dongshuyan/compass-skills / skills/user-profile-keeper
user-profile-keeper
VerifiedLocal user-profile maintenance skill for Codex, Claude Code, OpenClaw, OpenCode, and other agent harnesses. Use only when the user explicitly invokes $user-profile-keeper to create, initialize, update, query, correct, delete, export, or audit a local persistent user profile. It can extract durable collaboration preferences, requirement-expression habits, capability boundaries, recurring omissions, risk preferences, privacy boundaries, and typical events from the current session into auditable, confirmable, retractable local profile data. Do not auto-invoke, upload profile data, or replace task-clarifier's normal clarification flow.
| File | Kind | Size |
|---|---|---|
| SKILL.md | skill md | 9.4 KB |
| agents/openai.yaml | config | 285 B |
| references/examples.md | markdown | 4.7 KB |
| references/privacy-boundary.md | markdown | 3.4 KB |
| references/profile-schema.md | markdown | 4.6 KB |
| references/questionnaire.md | markdown | 3.8 KB |
| references/task-clarifier-integration.md | markdown | 2.4 KB |
| references/update-policy.md | markdown | 4.9 KB |
| scripts/onboarding_webui.pyexecutable | script | 24.4 KB |
| scripts/profile_store.pyexecutable | script | 42.0 KB |
| scripts/smoke_test_onboarding.py | script | 8.2 KB |
references/update-policy.md
View on GitHub# Update Policy
Profile updates must preserve evidence, confidence, sensitivity, and reversibility. A single session does not create a permanent character judgment.
## Evidence Types
- `durable_profile_evidence`: explicit long-term or default collaboration preferences, risk boundaries, privacy boundaries, capability boundaries, recurring omissions, or anti-bubble rules stated by the user.
- `operational_instruction`: current task instructions, repository rules, AGENTS rules, skill operating steps, privacy rules, and tool-use constraints. These guide the current task and do not prove a durable profile by themselves.
- `agent_observation`: the agent's observation of the current session. Use it only as a proposal unless it is repeated, direct, low-ambiguity, and limited to collaboration process.
## Auto Apply
The script may write a candidate as `active` only when all conditions hold:
- The user explicitly self-reported a long-term/default collaboration preference, or the session contains repeated, direct, low-ambiguity observation.
- `sensitivity` is `low`.
- `confidence >= 0.75`.
- The candidate does not conflict with an existing active assertion.
- The assertion affects communication, need alignment, workflow, evidence preference, privacy boundary, or risk confirmation.
- The evidence is not AGENTS rules, system/developer instructions, skill operating steps, repository constraints, or current task constraints.
User-provided background information, including age range, education, field, role, experience stage, and long-term goals, starts as a pending proposal with `sensitivity=private`.
`--auto-apply-safe` asks the script to consider active write. It is a safety-gated path. Candidates that fail the checks become proposals when proposal mode is enabled.
## First-Run Gate
Use one first-run gate:
- If an active profile exists, handle the request as an incremental update.
- If no active profile exists, recommend the onboarding questionnaire.
- If the user asks for the questionnaire or WebUI, run it.
- If the user declines the questionnaire, continue with the current task and create proposals for durable candidates.
Do not skip the questionnaire by counting how many questionnaire modules the current session appears to cover. Do not initialize active profile data from operational instructions.
Current task constraints can govern the current task. If they may have durable value, create a pending proposal and let the user confirm.
## Session Inference
Inference candidates must satisfy:
- `source_type=inferred`.
- Default `sensitivity=private`; identity, health, finance, legal, political, religious, family, intimate, or location claims are at least `sensitive` and should usually be avoided.
- Default status is pending proposal.
- `value` contains `summary`, `basis`, `reasoning`, `counter_evidence`, `usefulness`, and `review_question`.
- Evidence comes from concrete current-session behavior or explicit text, not model stereotypes, path guessing, timezone guessing, or language guessing.
Confidence caps:
- Single weak signal: `confidence <= 0.4`.
- Multiple consistent session signals with alternative explanations: `confidence <= 0.6`.
- User-confirmed inference can be rewritten as `self_report` or `correction` with confidence above `0.75`.
Never silently write these claims as active:
- real location, ethnicity, political/religious position, health, finance, legal risk, or intimate relationship;
- strong labels about ability, personality, values, psychological state, or morality;
- high-impact judgments that could change how the system treats or evaluates the user.
## Pending Proposal
Use pending proposal for:
- `source_type=inferred`;
- `sensitivity` of `private`, `sensitive`, `intimate`, or `secret`;
- self-reported background information such as age range, education, field, role, experience stage, and long-term goals;
- conflict with an existing active assertion;
- claims that could affect how the user is treated or judged;
- uncertain user phrasing;
- sources that are mainly current task constraints, skill operating rules, or one-off context, when the agent sees possible durable value.
## Conflict Handling
If a new candidate differs from an active assertion with the same `category + claim + scope`:
1. Keep the existing active assertion.
2. Create a pending proposal.
3. Mark the conflicting assertion.
4. Supersede only after user confirmation.
## Correction
User correction has highest priority:
- Set the old assertion to `retracted` or `superseded`.
- Save the new assertion with `source_type=correction`.
- Record the correction in the audit log.
## Anti-Staleness
Preferences, boundaries, work style, and risk tolerance can change. When reading the profile, use this order:
1. Current session explicit statements.
2. Recent user corrections.
3. High-confidence active assertions.
4. Older or lower-confidence observations.
Current explicit requirements override stored profile data.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| rea-analysis-optimizationmorluto/reaOptimize REA providers and shared analysis workflows for agent context, latency, CPU, and memory while preserving correctness and Evidence. Use for analysis bottlenecks and cache or query design; CI/test-suite optimization belongs to rea-ci-tests-audit. | Stars: 472k | Updated | Claude CodeCodex | Verified | Install |
| rea-ci-tests-auditmorluto/reaAudit REA CI and test suites for correctness, fixture drift, false passes, and measured performance; use for suite or workflow investigations and efficiency changes, rather than ordinary feature testing or a single test failure. | Stars: 472k | Updated | Claude CodeCodex | Verified | Install |
| rea-tool-designmorluto/reaAudit, investigate, design, or change REA investigation tools, CLI/MCP contracts, provider capabilities, and Evidence semantics. | Stars: 472k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-auditDietrichGebert/ponytailQuality audit of the whole repo: bugs, security, real load, missing tests, speed, and what to delete. Most important first. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |