feat(researcher): source-grounded research report plugin (debut)

[grixu]

Introduces the researcher plugin to main — an interactive skill that gathers a research brief (depth, recency, source mix, audience) and launches a bundled Dynamic Workflow that fans out firecrawl retrieval into cited findings, gates rounds on coverage + contradictions, synthesizes one cited answer, and renders an evolving HTML report. Follow-ups extend the same report.

Commits

• d7797fd — initial plugin (skill + workflow + assets + ADRs 0001–0009).
• 6429514 — cross-run I/O rework + checkpoint/depth fixes (below).

What changed in 6429514

• Cross-run state I/O (ADR-0010, amended ADR-0005): on extend the Synthesizer reads the prior findings shards directly — no read-back swarm through the orchestrator. Writes are append-only: this run's new findings go to new shards continuing from priorShardCount, prior shards are never rewritten, and the HEAD counts accumulate.
• Follow-up checkpoint → structured AskUserQuestion (multiSelect): the skill now hands the decision back as selectable options (one per proposed follow-up, full text in the description) instead of a prose offer. Fixes the observed flow where the report ended with a free-text question and the checkpoint was effectively skipped.
• Capability-confirming depth: the Plan stage dedicates a primary/official-docs angle per named product/vendor/standard whose capability the goal hinges on; the Assessor treats an unconfirmed capability claim (secondary/community only, or inferred) as a material, resolvable gap rather than green-lighting on inference (high bar on deep); the Verifier flags such claims as needs-evidence. Addresses a deep run that settled "does X support Y" by inference instead of fetching X's own docs.

Requirements

Dynamic Workflows (Claude Code v2.1.154+, paid plan) + firecrawl MCP with firecrawl_search/firecrawl_scrape allow-listed. mmdc optional (diagrams). See the plugin README.

Release (after merge)

Debut as researcher/v0.1.0: reset the version baseline to 0.0.0 (matches every other plugin's pre-debut state), then ./scripts/release.sh researcher minor → tags researcher/v0.1.0.

[pullfrog]

ℹ️ No critical issues — two points worth a look before the post-merge release.

Reviewed changes — debut of the researcher plugin: an interactive research skill that gathers a brief and launches a bundled Dynamic Workflow producing a source-grounded, cited, evolving HTML report.

• Add research orchestrator skill (skills/research/SKILL.md) — infers the brief, resolves the report folder/slug, detects mmdc, launches the Workflow, and runs the follow-up checkpoint as a structured AskUserQuestion.
• Add bundled Dynamic Workflow (workflows/research.js) — Setup → Plan → Assessor-gated round loop (parallel firecrawl retrievers → Conflict-scout → deep-only Verifier → Assessor) → Synthesize → Edit → append-only sharded Persist → HTML Compose.
• Cross-run append-only state I/O — extend reads prior findings/NNN.json shards directly in the Synthesizer; new findings append as new shards from priorShardCount; the state.json HEAD counts accumulate.
• Shipped assets + docs — version-pinned chart.umd.js (Chart.js 4.5.0), report.css (whose class vocabulary matches the Composer's emitted HTML), ten ADRs, README/CONTEXT, and a workflow diagram.
• Marketplace + plugin manifest — registers researcher at 0.1.0 with a productivity category.

The workflow orchestration is well-structured and the append-only shard design is sound. The two findings below are inline: a version-baseline inconsistency that will mis-tag the release, and a low-severity index-coherence note on deep runs.

  ^{｜ Fix all ➔ ｜ Fix 👍s ➔ ｜ View workflow run ｜ Using Claude Opus ｜ 𝕏}

[pullfrog]

✅ No new issues found.

Reviewed changes — re-review of the single new commit since the prior pullfrog review (6429514), which addresses both findings from that review.

• Reset the version baseline to 0.0.0 — both plugin.json and the marketplace.json researcher entry now ship 0.0.0, so the documented post-merge ./scripts/release.sh researcher minor produces 0.1.0 / tag researcher/v0.1.0. Bakes the pre-debut baseline into the commit and removes the manual reset step.
• Reordered the round loop to Verifier → Conflict-scout → Assessor — the Conflict-scout now builds its #index references on the post-splice corpus the Assessor reasons over, keeping them aligned. The Verifier still reads its own pre-splice corpus, so its descending-order splice remains internally consistent; quick/standard runs (no Verifier) are unchanged.

Both prior threads are resolved. No new issues introduced.

  ^{｜ View workflow run ｜ Using Claude Opus ｜ 𝕏}