Skip to content

commands.md

Latest commit

 

History

History
423 lines (276 loc) · 38.4 KB

commands.md

File metadata and controls

423 lines (276 loc) · 38.4 KB
Raw

TLH slash commands

This document lists slash commands available in an interactive TLH session, grouped by origin.

Type any command name with a leading / in the TLH TUI to trigger it. Autocomplete surfaces most commands as you type; a small set of upstream and bundled commands is hidden from autocomplete but remains triggerable by typing — see Hidden autocomplete commands below.

On interactive startup, TLH may also show one quiet hand-curated random tip line for that process launch. Those tips are just lightweight discovery hints for real TLH commands and workflow affordances; they are launch-scoped and not LLM-generated.

Note: autocomplete hiding is not an execution block. Any command listed under Hidden autocomplete commands can still be run by typing it in full.

Upstream Pi built-ins

These commands are provided by the upstream Pi runtime. They are available in every TLH session unless noted otherwise.

Command Description
/changelog Show upstream Pi changelog entries — hidden from TLH autocomplete; use /tlh-changelog for TLH release notes
/clone Duplicate the current session at the current position
/compact Manually compact the session context
/copy Copy the last agent message to the clipboard
/export Export the session (HTML by default; pass a .html or .jsonl path to specify format)
/fork Create a new fork from a previous user message
/hotkeys Show all keyboard shortcuts
/import Import and resume a session from a JSONL file — hidden from TLH autocomplete
/login Configure provider authentication
/logout Remove provider authentication
/model Select the active model (opens the native selector; Enter is session-only and Ctrl+S saves the default; see model selection)
/name Set the session display name
/new Start a new session
/quit Quit the TLH TUI
/reload Reload keybindings, extensions, skills, prompts, and themes; also recapture TLH experimental-flag state for the active session
/resume Resume a different session
/scoped-models Enable or disable models for Ctrl+P cycling — hidden from TLH autocomplete
/session Show session info and stats
/settings Open the settings menu
/share Share the session as a secret GitHub gist
/thinking Select the model thinking level with Pi's native picker and controls
/tree Navigate the session tree and switch branches
/trust Save the current project trust decision for future sessions

TLH commands

These commands are registered by the TLH extension bundled with this profile.

Command Description
/effort TLH behavioral alias for Pi's native thinking-level picker
/experimental Open the TLH experimental-feature picker in TUI, or list/change TLH experimental features via typed subcommands (delta-follow-up-reviews and ci-failure-investigation are currently registered)
/tickets Show the read-only tk-backed TLH ticket workflow details for the current repo/worktree
/review Open an interactive code-review mode picker (available with the architect or disabled primary agent)
/switch-primary-agent Show or switch the active TLH primary agent (architect, rush, product, bug-hunter, disabled)
/reconcile Review and resolve model/effort override drift from TLH packaged defaults
/subagent-settings Show or edit persisted TLH bundled minor-agent model and effort overrides
/tlh-changelog Show TLH release notes from the packaged CHANGELOG.md
/tokens Generate and open a single no-flags local HTML token-spend report for the current session
/what-consumed-my-session-limit-and-tokens Generate and open a local HTML session-limit usage report across all in-window TLH sessions
/toggle-context-cap Toggle the 200k effective context-window cap for auto-compaction
/toggle-tlh-git-attribution Toggle the TLH commit attribution footer for agent-created git commits
/usage Show or change TLH subscription usage-limit footer preferences
/version Show the installed TLH version and the upstream Pi runtime version

/thinking (native) and /effort (TLH alias)

/thinking is Pi's built-in command. TLH does not register, route, or intercept it. /effort is the TLH behavioral alias: it uses Pi 0.84.4's exported ThinkingSelectorComponent and the current model's native supported thinking levels.

With no level argument in the interactive TUI, both commands use the native picker and visibly show the same controls:

  • Enter applies the selected level to this session only.
  • Ctrl+S applies the selected level and saves the future-session default. Pi owns this write for /thinking; /effort uses TLH's guarded isolated-profile settings path, which backs up changed settings, preserves unknown fields, and reports a session-only fallback if saving fails.
  • Esc closes the picker without changing the active level or persistent default.

Typed /thinking <level> and /effort <level> values, plus native thinking cycles/shortcuts, are session-only. The picker and typed commands use levels supported by the active model. For an enabled primary, TLH retains explicit choices through later turns and session-tree reapplication while that primary remains selected; a model switch clamps retained intent to a supported level. A new session or explicit primary-agent mode change clears session-only state, then uses the persisted default when present or the provider-aware packaged default otherwise.

