Execute qualquer Skill no Manus
com um clique

Execute qualquer Skill no Manus com um clique

research

Estrelas2

Forks0

Atualizado27 de abril de 2026 às 08:11

Research codebase comprehensively and emit SHA-tracked, line-referenced research docs at docs/research/. Decomposes the user's question, spawns parallel sub-agents to investigate, synthesizes findings with concrete file:line references, computes git hash-object SHA for every referenced source file, and writes a YAML-frontmatter doc. If a doc with the same topic/slug already exists, UPDATES that doc instead of duplicating (refreshes SHAs, appends a Follow-up Research section). Companion to the `docs` plugin (C4 docs). TRIGGER on phrases like "research X", "investigate Y", "deep dive into Z", "how does X work in this codebase", "audit X", "explore the codebase for X", "/research". SKIP if the user only wants a quick answer with no doc artifact, or for non-codebase questions.

Instalação

Instalar com Codex ou Claude Copie este prompt, cole no Codex, Claude ou outro assistente e deixe que ele revise a página da skill e instale para você.

Executar no Manus

Fonte

fkesheh

fkesheh/fmk-plugins

Abrir repositório GitHub Ver repositórios do creator

Download

Executar no Manus

Ocupações relacionadasSOC

Baseado na classificação ocupacional SOC

Desenvolvedores de softwareInformática e Matemática·SOC 15-1252

SKILL.md

readonly

Mais deste repositório

mesmo repositório

fticket

fkesheh/fmk-plugins

Operate fticket (FTS, the Factory Ticket System) through its `fts` CLI — the SQLite-backed transactional control plane that coordinates concurrent agents in an AI software factory. Covers initializing a staged pipeline database, creating tickets, atomic claim/renew/start/complete/advance through stages, dependencies, resource leases with FIFO waitlists, the scheduler sweep, and diagnosing stuck work (board, why, triage, doctor, events, serve). TRIGGER whenever the user or task mentions fticket, FTS, `fts` commands, factory.db, factory tickets, ticket stages/queues, claiming or advancing a ticket, worker claims, resource leases, dead-letter or orphan triage, or wiring agents/workers to a ticket pipeline — even if they just say "check the board" or "why is ticket N stuck" in a factory context. SKIP for GitHub/Jira/Linear issues or generic project-management tickets that are not backed by an fticket database.

2026-07-102

check-readiness

fkesheh/fmk-plugins

A Definition-of-Ready gate run on a task/ticket BEFORE decomposition or implementation begins. Restates the goal as observable behavior, walks a readiness checklist item by item, and classifies every gap as blocking (needs a human decision), resolvable-by-research (answer exists in the codebase/docs — go find it), or deferrable (a conservative default exists — propose it). Prevents the failure mode where a capable model receives a vague ticket, silently fills the gaps with plausible assumptions, and builds something that compiles and is confidently wrong in ways nobody decided. Moves ambiguity discovery to the cheapest point — before any agent is spawned. TRIGGER on phrases like "is this ticket ready", "definition of ready", "readiness check", "can we start on this", "is this ready to implement", "check this ticket before we build", "DoR gate", "is this well-specified enough", "ready for decomposition", "vet this ticket", "any blockers before we start", "/check-readiness". SKIP when the work is already in fli

2026-07-062

decompose-task

fkesheh/fmk-plugins

Executor discipline for splitting a feature or change into parallelizable tasks for multiple agents. Enforces disjoint file ownership (a verified zero-intersection ownership table), explicit dependency edges between every pair of tasks, an independently-completable bar (each task finishable and gate-passable with only its contract, spec, and file list), right-sizing into vertical slices, per-task specs in the standard template, a contract frozen FIRST when tasks share an interface, and an explicit integration task at the end. Prevents the two agents editing the same file / task B silently needing task A's output failure that causes collisions, merge conflicts, and integration failure. TRIGGER on phrases like "decompose this task", "break this into parallel tasks", "split this feature for multiple agents", "how do I fan this out", "plan the parallel work", "divide this across agents", "who owns which files", "make a task breakdown", "parallelize this change", "carve this into tasks", "assign work to subagents"

2026-07-062

design-api

fkesheh/fmk-plugins

Executor discipline for designing a PUBLIC or consumer-facing API surface — HTTP/REST, RPC, GraphQL, webhooks/events, or an SDK surface that callers you do not control will depend on. Forces a surface-wide inventory FIRST so new endpoints match existing naming, envelope, pagination, error, and auth conventions instead of each being locally elegant and collectively incoherent; makes auth (authn + object-level authz), idempotency, pagination, a single error envelope, and a stated compatibility stance mandatory per endpoint; and runs a consumer-journey walkthrough before freezing. Emits the surface as an executable spec (OpenAPI/proto/GraphQL SDL or a typed endpoint table). Exists to prevent the verb-soup RPC surface, the unguarded endpoint, and the accidental forever-contract. TRIGGER on phrases like "design the API", "design the REST endpoints", "design the public API surface", "spec the HTTP API", "design the GraphQL schema", "write the OpenAPI spec", "design the webhook/events contract", "design the SDK surf

