Tiangong Auto Research

Tiangong Auto Research

Route feedback before research

For requests to report a problem, suggest a capability, or prepare an Issue, read references/issue-reporting.md and follow that reporting workflow. Reporting does not require a valid research workspace, setup, or scientific-question approval. Return the report without entering the research workflow unless the user also requests research execution.

Own the accepted research task

The user supplies goals, material scope preferences and necessary permissions; the native assistant owns operationalization, planning, evidence work, technical choices, checks, integration and an honest completion judgment. Preserve an accepted goal through the existing task intake; do not ask the user to author its JSON or repeat the same contract. No goal yet means no invented task facts.

On “continue”, a new session or context recovery, first follow task recovery before choosing another local experiment. An unchanged valid accepted question does not need repeated approval; apply the question gate to a new substantive request or new biased constraint. Project instructions installed on disk are not proof that an already-running host loaded them; honor reported restart/compatibility requirements and user overrides.

Gate the research question before acting

Before any CLI, browser, search, database, or file operation, inspect the user's research request. Continue when it leaves the result open to evidence. Pause and require a rewrite when it assumes the conclusion, asks to prove a predetermined position, requests only supporting evidence, excludes contrary evidence, or treats an unsupported causal or normative judgment as fact.

When paused:

1. Do not call tools or begin setup.
2. Identify the problematic assumption in one concise sentence.
3. Offer one testable rewrite that preserves the intended topic, scope, population, geography, and time period.
4. Wait for the user's explicit confirmation or edited question.

Do not reject a controversial topic or directional hypothesis merely because it has a position. It may proceed when null results, alternative explanations, and counterevidence remain testable. Refuse requests to fabricate, conceal, or misrepresent evidence.

For an existing managed directory, use the bundled resolver for ordinary CLI operations. An explicitly reviewed CLI upgrade uses the candidate entrypoint in references/setup.md, then returns to the activated lock. Resolve AUTO_RESEARCH_CLI from this loaded Skill's absolute directory; do not guess a global Skill path. The resolver accepts only the CLI package and exact stable version from the regular non-symlink runtime lock, or from the immutable setup plan while installation is still pending or blocked:

AUTO_RESEARCH_CLI=/absolute/path/to/installed/tiangong-auto-research/scripts/research_cli.mjs
node "$AUTO_RESEARCH_CLI" --workspace /absolute/path/to/workspace -- research --help

For a clean directory without a runtime lock or setup plan, read references/setup.md and ask the user to choose one reviewed exact bootstrap CLI version. Never substitute latest, a range, a tag, a path, or a command fragment. After setup creates the runtime lock, use the resolver above.

If that clean directory contains the fixed workspace-local .tiangong-research/setup.yaml, bare research setup must use its declarative path without a TTY. It never scans parent directories and must not fall back to the Wizard after a declaration error. Use research setup init to generate the no-overwrite YAML and env examples. The YAML and env example explicitly expose all current catalog Skills, credentials, and settings, including disabled optional entries. Only declaration schema v2 is accepted; omission is invalid and a disabled credential must remain empty. Use explicit research setup wizard when the user chooses the interactive path. Follow references/setup.md and references/env.md; the generated CLI template is the only authoritative declaration schema.

The CLI is the deterministic control plane: it owns setup, locks, brokered evidence, schemas, coverage, budgets, admission, the independent review process, the journal, and closure. The current interactive Codex, Claude Code, WorkBuddy, or CodeBuddy session owns producer reasoning. The CLI must never launch a nested producer process. Do not reproduce control-plane contracts or edit control files by hand.

During an accepted apply, the CLI may temporarily create the project Skill tiangong-auto-research-recovery. That generated recovery-only entry can inspect context and setup status and execute only the exact pinned retry command. It must never perform research or standalone evidence search. The CLI removes only its own exact plan-bound recovery bytes after this full external orchestrator matches the reviewed tree hash; follow references/setup.md for the detailed stop and recovery rules.