Disabled primary-agent mode

Use /switch-primary-agent disabled to disable the primary persona for the current session; use /switch-primary-agent default disabled to make it the persistent default. Shift+Tab also cycles into disabled mode.

Disabled mode retains TLH's base defaults/infrastructure and the architect-equivalent configured tool surface, including subagent safety/authorization checks and provider auth-health preflight. Canonical bundled minor agents and persisted-trust-authorized project custom embedded.<slug> agents remain available; new embedded runs are forced to the validated Git-root project scope and a fresh context. It does not inject the architect persona, architect-only project append or experimental guidance, automatic primary model/thinking defaults, any minimum thinking floor, or per-primary model override. The current session's model and thinking level stay unchanged unless you explicitly use /model, /thinking, or /effort.

/review is available while architect or disabled is active (and requires the interactive TUI); rush, product, and bug-hunter remain blocked. It gathers the selected review target, sends a [/review] handoff, and the active primary delegates it to code-reviewer in a fresh isolated context, digests findings, and keeps the request review-only. Disabled mode does not regain architect planning, approval, ticket, or implementation orchestration.

/experimental

/experimental currently registers delta-follow-up-reviews and ci-failure-investigation. In the interactive TLH TUI, running /experimental with no arguments opens a picker that shows current feature state and lets you toggle flags; outside the TUI it falls back to the status list. Typed subcommands remain available: /experimental list, /experimental status [feature], /experimental enable <feature>, /experimental disable <feature>, and /experimental toggle <feature>. contrarian is a bundled default minor subagent for sparing pre-ticket planning stress-tests when a proposed change genuinely warrants an adversarial brief; it is not part of the /experimental toggle surface, not the routine code-reviewer diff pass, and not the broader oracle second-opinion path. delta-follow-up-reviews is an opt-in flag that adds architect and code-reviewer guidance for delta-scoped follow-up reviews after fixes. ci-failure-investigation is an opt-in flag that lets the architect primary agent do read-only failed CI/status-check investigation after TLH opens a PR, then summarize and ask whether to proceed before any edits, commits, pushes, reruns, PR changes, or other follow-up changes. Both flags are disabled by default. These prompt-only flags are re-read from settings on each agent turn, so enabling or disabling one applies on the next agent turn; no new session or /reload is required. Project custom subagents use the exact direct Git-root .tlh/agents/custom/<UPPERCASE-SLUG>.md path and persisted project trust rather than an /experimental flag; see custom-subagents.md. Stale run-tests-last and embedded-subagents values in tlh.experimental.enabledFeatures are inert and do not re-enable retired behavior. Users upgrading to stable custom subagents do not need to edit the stale embedded-subagents value; if you choose to clean it up manually, remove only that value and preserve every other enabledFeatures entry.

/subagent-settings

/subagent-settings configures persistent model and effort overrides for the nine bundled TLH minor-agent roles: code-reviewer, contrarian, developer, test-runner, diff-summarizer, librarian, oracle, repo-scout, and web-scout. These roles are dispatched through TLH's first-party extensions/subagents runtime; they are not a separately installed extension. The command-only test-runner is intended for exact commands from final-validation tickets.

Grammar

/subagent-settings
/subagent-settings status [role]
/subagent-settings set <role> [model <provider/id>] [effort <off|minimal|low|medium|high|xhigh|max>]
/subagent-settings reset <role> [model|effort]
/subagent-settings reset-all

With no arguments, the command opens the picker in the interactive TUI. In a non-TUI or headless context, it reports status instead. set requires at least one model or effort pair; the two pairs can be supplied in either order. A model must be currently available and should be written as provider/id; do not put an effort suffix on the model override—use the separate effort field. status shows every bundled role, or one role when supplied. reset <role> clears both saved fields for that role, while the field-specific forms clear only model or effort.

Storage and backups

Overrides are global settings stored under subagents.agentOverrides in the active isolated TLH profile's settings.json. The default profile is ~/.the-last-harness/agent/settings.json; when PI_CODING_AGENT_DIR is set, the profile it names is used instead. The TLH wrapper sets this variable for the isolated profile. Writes are refused when the selected directory is the normal Pi profile or otherwise outside the isolated TLH profile, so the command will not modify normal user configuration. The user-facing effort field is stored as the runtime's thinking field.

TLH preserves unrelated settings and per-role keys. When a write replaces existing settings content, it first creates a collision-safe settings.json.bak-* backup (the notification shows the exact backup path); a new settings file with no previous content has no backup to report. To roll back a whole settings change, restore the desired backup over the active profile's settings.json, for example:

cp /path/to/settings.json.bak-<timestamp> /path/to/settings.json

Use the path shown by TLH, or substitute the directory selected by PI_CODING_AGENT_DIR for the default path above.

Dispatch precedence and warnings

For each subagent dispatch, the effective choices are resolved from highest to lowest precedence:

  1. A direct dispatch with an explicit model that already has a recognized :effort suffix is left unchanged; the suffix wins over any saved effort.
  2. A direct dispatch with an explicit model but no recognized suffix keeps that model. A saved role effort may be appended when the model supports it; a saved role model pin does not replace the caller's explicit model.
  3. When the dispatch has no explicit model, stored /subagent-settings values are resolved before bundled defaults. An available stored model pin is used, and a stored effort is applied when supported. A stored model: false means inherit the current session model and then apply the stored effort. A stored effort without a model pin applies to the resolved role model.
  4. With no applicable stored override, TLH uses the bundled provider-aware role defaults. code-reviewer, oracle, and contrarian prefer an available opposite provider for independence; the other bundled roles generally follow the current session provider when TLH injects defaults.

A saved model pin that is no longer available is not silently replaced: TLH warns that TLH saved minor-agent model override "<model>" for <role> is not currently available; forwarding the saved pin unchanged instead of swapping in bundled defaults. Update it with /subagent-settings set <role> model <provider/id> or clear it with /subagent-settings reset <role> model. During dispatch, the subagents runtime uses fallback models from the role's saved subagents.agentOverrides.<role>.fallbackModels field or bundled agent frontmatter, together with TLH-generated provider-aware candidates. Caller-supplied fallbackModels are no longer accepted. Because TLH cannot observe every runtime fallback source, this warning does not claim that a dispatch will fail closed.

Stored effort overrides are capability-checked on each dispatch. When a recognized stored value is unsupported, TLH neutralizes it with the provider-resolved bundled level when possible and warns TLH stored minor-agent effort "<level>" is not supported by <provider/id>; using bundled defaults for this run. If the bundled level is unavailable but off is supported, it warns TLH stored minor-agent effort "<level>" is not supported by <provider/id>; using explicit off for this run. A nonstandard stored value uses the same two outcomes: TLH ignored unsupported stored minor-agent effort "<value>" for <role>; using bundled defaults for this run. or TLH ignored unsupported stored minor-agent effort "<value>" for <role>; using explicit off for this run.

Generated opposite-provider fallbacks use the corresponding wording with generated fallback <provider/id> and that fallback will use, for example: TLH stored minor-agent effort "<level>" is not supported by generated fallback <provider/id>; that fallback will use bundled defaults for this run. or TLH stored minor-agent effort "<level>" is not supported by generated fallback <provider/id>; that fallback will use explicit off for this run. If neither bundled nor off is supported, TLH warns TLH stored minor-agent effort "<level>" is not supported by <provider/id>; no supported suffix can neutralize it, so the subagents runtime will drop the stored value for this run. If no model can be resolved for a nonstandard stored value, it warns TLH ignored unsupported stored minor-agent effort "<value>" for <role>; no supported model suffix could be emitted, so the subagents runtime will drop the value for a known model and fail open for an unknown model if this role is dispatched. When no bundled or current-session model is available for a thinking-only override, TLH warns TLH stored minor-agent effort "<value>" for <role> could not be capability-checked because no bundled or current-session model is available; the subagents runtime will apply its capability gate if the model resolves and fail open otherwise. The runtime gate resolves the known-model case by dropping unsupported values rather than emitting them; unknown or unresolvable models remain fail-open capability uncertainty, not residual known-model application.

Provider independence and headless use

A fixed model override cannot guarantee provider independence for code-reviewer, oracle, or contrarian. In a UI-enabled session, TLH asks for confirmation with the warning Provider independence is not guaranteed when a fixed model override is configured for this role. Status output repeats that warning for a saved fixed model. In a headless or otherwise non-UI context, TLH refuses the write because it cannot show the confirmation, reporting Cannot confirm the independence warning for <role> in this mode.

Supported effort values and undo

The complete set of values accepted by /subagent-settings set ... effort ... is off, minimal, low, medium, high, xhigh, and max. The command validates a new value against the currently effective model before writing. A stored value is applied when the resolved model advertises support; max specifically requires a thinkingLevelMap entry advertising it. If a model later stops supporting a stored value, TLH warns and uses bundled effort or explicit off when possible; if neither is supported, the runtime drops a known-unsupported value instead of emitting it, while unknown or unresolvable models fail open. max is a live value in bundled behavior too: the bundled developer role already configures max for its OpenAI default.

To undo a persistent change:

  • use /subagent-settings reset <role> model or reset <role> effort to clear one field;
  • use /subagent-settings reset <role> to clear both model and thinking for one bundled role while preserving any other keys in that role's entry;
  • use /subagent-settings reset-all to clear only model and thinking for bundled roles. It preserves other keys on those role entries and leaves unknown/non-TLH entries under subagents.agentOverrides untouched; or
  • restore the settings.json.bak-* file shown after a write by copying it back over the active settings.json.

The reset commands clean up empty role and override containers but do not remove unrelated settings. When diagnosing whether a saved setting took effect, /subagents-doctor shows first-party runtime diagnostics and subagent({ action: "status", view: "fleet" }) shows active dispatch status; neither command changes these overrides.

/reconcile

/reconcile lets you review and resolve drift between your saved model/effort overrides and TLH's current packaged defaults for the affected roles.

When the notice appears

At startup, TLH checks whether any packaged model or effort default has changed for a role you have overridden since you last ran /reconcile. When it finds changed roles it shows a one-line non-blocking notice:

TLH default model/effort changed for <role> — run /reconcile to review

The only trigger is TLH shipping an update that changes a packaged default for a role you have overridden. There is no periodic or scheduled reminder. An ignored notice reappears each launch until you make a decision via /reconcile.

When the session provider is unknown at startup, TLH defers: no comparison is performed, no baseline is written, and no notice appears until a provider is active.

Overrides that existed before baseline recording was introduced are silently backfilled on your first startup with a known provider, using the current packaged default as the baseline. The notice will fire on the next packaged-default change after that point — not for any changes that occurred before the backfill.

Acknowledgments are per-provider. A decision recorded under one provider does not suppress the notice for a different provider — switching providers means TLH treats you as not-yet-acknowledged for that provider until you run /reconcile under it, so a genuine later change to that provider's packaged default still notifies you.

Using /reconcile in the TUI

In an interactive TLH session, /reconcile opens a picker listing every affected role with both your override value and the current TLH packaged default. You can act on one role at a time or choose a bulk action for all.

For each role you have two choices:

  • Keep — acknowledges the new TLH packaged default and preserves your override unchanged. Non-destructive: nothing is written to or cleared from your settings. Keep requires a known session provider; when no provider is identified, TLH declines and asks you to rerun /reconcile in a session where a provider is active.
  • Reset — clears your stored override so the role resolves to TLH packaged defaults on the next dispatch. For primary agents, the packaged default is also applied to the active session immediately (subject to your tlh.primaryAgent.applyModel setting). Each reset is an independent write that creates a settings.json.bak-* backup; the notification shows the exact backup path.

Keep is always undoable because your override is never touched.

When resetting all, per-role failures are isolated — errors and unrecognised overrides are reported as a distinct category so a single failed clear does not abort the rest.

Reset is also undoable: restore the value through the native /model picker and press Ctrl+S to save the desired default for primary-agent overrides; use /subagent-settings set <role> model <provider/id> or /subagent-settings set <role> effort <level> for subagent overrides, or copy back the settings.json.bak-* backup shown in the notification.

Outside the TUI

In headless or non-TUI contexts, /reconcile prints a read-only drift summary showing all overridden roles, your values, and the current TLH packaged defaults. No state is written in this mode.

Packaged-default semantics

The packaged default shown is the canonical value for the active provider, resolved from TLH's own bundled agent catalog restricted to that provider's models. It is environment-independent — it reflects what this TLH release declares, not what models your current environment has available. A reported default may therefore name a model you cannot currently reach, and for roles that prefer an opposite-provider model (such as code-reviewer) the displayed value is a same-provider fallback rather than the exact model Reset would produce in a live dual-provider session. This does not block the decision; Keep and Reset work regardless of model availability.

Reconcile state

Acknowledgments are stored in the isolated TLH profile under tlh/reconcile-state.json (alongside settings.json). Each entry records the packaged defaults you acknowledged, keyed by role name and provider (byProvider), so a later provider switch does not inherit a prior acknowledgment. This file is written in three situations: when a correlated persisted model write creates or updates a primary override (normally the native /model picker Ctrl+S action, but Pi provider-auth and other programmatic persist: true writes are indistinguishable), via /subagent-settings, on startup when TLH backfills a missing baseline for a pre-existing override, and when you run /reconcile. Enter/session-only selections and direct session model applications do not write an override baseline. Do not edit it by hand.

/tickets

/tickets shows the current repo/worktree's read-only ticket workflow details from tk: ready, blocked, in-progress, active, and total counts, followed by one in-progress detail line or an In progress: list when multiple tickets are in progress. Each detail includes the ticket ID and its title when available. TLH strips terminal control sequences from titles and falls back to the ticket ID when a title cannot be resolved or is empty after sanitization.

/tokens

/tokens takes no flags or subcommands. Run it as /tokens to generate one local HTML token-spend report for the current session and open it on your machine.

The report is built from sanitized session analysis only. It omits raw transcript text, raw tool arguments, and raw tool-result payloads, and TLH tells you where the private local report directory lives so you can delete it when you no longer need the report.

/what-consumed-my-session-limit-and-tokens

/what-consumed-my-session-limit-and-tokens takes no flags or subcommands. Run it to generate a local HTML report covering TLH sessions within the current provider session-limit window across all projects under the same session root.

The report is built from sanitized session analysis only. It omits raw transcript text, raw tool arguments, and raw tool-result payloads. It uses the current subscription usage snapshot when available, with a trailing five-hour fallback when TLH cannot resolve an exact provider window.

Model-facing subagent tools

TLH ships the subagent tool as first-party runtime functionality. It is a model-facing tool, not a slash command you need to invoke manually. subagent supports only direct single or parallel execution, in the foreground by default or through TLH-tracked background mode with async: true, plus the closed action set list, get, status, interrupt, resume, steer, and doctor. Each child starts a fresh session; caller context, agent defaultContext, and turn budgets are not supported. TLH-tracked async work uses a detached OS child process managed by TLH, not the removed external pi-intercom detach request/result/control integration. Saved chains and mutating agent-management actions are not in the model-facing TLH contract.

The architect normally handles these tools for you. See subagents.md for dispatch fields, fresh-session/user/project-scope isolation, async control and durable resume behavior, native contact_supervisor, tool budgets, timeouts, diagnostics, acceptance, artifacts, migration, and undo steps. Project custom embedded-agent paths and trust rules are in custom-subagents.md.

Project custom subagents

Custom subagents are not a slash command. Project-owned agents use the exact direct Git-root path .tlh/agents/custom/<UPPERCASE-SLUG>.md; persist trust for the validated Git root, then request one naturally by slug from the architect. Generic profile, package, project .pi/.agents, configured-directory, and settings-based custom sources do not authorize targets. See custom-subagents.md for strict frontmatter, trust boundaries, snapshot behavior, resume limits, troubleshooting, and removal.

Packaged first-party extension commands

These commands ship inside the TLH package itself rather than through separately managed default-extension installs.

Command Extension Description
/annotate-last-message the-last-harness Open a native annotation window for the latest assistant message and send submitted feedback to the agent
/annotate-git-diff annotate-git-diff Open a native git-diff review window; clicking Submit sends review feedback to the agent, closing with unsent comments pastes a draft to the editor
/subagents-doctor subagents Show first-party subagent runtime diagnostics

/annotate-last-message

/annotate-last-message opens a lightweight native Glimpse annotation window for the latest completed assistant message on the current session branch.

Requirements

  • Run it from an interactive TLH session with editor access; it will not work in non-interactive/headless command contexts.
  • The current branch must already contain a completed assistant message with text. If the latest assistant turn is still running or has no text content, wait for a normal text reply before rerunning the command.
  • It needs a desktop session that can open a local native window. Headless shells and SSH-only sessions will not work unless they can display that window locally.
  • TLH packages the UI assets locally. /annotate-last-message does not require CDN access or general internet access just to render the annotation window.
  • Like the rest of TLH, it stays inside the isolated TLH profile and does not read or write normal Pi config under ~/.pi/agent.

Behavior

  • Finds the latest completed assistant message on the active session branch and shows it with line numbers plus section-level grouping.
  • Lets you leave overall, section, and inline comments in one lightweight first-party TLH window.
  • When you submit, TLH sends a structured planning-oriented feedback prompt directly to the agent as a follow-up message. Your existing editor text is left untouched.
  • Blank lines cannot be annotated inline; the inline-note button is not shown for empty lines.
  • It does not auto-apply code changes or silently mutate prior messages.
  • Use /annotate-last-message directly when you want to annotate the latest assistant reply.

Extension-local notes

See extensions/the-last-harness/annotate-last-message/README.md.

Troubleshooting and recovery

  • annotate-last-message requires interactive mode. → run the command from the TLH TUI rather than a non-interactive command context.
  • No assistant messages found on the current session branch. → wait until the branch has an assistant reply, then rerun.
  • Latest assistant message is incomplete (...) → wait for the assistant turn to finish, then rerun.
  • Latest assistant message has no text to annotate. → rerun after a normal text reply; tool-only or empty assistant turns cannot be annotated.
  • A last-message annotation window is already open. → return to the existing window or close it before opening another.
  • Annotation failed: Glimpse host not found ... → the local native window runtime is unavailable. Run tlh update (or reinstall TLH) to restore the packaged dependency, then rerun from a machine/session that can open native windows.
  • There is no separate tlh defaults toggle for /annotate-last-message because it ships inside TLH itself.

/annotate-git-diff

/annotate-git-diff opens a native Glimpse review window for the current git repository.

Requirements

  • Run it from inside a git repository.
  • It needs a desktop session that can open a local native window. Headless shells and SSH-only sessions will not work unless they can display that window locally.
  • TLH packages Monaco and Tailwind locally for this UI. Monaco editor, syntax-highlighting tokenizers, and the worker source are all inlined into the review window's HTML at build time, so the window works from any WebView origin (including null-origin WebViews) without runtime file-system fetches. /annotate-git-diff does not require CDN access or general internet access just to render the review window.
  • Like the rest of TLH, it stays inside the isolated TLH profile and does not read or write normal Pi config under ~/.pi/agent.

Behavior

  • Review branch diffs, individual commits (including working-tree changes), or the full file snapshot from one window.
  • Leave inline, file-level, and overall comments.
  • Clicking Submit sends a structured review-feedback prompt directly to the agent. If you close the window with comments not yet submitted, TLH pastes the draft prompt into the editor instead — so an accidental window close cannot fire a new agent turn. It does not auto-apply code changes or mutate your normal Pi profile.

Attribution

TLH's first-party implementation adapts the MIT-licensed @ryan_nookpi/pi-extension-diff-review work and preserves inspiration credit to badlogic/pi-diff-review. For extension-local notes, see extensions/annotate-git-diff/README.md.

Troubleshooting and recovery

  • Review failed: Not inside a git repository. → change into a git repo and rerun /annotate-git-diff.
  • No reviewable files found. → make or fetch reviewable changes/commits, then rerun /annotate-git-diff.
  • Review failed: Glimpse host not found ... → the local native window runtime is unavailable. Run tlh update (or reinstall TLH) to restore the packaged dependency, then rerun from a machine/session that can open native windows.
  • If the window says TLH could not load its packaged review assets, the monaco-editor package is missing or corrupt in your TLH install (Monaco editor, syntax-highlighting tokenizers, and the worker source are all inlined at build time, not fetched at runtime). Reinstall TLH (or run tlh update) to restore the package, then rerun /annotate-git-diff. If the problem persists after reinstalling, please file an issue.
  • There is no separate tlh defaults toggle for /annotate-git-diff because it ships inside TLH itself. If you customize TLH packages manually, keep that change inside the isolated TLH profile rather than ~/.pi/agent.

Packaged prompt templates

These slash commands come from prompt templates bundled inside TLH. They insert canned prompt text into the editor for you to review and send.

Command Description
/analyse-tlh-sessions Analyse the past week of tlh sessions for notable issues without changing files.
/investigate-pr-comments Check PR comments and verify whether each one is valid.
/merge-origin-main-into-this-branch Merge origin/main into this branch.
/rebase-this-branch-onto-origin-main Rebase this branch onto origin/main.

Bundled extension commands

These commands are provided by bundled default extensions and are visible in TLH autocomplete. They are available when the relevant extension is installed and active.

