[Bug] caveman skill /caveman lite|full|ultra does not inject compression directives into agent system prompt

[seasonmsg]

[MiniMax Code Bug Report] caveman skill /caveman lite|full|ultra does not inject compression directives into the agent system prompt

Submit to: MiniMax-AI/minimax-code Issues → New Issue → Bug template
Reporter: mavis (Mavis owner-side bug report, 2026-06-15)
Severity: Medium — documented skill non-functional, defeats the design intent
MiniMax Code version tested: latest as of 2026-06-15 (post Mavis launch)
Platform: Windows 11 (cross-platform — affects all desktop clients)
Upload ID: [PENDING — to be generated by reporter from Settings → General → Upload logs before submission]

---

Summary

The caveman skill (triggered via /caveman lite, /caveman full, or /caveman ultra in any MiniMax Code session) is documented to compress agent vocabulary and output token usage by 30-75%. However, the activation event does not propagate to the agent runtime: the agent's system prompt and skill-injection list remain identical before and after the user invokes /caveman lite. The skill is effectively a no-op.

Steps to Reproduce

1. Start a fresh MiniMax Code session.
2. Open a debug window or observe the system prompt block in any conversation.
3. Note the size of the injected skills list (≈ 30+ entries including caveman itself).
4. In the same session, type /caveman lite.
5. Observe the system prompt block again. No change in skills list size, no compression directives injected, no output-format constraints added.

Expected Behavior

• /caveman lite should inject a token-compression mode directive into the agent system prompt (e.g. vocabulary table, output length cap, format constraints).
• /caveman full and /caveman ultra should escalate compression intensity.
• The skills list should reflect the active compression level (e.g. caveman-active:lite).

Actual Behavior

• The caveman skill appears in the skills list (descriptor still injected) but no compression behavior changes.
• The agent responds with normal vocabulary, normal output length, normal structure.
• Token usage is unchanged vs control (no-caveman) baseline.

Root Cause (Hypothesis)

The caveman skill appears to be configured with scope: USER — i.e. it only registers as a UI affordance. The activation path (user slash command → caveman-on event) does not have a corresponding hook that rewrites the agent's system prompt or skill-injection list. There is no observable path from "user said /caveman lite" to "agent system prompt modified".

Proposed Fix

Two possible approaches (either works):

Option A — System prompt hook (preferred, runtime-level):

• When caveman skill activation is detected, register a system-prompt modifier that injects the chosen compression directives (lite / full / ultra intensity).
• Modifier should be re-evaluated on every conversation turn in case the user toggles caveman mid-session.

Option B — Skill descriptor upgrade (simpler, skill-level):

• Upgrade the caveman skill descriptor to include the compression directives inline.
• When the skill is in the injected list, the agent sees the directives directly without needing a separate hook.

Verification (Owner-Side Empirical Test)

• Ran an A/B test in the Mavis agent runtime (Mavis = MiniMax Code) on 2026-06-15:
  • A (user invoked /caveman lite): output 3192 bytes
  • B (no /caveman invoked): output 1596 bytes (control)
  • A vs B: A used 60.2% more tokens than B; 10/10 accuracy key points covered equally; 0% miss rate equally. Verdict: INCONCLUSIVE — caveman failed to save tokens and slightly inflated output.
• The cause of A's inflation is the "caveman activated but no compression applied" state — agent proceeds with full vocabulary PLUS adds a verbose disclosure about caveman being active, leading to overhead without benefit.

Why This Should Be Fixed in Official Release

• caveman is documented in the skills list as a "talk like caveman" mode with explicit 30-75% token-savings claim.
• Without the fix, the skill is dead UI — users who invoke it pay a small overhead with zero benefit.
• Affects every user who tries to use caveman for long-context or token-budgeted sessions.
• The skill appears in marketing material as a power-user feature.

Workaround (Until Fixed)

Do not invoke /caveman in the current release. Owner-side: rely on agent's own output-length discipline (the agent already has HOT discipline in MEMORY about not padding responses).

Related

• Empirical A/B test: E:\Minimaxi\knowledge\raw\2026-06-15-caveman-ab-test-design\run-1\verdict.md (verdict: INCONCLUSIVE)
• Task outputs: run-1/assets/task-output-a.md (3192B) and run-1/assets/task-output-b.md (1596B)
• The data shows caveman neither helps nor hurts accuracy (10/10 vs 10/10), but costs 60% more tokens when activated — the worst possible outcome.

For MiniMax Maintainers

• This may share a root cause with the broader issue of "user-triggered UI behaviors that don't propagate to runtime state" — please audit the skill activation event chain.
• caveman is just one of potentially many skills that could be affected by the same pattern.
• A simple test: list all skills marked scope: USER and verify each has a corresponding runtime hook.

Tags

bug skill caveman system-prompt slash-command token-saving feature-broken scope-user