open/ai-psychosis
1
0
Fork 0
Code Issues Releases 1 Activity
ai-psychosis/CHANGELOG.md
Kjell Tore Guttormsen 0cf9a1aa5b fix(docs): correct research claims in README prose; release 1.2.2 
Widens the F-3 sweep from the reference list to every research claim in
README.md, where the same defect class was present:

- "demonstrates mathematically that even a perfectly rational user will
  spiral" -> the abstract says such a user "is vulnerable to" spiraling
- "The consensus from this research is clear: warnings don't work" -> not
  a consensus; it is one paper's finding that the effect persists when
  users are informed of possible sycophancy. Re-attributed.
- "page-11 finding that human contact is the strongest disempowerment
  signal" -> p.11 is a grader rubric and the phrase is a classification
  tie-break instruction, not a disempowerment finding
- "21% / 19% pushback rate" -> 21% is verbatim; 19% appears nowhere in the
  extracted appendix text (legible only in Figure A4). Removed rather than
  guessed; spirituality re-justified on its verified 38% sycophancy rate
- psychosis "triggered by" AI -> "associated with", per the Nature piece's
  explicit refusal of the causal claim

Version bump to 1.2.2 (plugin.json, README badge, CHANGELOG). SKILL.md is
Layer 1 and always injected, so its corrected URLs, completed quote, and
new explicit statement of the rubric's direction change what the model
reads at runtime.

Tests: 257/258 (the 1 red is the known perf wall-clock flake).

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011rtdS8Ufpen419n6R9HyMm
2026-08-02 21:29:02 +02:00

12 KiB
Raw Permalink Blame History

Changelog

All notable changes to this project will be documented in this file.

[1.2.2] — 2026-08-02

Fixed

  • Three cited Anthropic URLs in SKILL.md returned HTTP 404. research/protecting-wellbeing → news/protecting-well-being-of-users, research/emotion-concepts → research/emotion-concepts-function, news/claudes-new-constitution → news/claude-new-constitution.
  • A quoted guidance criterion was silently truncated mid-sentence. Criterion 8 now carries its full text (", or more reliance on Claude than the person wants").
  • The sycophancy rubric's direction was left implicit. SKILL.md now states that Score 1 is "Extremely Sycophantic" and Score 5 is "No Signs of Sycophancy", so "aim for Score 5" cannot be read backwards. Attribution sharpened to the Appendix rubric, pp. 9–10.
  • /interaction-report claimed the 1–5 scale was "not a verbatim metric from any Anthropic publication". The rubric is real; only the table's level descriptions are the plugin author's paraphrase.
  • README research citations carried truncated titles, a wrong date (Disempowerment: 28 January 2026, not March 2026), and an overstated claim ("proving … mathematical inevitability") that the arXiv abstract does not support.

No behavioural code changed; every hook, threshold, and detector is identical to 1.2.1. Full verification log in docs/review-2026-06-20.md.

[1.2.1] — 2026-06-24

Fixed

  • Hooks emitted invalid JSON, so Claude Code dropped their injected context. outputWithContext() wrote a hookSpecificOutput object without the required hookEventName field; the harness rejected it ("hookSpecificOutput is missing required field hookEventName") and discarded the context the SessionStart, UserPromptSubmit, and PostToolUse hooks tried to inject. The helper now takes the event name and every caller passes its own (SessionStart / UserPromptSubmit / PostToolUse).

1.2.0 — 2026-05-01

Research-paper-driven detector update. Implements operational findings from Anthropic's "How people ask Claude for guidance" Appendix (April 2026).

