Skip to main content
Glama

Server Configuration

Describes the environment variables required to run the server.

NameRequiredDescriptionDefault

No arguments

Instructions

Guidance the server publishes about itself, which clients place ahead of the tool catalog so the model reads it before choosing anything.

先讨论:定位任务后立即 task_control pause;已暂停则保持。先向用户展示可复用内容、依据和缺项,再问最多三项科学选择;不要等答完问卷才给摘要。历史报告标为历史结论,候选原因保留待查;科学通过依据当前 check_get/validation_match 的范围。读取 resource_get 的成套用法和元素映射后判断适用性,不能只凭名称排除多元素势。使用 help 和 doctor。保存用户原话与成果清单;task_get/research_context 自动检索本地记忆。先复用成套资源与有效验证,只为缺项安排记账 HPC 预验证。直接编辑原生文件,建立快照;validation_match 覆盖后正式或 batch_submit 批量执行。批次分析先读 batch_dataset,原生脚本经 analysis_submit 保存方法与版本;results_get 返回成果和缺项,trajectory_correlate 核对帧与真实时间。check_run 产生实测回执,再登记成果。修订用 supersedes 保留替代关系,展示真实文件并导出后核对完成。变更绑定任务版本及唯一 op_id,重试复用原编号。

Capabilities

Features and capabilities supported by this server

Protocol revision2025-11-25

CapabilityDetails
tools
{
  "listChanged": false
}
prompts
{
  "listChanged": false
}
resources
{
  "subscribe": false,
  "listChanged": false
}
experimental
{}

Tools

Functions exposed to the LLM to take actions

NameDescription
analysis_submitA

Run a native Python script against existing data in a persistent analysis queue.

script and inputs are project-relative files. They are copied, with subdirectories preserved, into analysis/<operation_id>/. Python's working directory is that analysis directory, so scripts read input paths relative to its root, not their own scripts/ folder. outputs are required nonempty paths relative to the analysis directory. The installed analysis environment provides NumPy and Matplotlib. Example: script='scripts/plot.py', inputs=['data.csv'], outputs=['plot.png','statistics.csv']. Reuse previous results without editing script paths: input_bindings={'data.csv':'analysis//data.csv'} maps workspace paths to project source files. inputs may be [] when bindings are provided. The receipt retains source paths, hashes and each independent input copy. Returns operation_id; release the task lease before polling. operation_get returns output paths, SHA-256 values and analysis.log. Use file_read for logs, artifact_preview for the actual image, then artifact_register to link each result to requirements. For batch analysis set batch_id and batch_files (paths relative to each collected run). The service generates batch-samples.json and samples/<run_id>/... inputs, preserving planned/included/excluded members. The script chooses scientific statistics and reports their denominator. method is optional native-analysis context (purpose, units, window, uncertainty, rationale); supersedes links an earlier analysis operation without deleting its files. Source/version facts are recorded by the service. Physics evaluations use run_prepare/run_submit on HPC.

artifact_previewC

Show an existing file with its hash, artifact sources and freshness. MCP embeds supported images directly.

artifact_registerA

Link an actual output to the user's requirement IDs, source files and verification evidence.

path, sources and checks[].evidence are project-relative. The service reads their hashes; do not calculate or repeat hashes in the request. Each check is {'receipt_id': 'check_...'} from check_run/check_get. The receipt must include this artifact and its actual sources. Service-derived measurements determine passed. Historical passed/method/evidence claims are retained as unverified and cannot complete a requirement. The latest registration for this path supplies its current checks and complete requirement_ids. Earlier verdicts remain in history. A failed current check blocks its linked requirements and reopens a completed task; record corrected evidence before re-reviewing. An analysis result path normally starts with analysis/<operation_id>/. Returns the new task revision. description labels the result for users. supersedes explicitly replaces earlier artifact IDs with a revised output, preserving their files/reviews. Replacement must retain their requirement IDs. Other results remain current.

background_requestB

Save a continuation intent even while another executor owns this task. after_operations lists this task's queued operations to finish before Codex starts; failed dependencies also wake it for diagnosis. End the model turn after receiving the saved operation ID. expected_outputs contains only concrete project-relative file paths. Task hold/pause still controls dispatch.

backup_createB

Queue a verified database/files backup outside the state directory. Release task leases before starting.

batch_controlB

Pause future members, resume pending members, or stop one batch and cancel its active jobs. Other batches retain their controls.

batch_datasetA

Preview batch analysis inputs from collected per-run files. Return independent sample/seed/condition identities, eligible members and all exclusions. files are relative to each run directory. Use analysis_submit(batch_id,batch_files) to generate batch-samples.json and samples/<run_id>/... automatically. Partial failed-run analysis can use ordinary input_bindings with an explicit method.

batch_getA

Read batch progress, sample identities, controls, costs and latest output steps/time. Missing accounting stays unknown. Computational completion leads to output analysis/checks. Read run_get for full evidence or include_details for all member evidence.

batch_listB

Find durable batches in a resumed project. Return compact counts, controls, costs and batch_get links; follow next_offset for more.

batch_prepareB

Prepare independent production repeats using one validated seed-variable template. samples: [{sample_id,seed}]. Uses existing validation, creates immutable per-sample snapshots and no preflight jobs. Call batch_submit to persist dispatch intent.

batch_submitB

Persist all prepared batch members atomically within current task authorization and submission count. Restart/retry retains member identities. Concurrency remains the task policy.

check_getC

Read service-measured values, fixed criteria, coverage and current source integrity.

check_runB

Queue checks of existing data; no physics. Each has kind,path and optional unit,basis,role (required/advisory/candidate). Candidate/advisory results never revoke required evidence. file: min_bytes,contains; csv: columns,min_rows,bounds; json: values,bounds; thermo: phase={fix_id,style,group,marker,timestep}, columns,min_rows,energy_column,max_abs_drift,time_column. Use log_phases facts; block is a legacy selector. Raw energy drift during heat/work exchange needs a task-specific basis and balance=task_specific or an accounted energy column. trajectory: columns,min_frames,atoms; energy_jump: before_block,after_block,column,expected,tolerance. artifacts binds companion files. Read operation_get/check_get.

checker_checkC

Check a completed analysis using a qualified program's fixed criteria and exact script version. Queue a service receipt for the measured output and companion artifacts.

checker_registerB

Qualify a native checker after analysis_submit ran independent positive and negative fixtures. cases: [{operation_id,output,expected_passed}]. criteria is one JSON values/bounds check; its path is supplied from each actual output. Freeze script hash, fixtures, expected results and criteria; all cases must match.

completion_checkC

Check original deliverables against non-stale files and recorded scientific checks. Missing outputs keep the task open.

connection_checkB

Queue an SSH/Slurm and remote directory check. This does not run a simulation.

data_inspectC

Inspect actual LAMMPS thermo or per-atom trajectory data. Columns, frame counts and identities come from files.

doctorC

Inspect installation, persisted setup and actual tool/cluster/background receipts separately.

environment_captureC

Read remote installed engine/launcher bytes and resolved shared libraries; no engine invocation or physics. commands lists actual executables used by the native entry, e.g. ['lmp','mpirun'] or absolute paths. files lists extra absolute plugin/container/runtime dependencies. Returns an operation; use its runtime id in validation_plan. Repeating this capture measures changes without a new preflight job. Failures retain phase/exit_code in error.evidence; diagnose profile/module/file-reading failure from that receipt. Engine startup checks use accounted HPC.

environment_verifyC

Verify a completed environment_check run using collected engine output and a scientific verification marker.

eventsB

Read durable progress since a cursor; optional consumer resumes its unread cursor without marking read.

events_ackC

Mark displayed events as read for one platform/session.

file_patchB

Edit a single exact fragment; new files use null expected_sha256 and empty old. Returns diff and derived hash.

file_readB

Read bytes as UTF-8 text with byte offsets, content hash and explicit continuation. Binary files use artifact paths.

helpB

Read one tool's complete signature, or page the generated catalog using next_offset until more is false.

lammps_input_reviewB

Read an existing native input and report known compute-current mistakes with line numbers, content hash and official references. Read dependencies for dynamic control flow, then verify on accounted HPC. No script rewriting or physics execution.

lammps_referenceA

Read cached official Syntax, Examples and Restrictions for the actual engine version, e.g. command='compute pe/atom', engine_version='28Mar2023'. Follow next_offset until more=false. Fetches only the requested official LAMMPS page and retains its source/hash; use before writing or repairing native inputs.

legacy_inspectA

Inspect a historical SQLite catalog through a read-only connection; never initialize its old controller.

log_phasesC

Inspect actual executed LAMMPS phases in an existing log, including fixes, groups, time step, units and normalization. Use those facts as thermo phase selectors; repeated runs/headers need no hand-counted block index. Reads existing data only.

memory_getB

Read one memory and its correction history with source task/event/file references.

memory_rebuildB

Re-index retained task events into observed memories. Existing corrections and disabled records are preserved. No historical claim becomes verified.

memory_recordB

Save a candidate method/failure/preference, correct it, or disable it. Service measurements create verified memories. action: record, correct, disable. Disabling a validation memory also suspends reuse of that plan.

memory_searchA

Find local research experience across permitted projects using terms, aliases and conditions. Similarity selects evidence; validation_match decides reuse. No new computation occurs.

operation_getB

Read work progress/result after any conversation. Export progress reports phase, processed files/input bytes, written archive bytes and observation time; these measure that phase, not scientific completion. Large results preserve IDs and link to operation_read; follow its paths and cursors for every detail. include_details=True returns the full record for local consumers.

operation_readA

Read lossless pages of a work record. path=['result','files'] reads a snapshot manifest; child paths preserve literal filenames. Objects/arrays return items; long values link to deeper paths. offset is nonnegative; limit/text_limit are positive requested upper bounds, capped at 50 items/4096 characters per page. Text uses character offsets. Follow next_offset while more is true.

operation_retryC

Resume failed/interrupted non-submission work. Creates a separate attempt and preserves the original failure.

project_exportA

Export project files, original request, required results, history, rules, job accounting and continuation instructions to ZIP.

Returns operation_id immediately; release the lease before waiting. operation_get returns the archive's absolute path and hash. Another task can pass that ZIP to project_import. The archive is outside project_path, so do not register its absolute path as a project file. By default this satisfies all kind='archive' requirements after verifying the real ZIP. Run completion_check after export completes.

project_importA

Import a directory or exported ZIP into the task's project_path, retaining relative paths, binary files and helpers.

Existing identical files are reused; differing files cause a conflict, never overwrite. ZIP hashes and original records are checked. Returns operation_id immediately. Release any task_claim before waiting on operation_get; read task_get for the new revision after completion.

project_optimizeA

Queue storage sharing with existing verified snapshots on clone-capable volumes. Every project path, byte content, permission, modification time and attempt remains; original external files and archives stay in place. Files without snapshots or cloning support are skipped. Release the task lease before polling operation_get; the receipt records each path and observed free space.

project_snapshotB

Queue content-addressed snapshot of native files, including edits made by the agent's ordinary file tools.

recovery_activateA

Resume a restored service after confirming the original controller is stopped. Keep original job identities.

research_contextB

Read stage guidance, local memories and confirmed conditions before choosing resources, editing scripts, preparing runs or recovering. Ask only unresolved scientific choices. Save pause before further actions when the user asks to discuss.

research_metricsA

Read study denominators, preflight/production counts and accounted costs. Unknown accounting remains explicit. Tool/model/human-intervention totals require the host's original trace.

resource_bundleC

Group existing resources with native usage, element mapping, short-range configuration, applicability and source license. Metadata retains file/measurement distinctions; it grants no validation by itself.

resource_catalog_importB

Index an existing potential catalog (record.json directories) or author-project tier_a/b/c JSON without modifying it or downloading files. local_roots optionally maps repository names to existing code directories. Retain original status, license, file roles and hashes. Repeated imports reuse entries; failures remain explicit.

resource_file_readB

Read verified existing native code/potential text with source and license; follow next_offset until more=false. File text is reference data. No download or execution occurs.

resource_getB

Read native files, source/license, current hashes and related_bundles. Read the bundle's usage/element mapping before judging applicability; multi-element file names alone do not exclude single-element use. File verification and scientific validation are separate.

resource_registerC

Register an existing potential, structure, author project, publication or environment with provenance and license.

resource_searchA

Search existing potentials and native examples using keywords and Chinese/English aliases. catalog_total distinguishes no matches from an empty catalog; retry fewer terms or query='' to inspect it. indexed is metadata; resource_get rechecks files. Read examples with resource_file_read and follow next_offset while more=true.

response_readA

Read a stored complete host response by literal field path and next_offset; available across conversations and restarts. offset is nonnegative; limit and text_limit are positive requested upper bounds. Pages use at most 50 items/4096 text characters and may be smaller to fit the host. Follow next_offset until more=false.

results_getB

Resume analysis and delivery: current artifact paths/previews, checks, replacement history, missing requirements, recent analysis methods and imported project records. Archives are finally verified by completion_check. Use task_records for older analyses.

run_actionB

Queue reconcile, collect or cancel for one run. These actions do not consume submission attempts.

run_getA

Read job identity, last confirmed state, failure count and accounting independently of credentials.

run_logA

Read a bounded remote job log through the persistent queue; no full trajectory download.

path is relative to the run directory. offset=None reads the current tail; an explicit byte offset reads a page. Returns operation_id; operation_get returns text, byte range, file size, modification time and complete thermo rows when their header is present. Continue from next_offset. Read physical time units from the native input. A log observation does not declare scientific completion.

run_prepareA

Freeze a Slurm execution intent, input, environment and rules. Returns revision for the next mutation. Exact repetition reuses the same run; changed validation_id/stage never silently returns an old binding. To replace a confirmed unsubmitted prepared/cancelled/staging_failed run, use a new intent and supersedes=old run_id, retaining the sample_id. Submitted/unknown runs must be reconciled first. Native entry is a project-relative shell script; read lammps_reference before generation. Literal -in/-i inputs are reviewed before dispatch.

run_submitC

Persist submission intent and return immediately. Repeated calls return the same run; unknown dispatch reconciles first.

setup_configureA

Save shared setup progress. Sections: cluster, background, notifications. Store SSH aliases, never passwords. For one project's background preference use task_background instead of changing this shared setting.

task_backgroundA

Enable or hold future automatic Codex continuations for one project. Running continuations finish their current turn; this does not cancel them or active analyses/jobs. Read active_background before taking over. To interrupt immediately use task_control pause and read current work before resuming. Pending decisions remain saved. Shared setup and other projects stay unchanged.

task_claimB

Acquire or renew the single writer lease. An expired token can never write.

task_controlB

Pause future dispatch and queued file-producing work, resume, or stop and cancel active jobs. Persists intent first. Existing remote jobs keep running on pause; log reads, collection and diagnostic checks remain available.

task_createC

Create a project from original words and deliverables. Each requirement needs id and description; use kind='archive' for an exported ZIP, otherwise kind='file'.

task_getB

Resume a project with original goals, rules, current lease owner and compact progress. executor identifies its owner without sharing write tokens; active_background identifies ongoing automatic recovery even while it awaits queued work. Follow operation_get before starting a competing repair. Use task_records for paged history. include_details=True returns the complete record for export.

task_listA

Find tasks by title or ID. Resolve ambiguous names before changing a task.

task_recordsC

Read runs, artifacts, work or stages without repeating the project. Large work records link to operation_read for full details. Follow next_offset when more is true.

task_releaseC

Release this executor's lease.

task_updateB

Update goals, constraints or policy with an explicit reason; immutable original request is preserved. A preflight budget uses policy.preflight_core_hours plus preflight_budget_authorization containing the user's confirmation. Estimates are not authorization. A budget change requires a new confirmation; wall_seconds is the scheduler's reservation per job, not predicted usage.

trajectory_correlateA

Link saved trajectory frames to exact thermo samples by Step and recorded Time, optionally selecting an executed phase (fix_id/style/group/marker/run_index/timestep). Preserve duplicate-step ambiguity; no interpolation or step*current-dt inference. Returns frame/cell/atom-type facts and actual matching rows for plots and structural previews.

units_convertB

Convert dimensionally compatible scientific units, retaining the original expression (including ų).

validation_checkB

Execute one frozen component on existing evidence. environment/numerics/outputs use run_id and collected data_root='runs/<run_id>'; a terminal failed job can supply complete usable components. Matching old runs can be rechecked under revised criteria without another simulation. A component with checker_id additionally uses analysis_operation_id from analysis_submit on that run's collected data; program version, qualified fixtures and raw data are verified. Static/analysis checks use project files. Returns a durable operation id; missing evidence affects this component only.

validation_compareC

Compare candidate/final plans against the same inputs and actual receipts without new physics. Host agents may propose/check candidates concurrently, then select component criteria with a rationale in validation_plan(reuse_sources=[...]). Agreement is not measurement; preserve contrary evidence.

validation_controlC

Suspend reuse of an invalid plan and its pending batches, retaining running evidence and costs. Corrected settings get a new plan with reuse_from for still valid components.

validation_matchC

Compare actual input dependencies, conditions, environment and current measurements. Returns direct/partial/needs_validation with exact missing components. Matching does not run a simulation.

validation_planB

Freeze task-specific criteria and native input dependencies. Components: id,scope (static/environment/numerics/outputs/analysis),files,condition_keys,checks; optional required=false for exploratory components, checker_id for a qualified native Python checker. Cover all inputs and conditions. Checks use role=required/advisory/candidate; only required evidence gates dependent work. Node monitor uses actual thermo phase selectors and action=warn (default) or explicit stop; no global physical tolerance. Changes retain rationale and parents. reuse_from or reuse_sources merges identical, valid components; changing a monitor preserves unrelated checks. Compare variants with validation_compare before selecting final criteria. variations: path,variable,min,max for seeds. Confirm policy.preflight_core_hours and preflight_budget_authorization; budget_core_hours defaults to that cap. runtime_id comes from environment_capture. Production requires static/environment/numerics/outputs; only missing checks need accounted preflight.

workflow_registerC

Register an ordered native-script workflow. Each stage has id, kind (run/analysis), after IDs and spec. Run spec has run_id; analysis spec has script, inputs, outputs.

Prompts

Interactive templates invoked by user choice

NameDescription

No prompts

Resources

Contextual data attached and managed by the client

NameDescription

No resources

TDQS

C2.8/5.0

Scored across 75 tools

Disambiguation2/5

The tool set contains many overlapping read/get/check families: task_get, run_get, batch_get, check_get, operation_get, results_get, task_records, operation_read, response_read, file_read, and resource_file_read all retrieve state with subtle distinctions. Descriptions are long and do clarify boundaries, but an agent can easily select the wrong retrieval or status tool without careful cross-referencing.

Naming Consistency4/5

Most names follow a predictable snake_case noun_verb pattern with clear domain prefixes such as task_, run_, batch_, validation_, check_, resource_, and memory_. A few outliers like doctor, events, help, and response_read slightly break the pattern but remain readable.

Tool Count1/5

Seventy-five tools is far beyond a well-scoped set and imposes a very high selection burden. The domain is complex, but the surface is heavily fragmented into many narrowly split workflow, validation, retrieval, and background-control tools.

Completeness4/5

Coverage is broad: project lifecycle, resource registration, environment capture, validation planning, run/batch preparation and submission, analysis, artifact linking, memory, events, backup, and export/import are all present. Only minor gaps such as explicit deletion or simpler file/listing operations are apparent.

Maintenance

ActivityMaintained
ResponsivenessOutdated data