Route to the right reference

• Read references/setup.md for a new or clean directory, the guided Wizard, non-interactive setup, updates, or recovery.
• Read references/external-skills.md before selecting a recommended external Skill, provider, license, or execution role.
• Read references/env.md before configuring credentials, agent authentication, provider checks, wrappers, or the first material-bearing review; revisit its recipient guidance when routing or authorization conflicts.
• Read references/production-research.md before production preflight, execution, recovery, or closure.
• Read references/execution-assurance.md before original-task intake, fresh-session continuation, pre-analysis evidence correction, scope changes, planned-object fulfillment, observed native calculations, packet-only artifact reads, evidence-to-design decisions, pilot interpretation, researcher-facing explanations, reviewer-directed repair, or completion reporting. Prefer supported same-project operations for an unchanged study; preserve original/current task completion separately.
• Read references/bounded-investigation.md for multiple diagnostic trials on one computational requirement, investigation recovery, candidate selection or explicit promotion to a fresh Canary.
• Read references/evidence-pipeline.md before discovery, acquisition, evidence refresh, or an addendum.
• Read references/evidence-exhaustion.md before declaring a material evidence gap exhausted, requesting paid or authorized access, waiting for an external response, or narrowing scope.
• Read references/native-execution.md before preparing or submitting discover, acquire, analyze, or synthesize stages.
• Read references/sandboxed-ide.md when the producer runs inside WorkBuddy, CodeBuddy, or another outer sandbox, or when native reviewer isolation returns a nested-sandbox error. Its host-pause guidance also applies when an operation is rejected or a task is paused.
• Read references/publication-policy.md before a top-journal project, Policy approval, final manuscript freeze, four-role publication review, or readiness closure.
• Read references/scientific-design.md before top-journal admission, any scope change, the three early scientific gates, a target-specific recovery generation, or portable audit export.

Choose the mode before spending budget

• smoke-test is for deterministic mocks, routing/eval checks, workflow demos, and explicitly accepted low-cost canaries.
• production-research is for conclusions a user may rely on. It requires a locked independent public-internet evidence profile. Local evidence and an owner database can supplement that profile but cannot replace it.

Inspect before mutation

The workspace may be any user-selected directory; example paths are placeholders, never defaults. Resolve paths to absolute paths, then inspect the directory:

node "$AUTO_RESEARCH_CLI" --workspace /absolute/path/to/workspace -- \
  research context inspect \
  --path /absolute/path/to/workspace --json

Follow role, allowedOperations, and the structured setup preflight. For setup or a blocked workspace, execute only its exact-version-pinned next.retryCommand; never fall through to a standalone provider wrapper. Stop on invalid; never repair immutable state, locks, journal events, evidence objects, receipts, or outputs manually.

For a clean directory, show the read-only ecosystem catalog and run the guided setup only after the user asks to configure it:

node "$AUTO_RESEARCH_CLI" --workspace /absolute/path/to/workspace -- \
  research setup catalog \
  --workspace /absolute/path/to/workspace --json
node "$AUTO_RESEARCH_CLI" --workspace /absolute/path/to/workspace -- \
  research setup \
  --workspace /absolute/path/to/workspace --json

When the user asks for repeatable non-interactive configuration, first create the safe template with research setup init --workspace /absolute/path/to/workspace --json, then let the user review and complete it. Never pre-accept licenses, downloads, global writes, quota, or agent-smoke cost. Setup is complete only when the resulting overallReadiness is READY; a planned, skipped-check, partially ready, or blocked result is not success.

The user must explicitly confirm the recommended project-local tiangong-auto-research orchestrator and every external source, then accept its pinned license. The Wizard defaults evidence to Brave web/news; context and media remain visibly subscription-dependent choices. For each selected credential it offers hidden input, an owner environment variable, preloaded stdin/password-manager input, or an explicit skip; read references/env.md for the exact safe paths. It may create a plan without applying it. Never silently install a Skill, write globally, substitute a provider, downgrade a profile, or accept a license. Missing required credentials must block before any source download.

Preserve execution boundaries

• Brave and owner-whitelisted SCI/report/patent sources are evidence-capability Skills only after their exact capabilities are locked. Discovery calls them only through the scoped broker and locked manifest method. Never execute their standalone shell examples from a research workflow or expose broker credentials to an agent.

• Built-in structured data capabilities are native CLI capabilities, not external Skills. During discover, use only the packet's dynamic data catalog: choose from its source/capability/operation summaries, inspect the selected input contract through the workspace-locked resolver, and invoke the packet's runDataCapability command. That command promotes the shared TypeScript runtime result into Research evidence. Never use standalone data run for a project, copy provider logic into this Skill, or assume a fixed capability list. Suspended capabilities are not projected into this packet, even though standalone catalog and describe output keep them discoverable for diagnosis.

  runDataCapability.executionKind=workspace-cli-relative-argv means every packet command is relative to the workspace runtime lock; always prefix its resolver-relative argv, readArgv, or describeArgv arguments with node "$AUTO_RESEARCH_CLI" --workspace /absolute/path/to/workspace --; never execute a packet token as a PATH-resolved global tiangong-ai binary.

  Review the returned providerCoverage, limitCoverage, and contextView independently. Provider gaps, intentional operation bounds, and the smaller Agent-facing projection are not interchangeable. When contextView.nextCursor is present and the question requires exhaustive row-level review, invoke the packet's exact runDataCapability.readArgv with the returned receipt and cursor until nextCursor is null. This reads the immutable local evidence and does not consume another provider call. For a summary or adaptive task, stopping earlier is allowed only when the presented fraction is recorded as a limitation; never imply that every returned row was reviewed.

Use the same runtime lock for contract inspection and execution:

node "$AUTO_RESEARCH_CLI" --workspace /absolute/path/to/workspace -- \
  data describe <capability-id> --json

• document-granular-decompose is an input-preprocessor. Run it explicitly through research setup companion run, then admit the exact hash-bound output as a project input. Its output is not evidence merely because parsing worked.
• academic-paper-download is an acquisition-adapter. Its automatic OA order remains Unpaywall, Semantic Scholar OA, then arXiv. If all are exhausted, the CLI reports an explicit browser handoff; it never launches or chooses a browser automatically.
• Document/PDF/spreadsheet/presentation Skills are post-closure-authoring. They may format a closed report but cannot produce or alter admitted evidence, analysis, review, or closure.
• For PPT creation, prefer hugohe3.ppt-master. Keep anthropic.pptx as a compatible situational option; both may be selected explicitly in one plan.
• Selected optional preprocessors, acquisition adapters, and authoring Skills remain visible diagnostics. They become hard gates only when the current project lists them in requiredCompanionIds or explicitly invokes that operation. Their failure never authorizes a standalone evidence downgrade.

Preflight and initialize a project

For --goal top-journal, first use the project-scoped Policy Wizard and stop unless its exact reviewed hash is approved. Generic defaults are bounded feasibility guidance, not journal endorsement. Follow references/publication-policy.md; never invent a target-journal policy or bypass its status gate.

Before top-journal preflight, the current native host must write the closed scientific-design contract. It must distinguish observation, validation, cross-model comparison, scenario, and accounting roles; freeze units, denominators, independent clusters, thresholds, baselines, evidence roles, closest-work requirements, known gaps, and handoff conditions. Pass the same exact file to preflight and init with the native producer's opaque session ID. Before referencing a frozen model implementation or environment lock, use research scientific object register and copy its returned raw-byte hash and safe locator into the design; never hand-copy objects into the control store. The design must map every required evidence role to its lawful acquisition routes. All plan-bound lawful agent routes must have immutable terminal events before an evidence-exhausted handoff is valid. Follow references/scientific-design.md; never let the CLI invent a study design or describe resampling as new independent data.

Prepare evidence requirements with dimensions, sourceTypes, requiredCapabilityIds, requiredCompanionIds, requiredDiscoveryScopes, minSources, minFullTextSources, minDatedSources, and nullable date bounds. Require each owner database that the question must actually exercise; Brave or local files cannot mask an undeclared report, patent, or other whitelisted capability. For large local sources, create an immutable input plan with bounded context files or ranges. Use a projected data operation's exact data:<capability-id>:<operation-id> ID in requiredCapabilityIds only when that source is mandatory for the reviewed question. Do not require all built-in data operations merely because they are available.

First require the current setup generation and its real production checks to pass. Doctor probes required capabilities once, then runs only the independent reviewer CLI smoke after all blocking zero/low-cost checks pass. It reuses an unexpired attestation and never smoke-tests the current native producer as a child process:

node "$AUTO_RESEARCH_CLI" --workspace /absolute/path/to/workspace -- \
  research setup status \
  --workspace /absolute/path/to/workspace --json
node "$AUTO_RESEARCH_CLI" --workspace /absolute/path/to/workspace -- \
  research setup doctor \
  --workspace /absolute/path/to/workspace --live \
  --agent-smoke --confirm-agent-smoke-cost --json

Stop unless researchReadiness is READY and every companion explicitly required by this project or operation is READY. Keep degraded optional domains visible, but do not block unrelated research or weaken evidence coverage.

node "$AUTO_RESEARCH_CLI" --workspace /absolute/path/to/workspace -- \
  research project preflight \
  --workspace /absolute/path/to/workspace \
  --goal top-journal --policy-project PROJECT \
  --question "Research question" \
  --requirements /absolute/path/to/evidence-requirements.json \
  --input-plan /absolute/path/to/input-plan.json \
  --design /absolute/path/to/scientific-design.json --json

node "$AUTO_RESEARCH_CLI" --workspace /absolute/path/to/workspace -- \
  research project init PROJECT \
  --workspace /absolute/path/to/workspace \
  --goal top-journal \
  --question "Research question" \
  --requirements /absolute/path/to/evidence-requirements.json \
  --input-plan /absolute/path/to/input-plan.json \
  --design /absolute/path/to/scientific-design.json \
  --design-producer-agent codex \
  --design-producer-session OPAQUE_NATIVE_SESSION \
  --confirm-budget --json

Stop when capability/input coverage is insufficient or the projected cost has not been accepted. Use the same requirements and input plan for preflight and initialization. For an explicitly selected evidence-report goal, omit the top-journal Policy/design options; do not silently downgrade a requested top-journal study to make admission pass.

For a new project on a compatible locked runtime, register the original-task checklist immediately after init and before the first review or producer stage, following execution-assurance.md. Use a few meaningful user requirements and existing design/coverage IDs, not a duplicate workflow or one requirement per tool call. Never invent historical acceptance for a project that lacks those records.

Validate, run, and recover

Production requires explicit models/prices, different producer and reviewer agent families, a current setup doctor report, and real provider/reviewer attestations. After preflight, use dry-run, then let the control plane identify the next action:

node "$AUTO_RESEARCH_CLI" --workspace /absolute/path/to/workspace -- \
  research run \
  --workspace /absolute/path/to/workspace --project PROJECT --dry-run --json
node "$AUTO_RESEARCH_CLI" --workspace /absolute/path/to/workspace -- \
  research run \
  --workspace /absolute/path/to/workspace --project PROJECT \
  --max-cycles 20 --progress-jsonl --json

For native-stage-required, do not keep calling research run. Prepare the exact next stage, perform the returned prompt in this current app/session, and submit its JSON through the CLI. Discovery evidence must be fetched through the packet's broker command or its native runDataCapability command for a selected structured source. Use the dynamic data catalog and the locked data describe instead of memorizing operations. Follow runDataCapability.readArgv whenever the first data view has a cursor and exhaustive row-level review is required; otherwise disclose the exact presented fraction. Record discovery assessments incrementally instead of returning all source metadata in the final stage output. Record native Web/Browser activity, formalize useful leads through the broker, and bind every network download to its exact download event before artifact registration. After acquire, disposition every acquired content artifact, register exact line-range/JSON-Pointer evidence atoms, and freeze typed content. A stopped acquisition/content gate still preserves these frozen results but prohibits inference. Only then proceed through inference, analyze, the mechanically generated Claim-Evidence Graph, and synthesize. Only after that may research run launch the configured independent reviewer and perform mechanical closure. Follow references/native-execution.md for the exact commands and recovery rules.

At idle stage boundaries, record the checks actually performed against each current requirement before launching the existing independent review. Preserve failed, not-run and inconclusive outcomes; a supported negative finding is not automatically a failure. Bind computational acceptance to its actual native run when the locked CLI supports observation; a self-reported command remains unverified execution. Read large bound artifacts on demand instead of dropping requirements or changing frozen evidence to fit an arbitrary context limit.

For a top-journal project, research status may require research-design, evidence-construct, or pilot-methods review before it exposes the next native stage. Produce the bounded assessment in this native host, then use a fresh configured reviewer session through the CLI. A real-record construct canary occurs only after acquisition and typed-content snapshots are frozen, and an outcome-blind methods pilot occurs after that canary and before analysis. The construct assessment may reference only source IDs, exact atoms, and full-text/date states from the frozen chain. Pass its exact, external JSON canary files through --canary-artifacts; an unbound digest or invented source ID is a mechanical failure. Reviewer prose cannot override these failures.

Proceed only when doctor reports ready. Discovery uses only locked broker or native data evidence capabilities for formal retrieval. Acquisition uses its packet-governed download, parsing, and artifact-registration operations; analysis, synthesis, and independent review do not acquire additional evidence. Doctor, preflight, dependency, provider, and evidence-coverage failures must stop the workflow. Never silently downgrade a systematic task to a standalone SCI, report, patent, web, or paper operation; only the user may explicitly narrow the request to one isolated standalone operation.

Inspect state with research status; the default list contains only authoritative work. A recovery fork supersedes its source. Archive completed or superseded history, and abandon unfinished history, instead of leaving ambiguous project variants. Use the exact native stage abort, research project retry, or research project fork only with explicit user direction; do not reset or delete state.

When the next material step requires user authorization, login/MFA/challenge completion, or an external institution's response, request the packet's durable handoff and stop. Do not spend the remaining search budget on low-yield substitutes. Resume only after the handoff is explicitly resolved. A complete project has a passing independent review, outputs/report.md, and outputs/closure.json. Return the permanent evidence locators, review-packet binding, usage/cost, decision, and material limitations.

For a top-journal goal, that base closure is not publication closure. Author the final paper and role-complete submission files in this current native host, freeze their exact manifest with the assessment, inference chain, Claim-Evidence Graph, and reproducibility record, obtain fresh configured other-family evidence, methods/reproducibility, domain/novelty, and journal-editor reviews, then mechanically close the publication generation. Any Policy, manuscript, or submission-file change invalidates downstream approval/review. Return only the CLI-computed bounded readiness language; never promise acceptance. Report workflow status, publication verdict, task.originalScope and task.currentScope separately. A closed workflow or narrower completed scope does not answer withdrawn or unresolved original requirements. Export and independently verify a portable project audit bundle before external handoff or archival; it must contain the formal evidence bytes and review objects, not merely receipts or local-path references.