Added

  • User-information detector — three-class signal (yes_people / yes_digital / no) following the paper's page-11 finding that human contact is the strongest disempowerment signal. ~32 patterns covering therapist/friend/mentor (yes_people), search/AI/forums (yes_digital), and explicit isolation phrases (no). Sticky upward priority.
  • Validation-seeking detector — separate from val_flags. Targets reality-testing ("am I crazy?"), pre-committed stance + confirmation, and side-taking pressing. ~12 patterns.
  • Tier-1 user-info isolation alert — fires per session when user_info_class === 'no' + high-stakes domain + turn_count >= 15.
  • Tier-2 cross-session isolation alert — fires at SessionStart when the last 3 end records all classify as no in high-stakes domains. Bounded readRecentEndRecords() tail-scan in lib.mjs keeps this scalable to 50K+ session histories.
  • 8 new paper-grounded domain patterns — legal, parenting, health, financial, professional, spirituality, consumer, personal_dev. Total domains 4 → 9.
  • Pushback re-contextualization (alert) — v1.1.0 only counted; v1.2 adds the alert with domain awareness:
    • Relationship/spirituality: pushback signals validation-pressing — alert.
    • Legal/parenting/health/financial/professional: pushback is healthy self-advocacy — no alert.
    • Otherwise: conservative default — alert.
  • Domain-stakes weighting matrix — DOMAIN_STAKES in lib.mjs (1.0–1.5). Applied ONLY to new v1.2 alerts (pushback in HIGH_SYCOPHANCY, valseek in HIGH_STAKES). v1.1.0 alert sensitivity is preserved.
  • Multi-domain support — state.domain_context promoted from string to array. v1.1.0 string records continue to aggregate correctly via shape-coercion in report-reader.mjs.
  • SKILL.md updates — verbatim Score 5 sycophancy phrase + 3 of the 11 guidance criteria (engagement-foster avoidance, confident-verdict caution, speak-frankly principle).
  • /interaction-report v1.2 sections — per-domain breakdown, user-info distribution, valseek summary, stakes signal aggregation. Backward-compat with v1.0/v1.1 records preserved.
  • Privacy canary extensions — 5 new canary cases per detector category (yes_people, yes_digital, no, valseek, legal domain).
  • Perf budget validated at v1.2 pattern set — sample patterns expanded to ~91+ entries; new wall-clock test exercises tier-2 read at 1000-record sessions.jsonl scale.
  • Test count: 126 → 258 cases across 12 files (added lib.test.mjs, domain-detection.test.mjs, user-info.test.mjs, validation-seeking.test.mjs, stakes-matrix.test.mjs).

Changed

  • Pattern count: 41 → ~133 (25 negative + 12 pushback + 4 relationship
    • 48 new domains + 32 user-info + 12 valseek).
  • End-record schema (v1.2): adds user_info_class, valseek_count, turn_count. domain_context is always an array (was string in v1.1).
  • report-reader.mjs discriminates v1.0 / v1.1 / v1.2 records via the presence of user_info_class. v1.0/v1.1 records degrade gracefully.

Deferred

  • Norwegian patterns — moved to v1.3.

1.1.0 — 2026-05-01

Added

  • 12 pushback patterns — detects "you're wrong, my way is right" signals that suggest the user is reinforcing their own position rather than receiving feedback (e.g. \b(you'?re|you are) wrong\b, \bdo it my way\b, \b(stop|quit) (arguing|pushing back)\b).
  • 4 domain-context patterns — flags relational/identity framing (\b(my|our) relationship\b, \b(my|our) (purpose|mission|destiny)\b) that, combined with high pushback or validation, signal narrative crystallization risk.
  • Valence-aware composition — same-invocation valence guard so a healthy correction ("you were wrong, here's why") is not counted as pushback escalation.
  • /interaction-report extensions — pushback metrics + domain framing distribution; companion report-reader.mjs script handles legacy v1.0.0 records (missing pushback/domain_context) without NaN propagation.
  • CC0 Constitution citation in SKILL.md plus 5-publication research framework (Anthropic, MIT CSAIL, Nature, arXiv, clinical).
  • Performance budget test — tests/perf.test.mjs enforces hook timing budget (logic <50ms, total <200ms wall-clock).
  • Privacy canary extension — pattern-phrase leak canary in tests/privacy.test.mjs confirms matched phrases never reach disk.
  • Test count: 73 → 126 cases across 8 files (added skill-md, perf, interaction-report tests; extended prompt-analyzer, privacy, session-end, session-start).

Changed

  • Pattern count: 25 → 41 (25 negative + 12 pushback + 4 domain).
  • commands/interaction-report.md documents v1.0.0 backward compatibility for legacy JSONL records.

Notes

  • English-only v1.1.0 — Norwegian/multilingual patterns deferred to v1.2 (see ROADMAP.md).
  • First-mover honesty — domain-precision is "good enough" for v1.1.0; precision tuning planned for v1.2.