2026-07-062

design-contract

fkesheh/fmk-plugins

Executor discipline for authoring the FROZEN contract — the interfaces, types, error taxonomy, config schema, and constants that parallel implementers and test writers build against — with zero logic smuggled in. Enforces strong types (no any/dict/stringly-typed enums), first-class error shapes, boundary completeness between parallel tasks, a decision log, and a freeze protocol that makes the contract immutable once fan-out starts. Exists to prevent leaky, ambiguous contracts that cause parallel agents to diverge and integration to fail. TRIGGER on phrases like "design the contract", "write the contract first", "freeze the interfaces", "define the types and signatures", "contract-first", "set up the API contract before we split the work", "define the error shapes", "write the shared types for the parallel tasks", "spec the config schema", "the interface contract for the fan-out", "author the frozen contract", "define the seams between modules". SKIP when the user wants the implementation itself (this skill pr

2026-07-062

design-tests

fkesheh/fmk-plugins

Executor discipline for designing and writing tests (unit + integration) that catch real defects instead of restating the implementation. Forces every test case to be derived from a behavioral requirement in the contract/spec, a systematic edge-case enumeration before any test code is written, and a "can this test fail?" mutation check on each assertion. Kills the implementation-mirroring failure mode — tests that assert a mock was called or copy the code's own logic into the assertion and therefore pass forever while catching nothing. TRIGGER on phrases like "write tests for X", "add unit tests", "design a test suite", "test this function/module", "improve test coverage", "write integration tests", "what edge cases should I test", "add a regression test", "these tests pass but the bug shipped", "tests for the new endpoint", "cover this with tests", "test plan for this feature". SKIP when the user is running an existing suite with no new tests to author, when they want the red step of a strict TDD red-green-r

2026-07-062

name

research

description

research — incremental, SHA-tracked codebase research

Produces durable research docs at docs/research/ that reference source code by path + git hash-object SHA + line ranges. Designed to grow incrementally and stay coherent with docs/ (C4 hierarchy from the docs plugin) — research is the per-demand, narrower companion.

Output location

docs/research/YYYY-MM-DD-<slug>.md
docs/research/YYYY-MM-DD-<TICKET>-<slug>.md   # if a ticket id is provided

slug: short kebab-case derived from the user's query (≤6 words).
Date: date +%Y-%m-%d in the user's local time.

Frontmatter schema (authoritative)

---
type: research
topic: "Free-form question or research target"
slug: kebab-case-slug
date: 2026-04-27T14:32:00-03:00
researcher: <git config user.name>
git_commit: <HEAD SHA at time of research>
branch: <current branch>
repository: <repo name, e.g. owner/repo>
status: complete | partial
tags: [research, <component-tags>]
last_updated: 2026-04-27
last_updated_by: <name>
last_updated_note: ""        # short why-changed string when updating
references:
  - path: src/auth/handler.py
    sha: <git hash-object SHA-1>
    lines: "12-45"
    note: "Defines the auth middleware entry point"
  - path: src/auth/tokens.ts
    sha: <git hash-object SHA-1>
    lines: "100-130,200"
    note: "Token expiry logic"
related:
  - docs/research/2026-04-20-other-topic.md
---

Required: type, topic, slug, date, git_commit, branch, references. The rest are recommended.

Workflow

Follow these steps in order.

1. Read explicitly mentioned files first

If the user references specific files, tickets, or docs, read them FULLY in the main context (no offset/limit). This grounds the decomposition.

2. Dedupe — look for an existing research doc on the same topic

Before doing any new work:

python3 ${CLAUDE_PLUGIN_ROOT}/scripts/find_existing.py --topic "<user query verbatim or close paraphrase>"
python3 ${CLAUDE_PLUGIN_ROOT}/scripts/find_existing.py --slug "<derived-slug>"

If either prints a path, prefer updating that doc (jump to step 7b — Update existing). Otherwise continue with a new doc.

Also scan docs/ (the C4 docs from the docs plugin) for related context — list candidate paths, but do not block on reading them. Surface relevant ones to sub-agents in step 4.

3. Decompose — ultrathink the question

Break the query into independent investigation lanes. Examples of useful lanes:

Locator: where does feature X live? entry points, modules, files.
Dataflow: how does data move through X? upstream callers, downstream consumers.
Persistence/storage: schemas, migrations, queries.
Tests: what coverage exists? what's missing?
History: relevant commits, CHANGELOG entries.
External boundary: API surface, MCP tools, CLI entrypoints.

