[Bug]: ponytail does not detect caveman conflict — both inject contradictory instructions

[lg320531124]

Bug Description

When ponytail and caveman are both active, they both inject additionalContext on UserPromptSubmit. The model receives two contradictory style instruction sets every turn:

• Caveman: "CAVEMAN MODE ACTIVE (full). Drop articles/filler/pleasantries/hedging. Fragments OK."
• Ponytail: "PONYTAIL MODE ACTIVE (full). The ladder enforced. Stdlib and native first."

Previous issues (#18, #68, #193) asked "can they coexist?" and were closed with "yes, they stack". But the actual bug is that both plugins inject per-turn style reminders simultaneously, causing the model to receive contradictory instructions about output style. This results in unpredictable, inconsistent outputs.

What caveman already does

Caveman PR #577 added conflict detection: when ponytail (or other compression plugins) is detected in settings.json, caveman reduces its per-turn injection to a one-liner instead of the full ruleset. But ponytail does NOT reciprocate — it always injects its full ruleset regardless of whether caveman is active.

Proposed Fix

In ponytail-mode-tracker.js, before emitting the per-turn additionalContext, check settings.json for known compression plugins (caveman, etc.). If detected, reduce the per-turn injection to a minimal one-liner:

PONYTAIL MODE ACTIVE (full) — reduced injection (another compression plugin detected). Ponytail rules still apply per SessionStart injection.

The SessionStart full ruleset is unchanged — ponytail rules still apply, just the per-turn reinforcement is reduced to avoid contradictory instructions.

This mirrors the approach caveman already uses in its detectCompressionConflict() function.

Testing

With caveman installed in settings.json:

• Before: model receives both "CAVEMAN MODE ACTIVE (full). Drop articles..." AND "PONYTAIL MODE ACTIVE (full). The ladder enforced..." every turn → contradictory instructions
• After: model receives "CAVEMAN MODE ACTIVE (full) — reduced injection" + ponytail full injection = no contradiction

[changhee-somite]

Adding a data point, since this issue is the one that owns the contradictory-injection problem.

The contradiction also exists inside ponytail's own injected text, with no second plugin involved.

At 4.8.1 the SessionStart injection is 885 words, and two of its sections disagree about whether ponytail governs prose:

Boundaries

Ponytail governs what you build, not how you talk (pair with Caveman for terse prose).

Output

Code first. Then at most three short lines: what was skipped, when to add it. No essays, no feature tours, no design notes. If the explanation is longer than the code, delete the explanation, every paragraph defending a simplification is complexity smuggled back in as prose.

## Boundaries is 37 words disclaiming prose. ## Output is 80 words governing it: a line cap, a content template, and a rule that explanatory paragraphs are debt. Both ship in the same injection, at every level.

The maintainer's answers in #68 and #193 — "Ponytail cuts code. Caveman cuts user response." — match Boundaries rather than Output. #298 asked this question directly and was closed on the Boundaries line, without the Output section being opened.

It binds in practice. I split my own local Claude Code transcripts by whether a session carried the SessionStart injection, held the model fixed, and counted only turn-final assistant messages with a rule-based parser. Claude Opus 4.8 straddles my install date:

	no injection (n=935)	injected (n=1,736)
median words per message	286	240
sentences over 25 words	22.2%	20.8%
passive voice	8.5%	8.0%
typographic marks per 100 words	19.6	20.7

This is observational, not randomised. Injected sessions were coding-heavy by selection and coding turns run short regardless, so 16% is an upper bound rather than an estimate. The direction is the Output section's direction, though, and message length is not something the Boundaries line claims to touch.

Why it matters for the fix proposed here. Detecting caveman and reducing the per-turn reinforcement would not resolve it, because the length rule arrives in the SessionStart ruleset that the proposal explicitly leaves unchanged. Two changes would:

1. Bring ## Output into line with ## Boundaries: scope it to the code-adjacent note it describes — what was skipped, when to add it — and drop the general prohibitions on explanation length, leaving prose to whichever prose plugin the user chose.
2. Or keep the rule and drop the Boundaries disclaimer, so anyone pairing ponytail with a prose plugin knows there is a real overlap to arbitrate.

Either one resolves this without cross-plugin detection, and (1) also removes the need for ponytail to know what else is installed.

One note for #664 and #668. I checked whether #668 changes this and it does not: its mode markers wrap only the lite/full/ultra stance blocks, ## Output stays in the ungated core, the added text says "Core rules above always apply", and the lite override list does not mention the three-line rule. The prose cap survives at every level after that PR lands.

Entirely possible the answer is that the Output rule is the point of the plugin and the Boundaries line is what should change. Either way, having the two agree would help anyone pairing this with a prose style.

[changhee-somite]

One more piece, and it is the cleanest of the three, because it is a disagreement between two files rather than an interpretation of one.

4.8.4 added a prose exclusion to the skill description, and the injection strips it.

The current skills/ponytail/SKILL.md frontmatter ends:

Do NOT use for non-coding requests (general knowledge, prose, translation, summaries, recipes).

That is the right call, and it agrees with ## Boundaries. But filterSkillBodyForMode removes the frontmatter before anything is emitted:

const withoutFrontmatter = String(body || '').replace(/^---[\s\S]*?---\s*/, '');

So the exclusion governs only skill triggering. The SessionStart hook then injects the remaining body — ## Output included — into every session unconditionally, with no knowledge of whether the turn involves code. Ask for a summary or a translation and the three-line cap and "explanatory paragraphs are debt" arrive anyway, in the same session where the description says not to apply them.

Worth noting how stable that section is: across the 197 commits between 4.8.1 and 4.8.4, skills/ponytail/SKILL.md changed by 9 insertions and 6 deletions, and ## Output is byte-identical. The description grew a prose exclusion while the prose rules stayed exactly where they were.

So the scope is now disclaimed in two places — ## Boundaries in the body, and the description — and still shipped in a third. A fix consistent with both disclaimers would be to scope ## Output to the code-adjacent note it actually describes (what was skipped, when to add it) and drop the general prohibitions on explanation length. That needs no cross-plugin detection and no change to how the hook works.

Also relevant to the version question in this thread: SubagentStart landed in 4.8.2, so 4.8.1 is the last release that injects only at session start. Anyone comparing behaviour across installs should check which side of that boundary they are on.