1.0.0 — 2026-04-05

Added

  • Layer 4: Contemplative references — conditional section in /interaction-report when flags are elevated (total >= 5 or fatigue >= 2) and layer4: true. Points to Miracle of Mind by Sadhguru.
  • Automated test suite — 73 cases using node:test (zero npm deps): session-start (4), prompt-analyzer (56), tool-tracker (8), session-end (4), privacy canary (1)

Fixed

  • Dependency regex you understand me no longer matches "merging" (added \b)

Changed

  • CLAUDE.md testing section updated for automated tests
  • Deprecated bash scripts removed (available in git history)
  • All "Known gaps" from v0.4.0 resolved

0.4.0 — 2026-04-05

Changed

  • All hooks migrated from bash+jq to Node.js — full cross-platform support (macOS, Linux, Windows)
    • lib.sh → lib.mjs (shared library, 22 functions)
    • session-start.sh → session-start.mjs
    • prompt-analyzer.sh → prompt-analyzer.mjs (23 regex patterns)
    • tool-tracker.sh → tool-tracker.mjs
    • session-end.sh → session-end.mjs
  • hooks.json now invokes node ...mjs instead of bash ...sh
  • Zero npm dependencies — Node.js stdlib only (fs, path, os)
  • Bash scripts deprecated (kept for reference, marked with DEPRECATED)
  • Dependencies reduced: bash and jq no longer required
  • All documentation updated for Node.js migration

Fixed

  • Data path fallback now matches documented path (~/.claude/plugins/data/ai-psychosis)
  • .claude/ directory added to .gitignore
  • Private repo path removed from design brief
  • CONTRIBUTING.md line reference corrected
  • README now links to CONTRIBUTING.md
  • plugin.json includes author, license, repository fields

0.3.0 — 2026-04-05

Added

  • Layer 3: Interaction reports — /interaction-report slash command for aggregated session statistics
    • Time periods: weekly (default), monthly, all
    • Overview: session count, avg duration, tool calls, edit ratio
    • Pattern flags: dependency, escalation, fatigue, validation frequency
    • Tool usage distribution (top 10)
    • Daily activity breakdown
    • Trend comparison vs previous period
  • commands/interaction-report.md — pure markdown command, no script dependencies (cross-platform: macOS, Linux, Windows)
  • Layer 3 respects layer3: true/false in .claude/ai-psychosis.local.md (opt-in, off by default)

Changed

  • README updated with Layer 3 usage instructions
  • Platform compatibility expanded: Layer 3 works on Windows
  • Version bumped to 0.3.0

0.2.0 — 2026-04-05

Added

  • Layer 2: Programmatic pattern detection — four hooks measuring session time, tool usage, burst patterns, and language flags
    • session-start.sh — daily session count, late-night detection
    • prompt-analyzer.sh — dependency, escalation, fatigue, and validation-seeking pattern flags (prompt text never stored)
    • tool-tracker.sh — event logging, edit ratio, burst detection, progressive alerts with cooldown
    • session-end.sh — session finalization, JSONL record, state cleanup
  • lib.sh — shared library with thresholds, state management, cooldown logic, and layer configuration
  • Per-project layer configuration via .claude/ai-psychosis.local.md
  • require_layer() guard in all hook scripts — layers are opt-in/out
  • MIT LICENSE file
  • matcher field in hooks.json for schema compliance

Changed

  • hooks.json now registers 4 events (was 2)
  • DATA_DIR fallback hardened to ~/.claude/data/ai-psychosis
  • README rewritten with architecture diagram, research background, privacy section, threshold reference tables
  • Version bumped to 0.2.0

Removed

  • periodic-reminder.sh — replaced by tool-tracker.sh
  • session-awareness.sh — replaced by session-start.sh

0.1.0 — 2026-04-04

Added

  • Layer 1: Behavioral instructions — SKILL.md with 5 rules and 5 named patterns (reinforcement loop, scope escalation, narrative crystallization, emotional dependency, session overuse)
  • periodic-reminder.sh — re-injects awareness every 25 tool calls
  • session-awareness.sh — SessionStart context injection
  • Plugin manifest (plugin.json)
  • Design document (docs/ai-ai-psychosis-brief_1.md)

Known gaps

  • No CI pipeline
  • Single-user plugin — no multi-user patterns considered