Track lanes in TaskCreate so progress is visible.

4. Spawn parallel sub-agents

Use the Agent tool with subagent_type: "Explore" for codebase searches and subagent_type: "general-purpose" for synthesis-heavy lanes. Send all independent lanes in one message with multiple Agent tool calls so they run concurrently.

Each agent prompt should include:

The specific lane question (not the full user query).
Any pre-read files from step 1 quoted as context.
Instruction to return concrete path:line references and short quoted snippets.
Read-only stance (no edits).

5. Synthesize

Wait for all agents to finish. Compile findings:

Live code is the source of truth.
Connect findings across lanes.
Note patterns, conventions, surprising decisions.
Resolve contradictions by re-reading the relevant file (or spawning a focused agent).
Keep the main agent for synthesis, not deep file-reading.

6. Compute SHAs for every referenced file

For every file you cite in the doc:

git hash-object <path>

Batch where convenient. Record each result as a references[] entry with path, sha, lines, note. Use precise line ranges ("12-45") or comma-joined ranges ("12-45,80,120-130").

Also capture:

git rev-parse HEAD              # git_commit
git branch --show-current       # branch
git config user.name            # researcher
basename "$(git rev-parse --show-toplevel)"  # repository (or use owner/repo from `gh repo view`)
date +%Y-%m-%dT%H:%M:%S%z       # ISO date

7a. Write a new research doc

Filename: docs/research/YYYY-MM-DD-<slug>.md. Use the Write tool. Body sections:

# Research: <topic>

## Research Question
<verbatim or close paraphrase of user's query>

## Summary
<2–6 sentences answering the question>

## Detailed Findings

### <Component / lane 1>
- <claim> (`path/to/file.ext:LINE`)
- <claim> (`other/file.ext:RANGE`)

### <Component / lane 2>
...

## Code References
- `path/to/file.py:123` — <what's there>
- `another/file.ts:45-67` — <what's there>

## Architecture Insights
<patterns, conventions, design decisions>

## Related Docs
- `docs/<C4 path>` — <relation>
- `docs/research/<other>.md` — <relation>

## Open Questions
- <unresolved item>

If the current commit is pushed to a remote, you MAY also include GitHub permalinks (https://github.com/<owner>/<repo>/blob/<commit>/<file>#L<line>) alongside local refs. Always keep the local refs too.

7b. Update an existing research doc (same-topic match)

When step 2 found a match, do NOT create a new doc. Instead:

Read the existing doc fully.

Run the staleness check on it:

python3 ${CLAUDE_PLUGIN_ROOT}/scripts/stale.py <path>

Re-investigate the drifted refs (reuse step 4's parallel-agent pattern, scoped to changed files).
Update prior claims that are now wrong.
Recompute git hash-object for every reference (drifted + still-fresh, in case the doc had outdated SHAs from prior work). Update references[].sha.

Append:

## Follow-up Research <ISO timestamp>
<what triggered the update; what changed; new findings>

Update frontmatter: last_updated, last_updated_by, last_updated_note. Bump git_commit to current HEAD. Keep original date and filename.

8. Present a summary

Reply to the user with:

Path of the doc written/updated.
3–5 bullet key findings.
Any open questions or follow-up suggestions.

Companion commands

Command	Purpose
`/research <query>`	Run this skill on a query.
`/research-check [path]`	Staleness report. Empty arg → scans all of `docs/research/`.
`/research-update <path>`	Refresh a stale research doc; appends a Follow-up section.

Staleness algorithm

scripts/stale.py reads each doc's frontmatter and, for every references[] entry, runs git hash-object on the current file. Output JSON list per doc:

{
  "doc": "docs/research/...",
  "topic": "...",
  "slug": "...",
  "stale":   [{"path": "src/foo.py", "stored_sha": "...", "current_sha": "...", "lines": "1-30", "status": "stale"}],
  "missing": [{"path": "src/gone.py", "stored_sha": "...", "current_sha": null, "status": "missing"}],
  "fresh":   [{"path": "src/ok.py", "stored_sha": "...", "current_sha": "...", "status": "fresh"}],
  "is_stale": true
}

A doc is stale if any reference is stale or missing. Use is_stale to triage at a glance.

Same-topic detection rules

scripts/find_existing.py matches in this order:

Exact slug match (case-insensitive).
Exact topic match (case-insensitive).
topic substring match in either direction.

Slug match is preferred — keep slugs stable across updates so dedupe works.

Discipline

Always run fresh codebase research on update; do not trust prior conclusions.
Reference precise line numbers, not whole files.
Keep references[] minimal but exhaustive — every claim that cites code must trace back to a referenced file.
Tag thoroughly (tags) so future research can find this one via related.
Never write placeholder values (e.g. <commit>); compute them before writing.