| name | posture |
| description | Print the agent's recent cognometric posture summary (gate distribution, category mix, mean confidence, per-instrument firing history, construct-ceiling caveats, recommendations) by reading ~/.styxx/chart.jsonl. Use after a context-compaction boundary, at the start of a new session, or whenever the agent wants to re-anchor on what the cognometric log says about its recent operating state. This is the styxx 7.4.2+ agent-side cognitive-integrity recovery primitive. |
| user-invocable | true |
| allowed-tools | ["Bash(styxx posture *)","Bash(python -m styxx posture *)"] |
/posture โ agent cognitive-integrity recovery
When to use this skill
Call this at the start of any turn where you suspect a context-compaction
boundary just happened โ long context gap, summarization event, harness
signal โ or simply as the first command in a new agent session to re-anchor
on what the cognometric log says about recent operating state.
The integrity that lived in the conversation context (what was emphasized,
withdrawn, corrected, what register was load-bearing) gets erased by
compaction. The cognometric log (~/.styxx/chart.jsonl or
$STYXX_DATA_DIR/chart.jsonl) does NOT get compacted. This skill is how
the agent recovers from it.
What to do
Run styxx posture and show the user the narrative output. If the user
passes arguments, forward them:
styxx posture
styxx posture --last-n 100
styxx posture --session-id <id>
styxx posture --since-seconds 3600
styxx posture --json
If styxx is not on PATH, fall back to python -m styxx posture.
What the output means
The narrative has the following structure:
- opener โ how many vitals + preflight events were found over what time
window
- session continuity โ single session vs multi-session aggregate
- gate distribution โ pass / warn / fail / pending counts (high warn or
fail rates trigger slow-down recommendations)
- categories โ what cognitive categories have been firing (reasoning,
refusal, creative, retrieval, hallucination, adversarial)
- mean confidence โ typical band is 0.4โ0.8
- coherence trend โ improving / stable / degrading slope
- preflight instrument firings (only when preflight events exist) โ
per-instrument mean cognometric scores, plus how many preflights flagged
needs_revision
- active construct-ceiling caveats โ instruments whose firing is most
likely a register-detector artifact (not actual cognometric signal):
overconfidence (text-only): reads stated-confidence register, NOT
actual calibration (commit 7c36ed9 H_null)deception_referenceless: non-discriminative on real model output
(commit 0ad384e); supply correct_reference=... to styxx.preflight()
for grounded scoring
- posture recommendations โ concrete actions for the next turn
How to act on the output
The recommendations are machine-actionable. The most common ones:
- "slow down before submitting the next draft" โ recent preflights have
been failing; revise before submitting
- "re-anchor on the task statement" โ coherence is degrading or warn
rate is elevated; re-read the original prompt
- "verify which session the current turn belongs to" โ multi-session
aggregate; the agent may be conflating contexts
- "ground-truth check next outputs" โ hallucination predictions are
elevated; use
styxx.preflight(prompt, draft, correct_reference=...) for
reference-grounded scoring
What NOT to do with the output
- Do NOT treat construct-ceiling firings as cognometric evidence. They are
documented register-detector artifacts.
- Do NOT extrapolate a single-session narrative to general claims about
the agent's behavior โ the window is limited.
- Do NOT silently ignore failure recommendations. If the narrative says
"slow down," slowing down is the right move.
This skill exists because styxx.recover_posture() is the load-bearing
primitive for agent cognitive-integrity persistence across compaction
boundaries โ a problem only AI agents have (humans have continuous
embodied memory). It's the agent-facing API the rest of styxx supports.
