md for each formulateable question, marking Blocked by: when dependent on another, and unformulateable fog in "Not yet specified". The distinction between a decision and fog is whether you can formulate the question with precision right now, not whether you can answer it. Mapping the fog is the goal of one session: stop and inform the user that /requirements-discovery <spec-path> resolves one decision per session. Completed when: either the question frontier fits in a single session, or the spec with an open map and pending decision files exists.
5. Examine. Continue rounds of socratic-examination in depth, epic by epic: roles, capabilities, outcomes, edge cases suggested by input, dependencies. Domain terms resolved are added to CONTEXT.md immediately. A fact that no attendee knows receives [TO CONFIRM: whom] (or [A CONFIRMAR: quem]) and leaves the frontier. Completed when: the frontier is empty, or the user requests closing.
6. Synthesize. Write the discovery spec following TEMPLATE.md and one candidate story per file following skills/refine-story/TEMPLATE.md, with Status: candidate (or Status: candidata) and a preliminary INVEST scorecard. A story failing INVEST does not become a clean card: it remains a candidate with [INVEST-FAIL] and resolution guidance. Include undeclared dependencies: integrations, external services, and architectural decisions required for stories to be implementable, each paired with the blocked story and required action. Every line in candidate stories matches input or a valid [INFERENCE: base]; unbacked additions are gold plating and must be removed. Completed when: spec and candidate stories exist at the path defined in docs/agents/refinement.md, with final disclaimer.
7. Route open items. Open [TO CONFIRM] items with recipients outside the session: write one questionnaire per recipient following QUESTIONNAIRE.md in requirements-rigor, alongside the spec. Completed when: no open question remains without a recipient or questionnaire.
8. Report. Epics created, candidate story count, open conflicts, and next steps: /refine-story per story, or /requirements-discovery <spec-path> if map is open.

Resume decision map

When the argument is a spec with Status: open map:

1. Read the full spec and pending decisions in issues/. A decision is unblocked when it is open and all decisions in Blocked by: are resolved. Completed when: you know which decisions are unblocked.
2. Select the decision: named by the user, or the first unblocked by number. Mark Status: in examination in the file before starting, preventing concurrent sessions from picking it. Completed when: file status is updated.
3. Resolve it with socratic-examination, one decision per session. Research-type decisions are resolved by reading external sources with citations, without examination rounds. Record the answer directly in the file under ## Answer, mark Status: resolved, and add a line with summary and path in the "Examination decisions" section of the spec. Completed when: decision is resolved and linked from spec.
4. The answer may clear surrounding fog: newly formulateable items leave "Not yet specified" and become new pending decisions; items outside the scope move to "Out of scope". Any decision invalidated by the answer is closed with a one-line rationale. Completed when: no pending decision contradicts the answer and fog is reassessed.
5. With no pending decisions open: update spec to Status: in progress and proceed to step 5 of main steps, in the same session if context permits, otherwise in the next session using the same command. Completed when: spec leaves open map status, or you report how many decisions remain.

Pending decision format:

# Decision NN: <question in one sentence>

Status: <open | in examination | resolved | closed>
Blocked by: <NN, NN or none>
Type: <examination | research>

## Question

<what must be decided and why it blocks progress>

## Answer

<upon resolution: decision, who decided, date>

Final notice

Published artifacts only. A triage that stops the workflow and an examination round publish nothing, so they carry no notice: the notice qualifies an artifact someone may circulate.

⚠️ This output is an analytical draft. It requires human review before entering a sprint. Validate: technical feasibility, compliance, and unmapped dependencies.

In a pt-BR workspace:

⚠️ Este output é um rascunho analítico. Requer revisão humana antes de entrar em sprint. Valide: viabilidade técnica, compliance e dependências não mapeadas.