Command Extension Description
/context pi-context-inspector Open a local HTML breakdown of where this session's context is going
/fast pi-fast Toggle OpenAI Codex Fast mode for eligible ChatGPT-auth GPT-5.4, GPT-5.5, and GPT-5.6 sessions
/mcp pi-mcp-adapter Show MCP server status
/mcp-auth pi-mcp-adapter Authenticate with an MCP server (OAuth)
/transcribe pi-transcribe Configure local speech-to-text model, languages, microphone, and shortcut settings

/transcribe and Ctrl+Alt+Z

Run /transcribe once to choose and confirm a local model; pi-transcribe downloads the model after confirmation. While TLH has terminal focus, press Ctrl+Alt+Z to start and stop microphone recording. The transcription is inserted at the editor cursor, and Esc cancels recording. This is a terminal shortcut, not a global OS hotkey. The extension also provides the transcribe_file tool for local audio or video; file transcription requires ffmpeg.

Bundled skill commands

These bundled skills ship with TLH and remain visible in TLH autocomplete.

Command Description
/skill:cmux-cli Load the bundled cmux CLI skill for socket, workspace, pane, browser, and automation workflows
/skill:herdr Load the bundled Herdr skill for explicitly requested pane, tab, workspace, and agent control
/skill:show-me Load the bundled show-me skill for visual explanations with diagrams, sketches, and focused HTML artifacts
/skill:tmux Load the bundled tmux skill for session/pane control, output capture, key sending, and prompt monitoring

Hidden autocomplete commands

These commands are registered and fully functional, but deliberately excluded from TLH autocomplete suggestions. Type them in full to invoke them.

Hidden upstream Pi built-ins

Command Description
/changelog Show upstream Pi changelog entries; use /tlh-changelog for TLH release notes
/import Import and resume a session from a JSONL file
/scoped-models Enable or disable models for Ctrl+P cycling

Hidden skill commands

Command Description
/skill:librarian Load the bundled librarian skill by name without surfacing it in TLH autocomplete

Hidden bundled extension commands

Command Extension Description
/curator pi-web-access Toggle or configure the search curator workflow
/quiet-tools pi-quiet-tools Toggle one-line collapsed invocations for built-in tool rows
/search pi-web-access Browse stored web search results
/websearch pi-web-access Open the web search curator

TLH CLI subcommands

These are tlh command-line subcommands, distinct from the /slash commands used inside a TLH session.

tlh sessions

tlh sessions is a read-only session analysis tool. It emits JSON to stdout so you can pipe to jq. Run tlh sessions --mode per-session for a per-session summary of tool-pair statistics and coverage, or tlh sessions --mode per-tool for aggregated per-tool statistics across all sessions. Raw paths, cwd values, and project labels are omitted by default; pass --include-paths only when you need them for concrete evidence. tlh sessions never reads run-history.jsonl and never writes to session files.

Managing bundled extensions

Individual bundled extensions can be disabled without affecting the others. Use tlh defaults to list and manage opt-outs:

tlh defaults list        # show installed defaults and opt-out status
tlh defaults disable <id>  # disable a separately managed default extension (e.g. tlh defaults disable context-inspector)
tlh defaults enable <id>   # re-enable a disabled extension

Disabling an extension removes it from the installed packages list on the next tlh update run. Its slash commands will no longer be available in new sessions after the extension is unloaded. This opt-out flow applies to separately managed default extensions only (those in config/default-extensions.json), not first-party packaged extensions like notify, /annotate-last-message, or /annotate-git-diff.

Configuring the notify extension

The notify extension is first-party and ships bundled with TLH. It cannot be managed with tlh defaults disable; instead, disable it by setting "enabled": false in its config file.

Config files are merged at runtime — project config overrides global config:

  • ~/.the-last-harness/agent/extensions/notify.json (default path; may differ if you used --agent-dir during install)
  • <project>/.pi/notify.json (only when the project is trusted)

Example to disable notifications entirely:

{
  "enabled": false
}

Key config fields (see extensions/notify/README.md for the full list):

Field Default Description
enabled true Master on/off switch
onlyWhenInteractive true Skip notifications in non-UI (print) mode
suppressWhileActive true Hold all notification channels while background subagent work is still running; a notification fires only once the session is genuinely waiting on you, not merely between turns. Set to false to notify on every turn completion regardless. Has no effect when the TLH activity tracker is absent.
title "tlh" Notification title
body "Ready for input" Notification body