rrrrrredy/research-toolkit / plugins/research-toolkit
research-toolkit
FlaggedUse for source-backed long-form research and writing in English or Chinese across fields, including revisions and continuation. Understand requirements, develop analysis through substantive research collaboration, and deliver a complete report after independent review, audit and acceptance.
references/task-continuity.md
View on GitHub# Continuing an existing research task
[English](task-continuity.md) | [简体中文](task-continuity.zh-CN.md)
Use the existing task when new material, changed instructions, a comparison or a revision arrives. A new report filename is not a new research task. Preserve originals and continue the affected work.
## Project entry
Resume from the original request and currently effective requirements, current outline and report, relevant sources, complete research contributions and unfinished work. Recover this material from existing agent tool conversations and files in Skill-only use, or through the managed task's `status` output. Use the completion summary to locate the current passages, then check the actual content against effective requirements.
New instructions replace conflicting old instructions; uncancelled goals remain effective. A requirement marked satisfied still constrains subsequent analysis and editing. An approved scope change preserves the actual user decision. For accepted corrections, inspect the changed passages and complete the applicable independent recheck.
When a requirement explicitly applies only to named research units, retain that assignment in its optional `unit_ids`; omitted or empty means global. All effective requirements stay readable. A unit's input binding includes global requirements and those assigned to it or its actual upstream units, so a local change does not invalidate unrelated branches. Do not infer applicability from wording, the assigned agent, or which requirements happen to be selected for reading.
## Resume and save
For managed tasks, `research_status.current_work` includes effective requirements, open work, scope decisions, the explicitly selected main artifact, registered contributions and prior delivery references. `status.work` prepares relevant originals and input versions for the assignment; open its `reading_ref` to read the files and records, rather than expanding all text in JSON or guessing from the latest filename.
Use `research_save` to change affected fields and records. Partial `brief` updates retain manually added specification sections and update the saved brief consistently. Stable requirement IDs connect changed instructions to existing work. If direct edits conflict with the saved brief, reconcile that conflict before saving; do not overwrite unrelated content.
The designated main writer registers the report with `draft.main=true`; other workers save separate contributions. `complete=true` marks a complete draft, not whole-task delivery. Keep contribution inputs, dependencies, real agent identity and a concise handoff with the artifact. Preserve previous content when revising; a changed source or requirement calls for inspection of its actual dependent reasoning, not an automatic rewrite of every branch.
Source/claim records and full local originals remain accessible, including counterevidence. Saved record versions support determining what changed. A summary does not replace required original reading or the full analysis that supports a conclusion.
## Handoff and context compression
Handoff applies across agents and across phases of the same agent. Reuse the current context when it contains the necessary work; when handing off or compressing, preserve:
- The actual question, effective instructions and working outline.
- Completed contributions, their conclusions and conditions, and their original-source locations.
- Current report/version, dependencies, unresolved issues and the next concrete action.
- Original reviews, required fixes, independent dispositions and the actual agent tool execution references.
This can be a concise note attached to existing artifacts; it is not a second tracker. Keep full files and raw responses. The receiver reads the relevant full work before relying on it, and opens original sources when needed to understand or revise the judgment.
Send material changes to affected agents explicitly. A subagent's launch snapshot does not include later messages automatically. Continue the same agent task where supported, or supply its full reusable output and new inputs to the next context. The agent tool handles notifications and execution recovery; Toolkit keeps research inputs and artifacts connected.
## Review the necessary scope
For a full comparison, read both current documents across the requested scope; differences in summaries are only leads. For a bounded revision with a reliable baseline, read the complete diff, changed passages with context, affected sources and requirements, and open findings. The reviewer receives the current full artifact and the precise changed scope; a local recheck is not relabelled as a new full review. Missing baseline coverage must actually be read.
Use separate native contexts for independent review. All production identities remain distinct from the grouped content reviews, domain reviewers, auditor and final independent checker. Preserve actual execution identity, original response and exact assignment when importing agent output. Never fabricate historical execution details or copy a prior PASS onto a changed artifact.
The existing independent reviewer checks affected changes; the auditor examines disputes, important no-change decisions and repairs; final acceptance reads the current complete report. The [role methods](subagents-and-review-loop.md) determine their responsibilities. A coordinator or author cannot clear required findings merely by updating a status.
## Editing without losing the analysis
Retain dates, prediction status, experimental conditions, actors and source limitations when removing repetition. Carry meaningful comparisons, calculations, mechanisms and counterevidence into revised text. Changes in a requirement alter the judgments that depend on it; preserve other valid conclusions. Keep facts, inferences and recommendations consistent across paragraphs, tables and conclusions.
Use the user's actual format and style requirements. For decision-oriented work, retain the choice, reasons and conditions that would change it. For explanatory work, preserve the relationships that make the explanation useful. Do not turn every revision into a recommendation or a universal report template.
## Recover a manual or historical project
Read the existing request, effective changes, full artifacts and unresolved feedback. Resolve contradictory file references using their contents and user decisions, not timestamps. Select the main artifact explicitly, repair only records needed for continuation and preserve originals. Do not infer unobserved source reading or reconstruct old execution receipts.
Completed historical tasks retain their original artifacts, labels and completion evidence. Continued research reuses valid materials and fills the current independent review, audit and acceptance responsibilities; old author checklists remain author checklists. See [historical task compatibility](../docs/usage-modes.md#historical-task-compatibility).
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| rea-analysis-optimizationmorluto/reaOptimize REA providers and shared analysis workflows for agent context, latency, CPU, and memory while preserving correctness and Evidence. Use for analysis bottlenecks and cache or query design; CI/test-suite optimization belongs to rea-ci-tests-audit. | Stars: 472k | Updated | Claude CodeCodex | Verified | Install |
| rea-ci-tests-auditmorluto/reaAudit REA CI and test suites for correctness, fixture drift, false passes, and measured performance; use for suite or workflow investigations and efficiency changes, rather than ordinary feature testing or a single test failure. | Stars: 472k | Updated | Claude CodeCodex | Verified | Install |
| rea-tool-designmorluto/reaAudit, investigate, design, or change REA investigation tools, CLI/MCP contracts, provider capabilities, and Evidence semantics. | Stars: 472k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-auditDietrichGebert/ponytailQuality audit of the whole repo: bugs, security, real load, missing tests, speed, and what to delete. Most important first. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |