zy839971925-zyy/Agent--deep-research-Workflow-Skills / skills/reasoning-workflow
reasoning-workflow
VerifiedDefault silent workflow for everyday questions, research, writing, design and engineering. Match effort to the request, resolve costly ambiguity, find specialist resources and verify the requested outcome. Explicitly selectable with “调用我的 workflow skills”.
CHANGELOG.md
View on GitHub# Changelog
## 3.2.0 · 2026-11-19
Adds targeted engineering acceptance, reference discovery, research reporting, reader-aware writing and focused interviews. Fixes standalone validation, unknown routing signals and the DSH index. [Release notes](docs/RELEASE_3.2.0.md).
## 3.1.0 — 2026-10-24 (scenario-led research routing)
- Infer real-world, pre-build engineering/architecture, or linked hybrid research paths from the scenario; no user lane label required.
- Clarify host-selected default model and automatic workflow-depth selection; explicit requested depth wins, while Ultra does not imply a model switch or extra authority.
- Add a controlled protocol case for real-world, engineering, hybrid, explicit Ultra, and Light controls. No model-behavior evaluation is claimed.
## 3.0.1 — 2026-10-24 (evaluation-fixture maintenance)
- Bind the stale-checkpoint fixture to a real archived v1 artifact and computed digest, so version/content mismatch is testable rather than signalled by an obvious placeholder.
- Add mixed qualitative-interview excerpts and a dedicated prevalence-boundary case.
- Add controlled sparse-search recovery and no-web negative-control scenarios.
- Deterministic checks remain protocol/fixture validation only; real-model behavior evaluation is still outstanding.
## 3.0.0 — 2026-10-24 (merged to main; no separate GitHub Release)
- Add triggered research modes for real-world empirical questions, pre-build technical architecture, and linked hybrid investigations; ground architecture comparison in project evidence and quality-attribute scenarios.
- Add a source-aware clarification gate: route uncertainty across conversation state, project materials, live host capability, external facts, derivation, and user-owned judgments; preserve material unresolved state and ask only after relevant checks.
- Strengthen long-horizon resume behavior with scoped artifact relocation/version checks and a focus-to-acceptance check after interruption, handoff, or checkpoint; preserve partial-access uncertainty and unaffected completed work.
- Add a runtime capability contract with explicit fallbacks, selective architecture decision records, and bounded psychology/survey-method guidance; do not treat famous-expert heuristics as guaranteed interventions.
- Add a causal-chain web-search method that turns answer-bearing links into adaptive, role-specific queries, engine-aware operators, source checks, and a proportionate search trace; add a protocol scenario, not a model-performance claim.
- Add seventeen V3 functional evaluation scenarios and concrete synthetic architecture/resume fixtures, including silent automatic entry, explicit invocation, product/runtime mismatch, stale completion-checkpoint detection, causal-chain web search, and safe recovery from an oversized handoff, runtime revision drift, and late worker results after root cancellation. They remain `protocol_only_not_model_evaluated`; no real-model improvement is claimed.
- Define the activation contract separately from task routing: strict every-turn matching requires an always-on host instruction/front router; the direct phrase “调用我的 workflow skills” selects the workflow but does not force Ultra or grant tools/permissions.
- Preserve the 2.2.0 research-frontier, harness, packaging, and test improvements; keep history and previously released artifacts intact.
## 2.2.0 — 2026-10-24 (local candidate; not remotely published)
- Added a decision-bearing research-frontier loop, passage locator guidance, targeted provisional-draft gap search, and gap-driven backward/forward citation tracing.
- Added optional comparison cards for empirical studies, model results, and benchmarks, preserving comparability, uncertainty, denominator, and provenance.
- Reworked active model adaptation for GPT-6+ and current frontier models as capability-first; removed unsupported GPT-6 Astra behavior claims.
- Added host profiles for Codex, Claude Code, ChatGPT app chat, Claude app chat, and Grok chat, including serial Ultra fallback and conditional Intelligent UI guidance.
- Fixed the standalone Modular Deep Research navigation gap.
- Made package tests non-mutating by default; added deterministic archive generation and source/archive parity verification.
- Model compliance remains unevaluated; feature access still depends on host configuration.
# Reasoning Workflow change log
## 2.1.1 — 2026-10-21
- Fix scoped review fingerprints to traverse singular/plural canonical references and transitive dependency cycles safely; missing/duplicate refs fail closed.
- Apply material delivery obligations across all lanes. Add an honest partial-handoff manifest that cannot claim full task completion.
- Enforce forbidden parallel nodes using actual dependency/wait ordering, regardless of optional concurrency groups; preserve independent legal parallel work.
- Make the Ultra method a required routed resource, and reject edited/stale routes or missing method files without silent depth reduction.
- Resolve promotion evidence by role, kind, lesson scope/fingerprint, current/pass status and file hash; require independence evidence for dual review. Tier1-only normal retrieval/advisory maintenance remain unchanged.
- Add optional `prepare_task.py` for profile → route → required/deferred resources, and 30 deterministic boundary tests. Final combined run: 69 passing tests / 38 actual CLI invocations. No model-compliance or external executor claim.
- Retain the 2.1 baseline test report and original regression suite; see the current test report and Chinese release guide.
## Historical 2.1 change log
# 2.1 implementation changes
This is a separately built upgrade, not a change to the audited original. Existing research methods, canonical work-state, reasoning families, verification records and learning promotion gates are retained. No third-party skill code was copied.
## Model-facing behavior
- Replaced the long root entry with a short result-oriented working loop and on-demand task strategies.
- Made explicit depth, scope, format, prohibitions and Stop/Cancel priority visible at the entry point.
- Separated depth from autonomy/persistence; general tasks need no compulsory project record, table, plan or delegation.
- Added practical artifact/code, coordination, continuity and external-action guidance without creating a plugin market, distributed bus, new state engine or mandatory fixed worker tree.
- Preserved research framing, ecology, missingness, source lineage, discriminating observations and affected-branch invalidation.
## Routing and resources
- Explicit Ultra requests apply from every depth and take precedence over automatic downgrade. An optional persistent requested-depth field carries explicit steering. API and CLI share normalization; strict bridge/resource selection reject inconsistent unnormalized profiles rather than accepting weaker review.
- Delegation and orthogonal-verification needs enable the necessary control families in Question and Action paths.
- Required controls remain reachable for Light/high-consequence and high-governance tasks. Typical-depth filtering and small reading budgets cannot silently remove a mandatory family. Ordinary short Light tasks still avoid reference-loading ceremony.
- Root/schema/index paths are derived from script locations rather than caller working directory at corrected entry points.
## Input, lineage and delivery
- Reject malformed raw/learning envelopes; validate explicitly supplied empty runtime objects.
- Bind content hash to supported algorithm and require known derived lineage.
- Check acceptance artifact references and manifest/canonical artifact identity, path, role and basis. Do not force all artifacts to current global state version.
- Reject manifest downgrades of canonical current/required material artifacts; digest checks are selected from canonical obligations, not an untrusted manifest flag. Keep a legal historical optional-artifact control.
- Add an opt-in strict workflow bridge: current profile/route, successful review strength, scoped canonical target fingerprint, canonical verification bindings, current-file manifest and actual SHA-256 bytes. A valid failed review is not a successful review.
- Separate strict structural validity from authenticated evidence and verified orthogonality. Strong orthogonal delivery requires both layer-target bindings and an exact trusted-host receipt; absent integration blocks strong credit. No identity/authentication platform is implemented.
- Preserve `passed` as a legacy record-validity alias while exposing `record_valid` and `verification_passed` separately.
## Runtime and host boundaries
- Proposal schema 1.1 specifies conservative whole-collection replacement arrays. Validate actual keys, declared write scope and allowed scope together; detect conflicts from actual replacements.
- Classification requires a trusted host submission verifier; public static CLI fails closed rather than guessing merge permission.
- Supply a helper for current ledger, executor, attempt, contract, lease and authorization callback checks; document atomic CAS/cancellation fence and durable dedup as host obligations, not implemented guarantees.
- Block terminal recovery, stale/unknown versions and unsafe external replay; idempotency requires host-bound key verification.
- Detect cycles on the joint plan/schedule waiting graph in serial, parallel and hybrid modes.
- Do not allow local isolation labels to bypass external resource lock conflicts. Missing replay safety on consequential/irreversible nodes is an error.
## Learning
- Share schema, semantic and validation-evidence admission between validation and retrieval.
- Prevent intrinsic/empty feedback from becoming validated merely by flipping a boolean.
- Require trusted resolution of validation evidence rather than accepting self-reported labels.
- Restrict ordinary retrieval to Tier1; explicit maintenance reading of Tier2/3 is advisory and grants no canonical-write authority.
## Evidence limits
The shipped regression suite runs real code, schemas and CLIs on synthetic inputs. No real host scheduler/executor, outside account, model adherence evaluation or quality/cost benchmark has been implemented or measured. Authentic authorization, review authorship, context-exposure receipts, semantic support and atomic commit remain trusted-host responsibilities. Strict file identity does not prove file correctness.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| brainstormingobra/superpowersYou MUST use this before any creative work - creating features, building components, adding functionality, modifying behavior, or planning anything new, in software or out of it (a talk, a business, a renovation). | Stars: 323k | Updated | Claude CodeCodex | Verified | Install |
| brand-guidelinesanthropics/skillsApplies Anthropic's official brand colors and typography to any sort of artifact that may benefit from having Anthropic's look-and-feel. Use it when brand colors or style guidelines, visual formatting, or company design standards apply. | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |
| frontend-designanthropics/skillsGuidance for distinctive, intentional visual design when building new UI or reshaping an existing one. Helps with aesthetic direction, typography, and making choices that don't read as templated defaults. | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |
| mcp-builderanthropics/skillsGuide for creating high-quality MCP (Model Context Protocol) servers that enable LLMs to interact with external services through well-designed tools. Use when building MCP servers to integrate external APIs or services, whether in Python (FastMCP) or Node/TypeScript (MCP SDK). | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |