Pinarak

Showing 25 of 36 from the live probe.

• Cancel a durable command

  Use this to stop a non-terminal durable job. MCPD validates the persisted boot ID and runner process start identity before signaling, so a stale/reused PID is never treated as the old job. Repeating cancellation of a terminal job is safe.

• Cancel a process batch

  Use this to cancel queued/running jobs in a batch created by start_process_batch. Running MCPD-managed process groups are terminated safely; completed jobs remain completed.

• Checkpoint engineering progress

  Use this after meaningful workflow transitions such as implementation green, CI result, merge, blocker, or release step. expected_revision provides optimistic concurrency: read the latest run before retrying a revision conflict. Keep summaries and next actions compact; full command logs belong in job logs.

• Preview or collect stale workflow state

  Use this to preview retention cleanup or explicitly collect old terminal durable runs. It never deletes active or actively leased runs. The default is preview-only; set execute=true to perform bounded restart-safe cleanup. Normal automatic GC already runs from configured retention policy, so manual execution is mainly for disk-pressure or operator verification.

• Create a directory

  Use this to create a directory and any missing parent directories. It is additive and safe to repeat for the same path. Do not use start_process with mkdir when this dedicated tool is sufficient.

• Create a durable engineering run

  Use this at the beginning of substantial or long-horizon engineering work that may span many jobs, PRs, waits, or client reconnects. Store the stable objective and success criteria once; use checkpoint_run for mutable progress. The returned run_id is the durable resume handle.

• Edit part of a file

  Use this for precise edits to an existing file. For one change, provide old_string and new_string. For a multi-hunk refactor, provide edits; MCPD validates every exact hunk and overlap before one write, so any validation failure leaves the file unchanged. Prefer this over write_file for localized source/config edits. If no exact match exists, inspect the returned closest-match hint before retrying.

• Stop a managed command

  Use this to stop a process session created by start_process. It targets the managed process group, sends SIGINT first, and escalates to SIGKILL if needed. Prefer this over kill_process for MCPD-managed sessions because it handles the process group and cleanup correctly.

• Inspect a durable command

  Use this with a durable job_id to read its authoritative compact state after client reconnect or MCPD daemon restart. The state includes process identity, terminal reason, exit code, and log path metadata but never the raw command.

• Inspect file metadata

  Use this when you need metadata for one known file or directory—such as size, permissions, timestamps, detected type, and text line metadata—without reading its contents. Prefer read_file when content is needed and list_directory when inspecting children of a directory.

• Read more search results

  Use this only with a sessionId returned by start_search. Read retained results from that existing search instead of starting a duplicate search. Non-negative offsets are absolute result positions; negative offsets return results from the tail. Continue while hasMoreResults is true and more evidence is needed.

• Resume a durable engineering run

  Use this with a known run_id to reconstruct current objective, work items, counts, blockers, and next actions without replaying chat history or command logs. A fresh client should prefer this over rediscovering completed work.

• Checkpoint an AI-session handoff

  Use this before a long wait, periodically during multi-hour work, before an anticipated agent/session/turn limit, or during error recovery. It stores a compact durable handoff with active handles, blockers, and ordered next actions. Do not paste full command logs; reference durable handles instead.

• Send input to a command

  Use this only for an interactive PID returned by start_process. By default MCPD appends a newline when absent; set raw_input=true for exact bytes/control-oriented input where no newline should be added. wait_for_prompt controls whether the call waits for another prompt or process completion. For input that must not be submitted twice after a lost response, supply operation_key and reuse the same key on retries.

• Terminate a Linux process

  Use this only to send SIGTERM to an arbitrary OS PID not managed by MCPD. Prefer force_terminate for MCPD sessions. PID reuse makes blind retries unsafe: copy start_ticks from list_processes into expected_start_ticks when retry safety matters; MCPD then rejects a recycled PID instead of signaling the wrong process.

• List directory contents

  Use this to inspect a known directory, optionally recursively. Recursive developer-noise directories are pruned by default and a global max_entries cap prevents response explosions; inspect pruned metadata or set includePruned=true only when dependency/cache internals are actually needed. Prefer start_search when looking for a filename or content across a larger tree.

• List durable commands

  Use this to rediscover recent durable job handles when job_id is unknown. Results are compact and paginated; read_durable_job_log is separate so listing jobs never inlines command output.

• List Linux processes

  Use this to inspect the operating-system process table visible to the MCPD daemon user, including PID, CPU, memory, and command line. Prefer list_sessions when the target was started through MCPD and you need its managed session state or retained output.

• List durable engineering runs

  Use this to rediscover recent durable run handles and compact status when run_id is unknown. The result is a paginated metadata-only summary: it never inlines objectives, work-item bodies, next actions, or job logs.

• List search sessions

  Use this to discover running or recently completed MCPD search sessions, including their session IDs, patterns, backend, runtime, and result counts. Prefer get_more_search_results when you already know the sessionId you want to continue.

• List MCPD command sessions

  Use this to discover PIDs and state for commands previously started through start_process, including retained completed sessions. Prefer list_processes only when you need the full operating-system process table rather than MCPD-managed sessions.

• Move or rename a path

  Use this to rename or move an existing file or directory with the operating-system rename operation. A lost response is ambiguous: verify source/destination state with get_file_info before retrying. Do not blindly repeat a move, and do not use it for copying.

• Read a durable command log tail

  Use this only when durable job state says output evidence is needed. Returns a byte-bounded tail from the disk log; increase max_bytes explicitly for deeper evidence instead of loading the full log into model context.

• Read a file or URL

  Use this to read one text file with line pagination, or to fetch a textual HTTP/HTTPS URL when isUrl=true. Prefer read_multiple_files when several known local files are needed, start_search when the file/location is unknown, and get_file_info when only metadata is needed. Local text returns a single line-oriented `lines` payload with pagination metadata; URL text returns a single full `content` payload. For local files, offset is zero-based; negative offsets read from the tail.

• Read multiple local files

  Use this when the exact paths of multiple local text files are already known and they can be read independently in one call. Prefer read_file for one file or paginated continuation, and start_search when you need to discover which files contain something.