Termind/Frameworks/TermindRemoteInfrastructure/Agent

Tool Output Policies

AgentKit owns whole-line truncation, optional output accumulation, scratch writes, and model-output markers. Termind owns its tool policies and generated paths. The default model ceiling is 64 KiB (65,536 UTF-8 bytes) and 2,000 logical lines, including JSON escaping and the response envelope in the byte budget. A line that cannot fit is omitted, never split. none explicitly bypasses this ceiling.

Model Strategies and Automatic Capture

Every assistant-visible tool is listed below. Capture is a tool implementation choice, not a model parameter or a Scratch enable/disable switch.

Tools	Strategy	Automatic capture
ssh_exec_run	tail	Cumulative command output
ssh_exec_output	tail	Complete snapshot of the command, not the already-shortened preview
Dynamic outbound MCP tools	middle	Received result; filename uses the called tool name
fetch, web_search	head	Received document/search output
read_pane	none	No
sftp_read, scratch_read	head	No; continue with the returned offset
skill_read_file	head for supporting files; trusted SKILL.md bypasses projection	No; supporting files retain their line pagination
read_chat	head	No; cursor continues at the first unreturned message/line
memory_search, session_search, list_chats	head	No; retain their query, result-count and pagination contracts
scratch_search	Default head	No; narrow the query/path or use scratch_read
get_metrics, list_hosts, list_panes, ssh_exec_list	head	No
sftp_list, sftp_stat, scratch_list, skills_list, list_models	head	No
scratch_diff	head	No; source files remain available
ssh_exec_stdin, ssh_exec_cancel, terminal_suggest	head	No
sftp_write, sftp_mkdir, sftp_move, sftp_copy, sftp_chmod, sftp_delete	head	No
scratch_pull, scratch_push, scratch_fetch	head	No automatic capture; these tools already transfer files to an explicit destination
scratch_write, scratch_replace, scratch_copy, scratch_move, scratch_delete	head	No
request_user_input, present_plan, manage_tasks	head	No
skill_manage, skill_install, memory	head	No
create_new_chat, send_to_chat, wait_chats	head	No
request_user_secret	Secret-result bypass	Never capture secrets

Planning wrappers preserve their wrapped tool's strategy. Provider-native results and already-frozen model projections retain their separate contracts.

External clients calling the Termind MCP server bypass model projection and automatic capture. This also covers the external-only sftp_upload, sftp_upload_status, sftp_download, and sftp_download_status tools. Intrinsic source limits, such as file-read pagination, are not removed.

Paths and Shared Scratch

Scratch is shared by all conversations on this device. Its root is Application Support / Termind / Scratch (inside the app container on sandboxed macOS). A conversation deletion or fork does not delete or copy these files.

The path factory returns a complete scratch-relative path. AgentKit does not prepend output/, outputs/, a conversation ID, or an extension. Termind's tool convention is:

Producer	Path
ssh_exec_run	<conversationID>/outputs/ssh-exec-<UUID>.txt
ssh_exec_output	<conversationID>/outputs/ssh-exec-output-<UUID>.txt
fetch	<conversationID>/outputs/fetch-<UUID>.txt
web_search	<conversationID>/outputs/web-search-<UUID>.txt
Outbound MCP	<conversationID>/outputs/<tool-name>-<UUID>.txt

Tool-name underscores become hyphens. Contexts without a conversation ID use their run ID. This is file organization, not conversation isolation; the agent must still provide explicit relative paths when reading or writing its own files. Parent directories are created automatically. Generated collisions are not overwritten. Existing containment, symlink and quota checks still apply.

A command's collector reuses one cumulative file. Each truncated ssh_exec_output call captures its own snapshot under its own tool name. Files contain redacted output through that snapshot, not future command output. Capture is lazy, limited to 8 MiB per file and the shared 64 MiB/512-entry quota. Quota or I/O failure must not claim that a complete saved copy exists.

Markers and Presentation

AgentKit supplies two unlocalized, bracketed marker forms, following Pi:

[Showing lines 101-2100 of 2100. Full output: <path>]
[Showing lines 1-2000 of 2100]

middle lists both retained ranges. Empty selections report zero returned lines; paginated sources with unknown total length omit the total. The Full output clause appears only after a successful capture. Failure details and continuation offsets/cursors live in separate result fields, not extra markers.

Presentation never generates these markers. It may display the marker frozen by AgentKit for actual tool/model truncation. Display-only folding adds none:

• Compact bodies share five complete logical lines/list items and 64 KiB. SSH keeps the tail; command metadata remains outside the body budget.
• Expansion uses the card's full retained content, without height clipping.
• Existing truncation markers appear above the left-aligned Show more/Show less controls.
• Other text previews use middle, 120 lines and 64 KiB, retaining full content for Copy.
• Display folding never rewrites model content or creates Scratch files.

Tool-specific source limits remain where they have meaning: file pagination, search match counts, metric process counts, skill validation quotas and hunk-aligned diff computation/output limits. They are not additional arbitrary character cuts. read_chat no longer uses its former 8,000/40,000-character cuts.