Add cave-code command and skill support

[romgenie]

Summary

• sync CompleteTech-LLC-AI-Research/letta-code main with upstream Letta through 31469dfc
• add the cave-code persona prompt and register it as the caveman internal personality preset
• include the cave-code bundled skill with parseable frontmatter
• add a built-in /caveman slash command that ships with installed Letta Code
• keep /caveman as a conversation-local mode switch and send its mode-switch request without advertised client tools
• preserve /caveman no-tools behavior through direct reentry, manual approval-result continuations, and queued approval continuations after interrupts
• support lite, full, ultra, wenyan-lite, wenyan-full, and wenyan-ultra as public modes, with wenyan kept as an alias for wenyan-full
• accept cave-code as the public nickname alias for the caveman internal personality preset
• update CLI help and personality tests for the new preset
• address PR Add cave-code command and skill support #14/Add cave-code command and skill support #15/Add cave-code command and skill support #16/Add cave-code command and skill support #17/Add cave-code command and skill support #18 cleanup: export caveman mode rules for alignment tests, bind each SKILL.md mode example to its matching quoted example, align bundled skill examples, document persona_caveman.mdx, remove redundant human-content overrides, make the non-interactive /caveman result explicit, remove the brittle server-tool denylist from the mode-switch prompt, document advertised-tool suppression and pending-approval lifecycle, fix Windows bun run build, clear existing lint warnings, split ref lifecycle comments, and reset cancelled approval queues back to normal client-tool advertising

Validation

• bun test src\tests\agent\personality.test.ts src\tests\cli\caveman-command.test.ts src\tests\channels\runtimeDeps.test.ts src\tests\cli\approval-recovery-wiring.test.ts
• bun run typecheck
• bun run lint
• bun run build

Replaces #18.
Replaces #17.
Replaces #16.
Replaces #15.
Replaces #14.
Replaces #13.
Closes #1

[romgenie]

PR #18 carryover is addressed:

• pendingTranscriptStartLineIndexRef and pendingApprovalSuppressClientToolsRef each now have their own one-liner comment (src/cli/App.tsx:1197-1202), with the transcript ref described as "set at user-turn start and kept through approval reentry" and the suppression ref scoped to the requires_approval → approval-result window.
• handleCancelApprovals explicitly resets both the ref and the queue metadata to suppressClientTools: false, and approval-recovery-wiring.test.ts pins that behavior with a new assertion.
• The SKILL.md alignment test now requires both opening and closing quotes ("${example}"), the server-tool denylist is gone from the mode-switch prompt (test asserts server-side tools such as is absent), the non-interactive /caveman branch returns success: false with an explicit "must be used inside the interactive CLI" message, and suppressPreparedClientTools carries a comment explaining why the context snapshot is preserved.
• Windows chmodSync in build.js, dead areToolSetsEqual dropped from reconcileExistingAgentState.ts, ConversationListItem replaces any in generate-memory-viewer.ts, and the persona_caveman.mdx row is in prompts/README.md.
• The personality.ts refactor from ternary chains to PERSONA_TEMPLATE_BY_ID / HUMAN_TEMPLATE_BY_ID lookups with a PERSONA_CONTENT_OVERRIDES map for codex/claude is a nice cleanup and keeps the caveman wiring symmetrical with the other presets.

One small observation, not blocking:

[romgenie]

The queued-approval path now preserves suppressClientTools across an interrupt + new-user-message resubmission by reading from queuedApprovalMetadataRef. That's correct as a preservation of state, but it does mean: if a /caveman turn ever reached requires_approval and the user interrupted with a fresh regular message, that fresh message's turn would inherit the suppression flag and be sent with clientTools: [].

In practice this can't happen — /caveman submits with empty clientTools, so the agent has no client-side tools to request approval for, and server-side tools don't route through this approval path. So this is preserving correctness for a case that shouldn't arise rather than a live bug.

If you ever widen suppressClientTools to a mode that can legitimately trigger approvals, this branch will silently extend the suppression into the interrupting message. Worth a short comment near suppressClientToolsForSubmission = queuedSuppressClientTools; that the flag intentionally rides along with the queued continuation — or, if you prefer, gating the carry-over on initialInput.length === 1 (queued-only, no user message appended) so a new user message always starts fresh. Either one is fine; happy to leave as-is.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: f54f3cd7ce

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Preserve tool suppression when resuming restored approvals

This continuation path reads pendingApprovalSuppressClientToolsRef.current, but restored approval flows (recoverRestoredPendingApprovals) never initialize that ref from the restored batch before users continue approvals. In a resumed conversation where pending approvals originated from a /caveman mode-switch turn, this call can send suppressClientTools: false and re-advertise client tools, letting the model invoke tools during a continuation that should remain tool-suppressed. Please carry suppression metadata through restored approval recovery (or pass it explicitly) before this reentry call.

Useful? React with 👍 / 👎.