self-improving-agent

name	self-improving-agent
description	Self-improvement logging and closed loop: triggered when commands/operations fail, user corrects, knowledge is outdated, external API fails, or a reusable better approach appears. Review historical learnings before important tasks. Avoid triggering during casual chat, temporary mistakes that don't need recording, or when user explicitly says not to record.
metadata	{"language":"en","scope":"minis"}

Self-Improvement Skill (Minis Edition)

This skill is used within the Minis environment to record errors, corrections, and reusable best practices, forming a traceable learning closed loop.

Minis Directory Convention

• Working directory: /var/minis/workspace/
• Skill default learning log directory: /var/minis/skills/self-improving-agent/data/
• Skill public learning log directory (after promotion): /var/minis/skills/self-improving-agent/data/public/
• Project-level learning log directory (optional): <project>/.learnings/
• Learning log files:
  • LEARNINGS.md (corrections, knowledge gaps, best practices)
  • ERRORS.md (command failures, exceptions)
  • FEATURE_REQUESTS.md (capabilities proposed by user)

By default, write to the skill's own data directory; when you explicitly specify a project, write to the project-level log; when the issue has been abstracted into a cross-project general rule, promote to the intra-skill public area or Minis memory system.

Current Final Rules

• Default recording location: /var/minis/skills/self-improving-agent/data/
• Skill public area: /var/minis/skills/self-improving-agent/data/public/
• Project-level recording: use <project>/.learnings/ only when --project <path> is explicitly passed
• Recommended public flag: --public
• Compatibility alias: --workspace still works but is no longer recommended
• Promotion behavior: promote <entry-ID> copies the entry to the skill public area, auto-marks the source entry as promoted, writes back **Promoted to** and ### Resolution
• Duplicate protection: if the entry already exists in the skill public area, repeated promote will not duplicate it

Quick Reference

Scenario	Action
Command/operation fails	Default record to skill dir data/ERRORS.md
User corrects you	Default record to skill dir data/LEARNINGS.md, category correction
User needs missing capability	Default record to skill dir data/FEATURE_REQUESTS.md
Explicit project context provided	Record to <project>/.learnings/
External API/tool fails	Record to current scope's ERRORS.md, include integration details
Knowledge outdated	Record to current scope's LEARNINGS.md, category knowledge_gap
Found a better approach	Record in current scope first, promote after confirming generality
Same issue recurs across multiple projects	Promote to skill public area data/public/
Similar to existing entry	Link with **See Also**, consider promotion priority
Widely applicable experience	Promote to skill public area or Minis memory (see "Promote to Minis Memory" below)

Trigger Recording Rules (Minis Runtime Convention)

Note: This skill does not auto-monitor in the background by default. When trigger conditions are met, the assistant (or you) proactively calls scripts/minis_auto_log.sh to write to disk.

Suggested "Must Record" Trigger Conditions

Any of the following should be recorded (unless you explicitly say "don't record"):

1. Command/operation fails non-trivially: e.g., permission, path, dependency, network, third-party API exceptions require investigation to locate.
2. User corrects: you point out where I misunderstood, logic doesn't match the software's actual behavior, path/spec is wrong.
3. Knowledge update/outdated correction: discovered previous assumptions don't fit Minis, or docs/implementation need correction.
4. Reusable better approach: forms a stable practice, convention, template, or workflow that significantly reduces rework.
5. Recurring pattern: same issue repeats within a task, or across tasks/projects.

Generally Not Recorded

• Casual chat, one-time small changes, trivial details without reuse value.
• You explicitly say "don't record".

Recording Location Suggestions

• Default write to skill area data/.
• After confirming cross-task reuse value, use promote to promote to skill public area data/public/.

Difference from Minis Memory and Promotion Standards

Difference (suggested understanding)

• This skill's logs (data/ and data/public/) are an editable work review repository: records context, troubleshooting process, solution evolution, allowing long text and details.
• Minis memory (memory_write writes to /var/minis/memory/) is cross-session long-term rules/preferences: should be short, stable, reusable; poorly written rules can "contaminate" future decisions long-term.

Write Choice (log first, then refine into memory)

• First write to this skill's log: when content needs context (error output, investigation path, comparison of solutions), unsure if generally applicable, or still iterating.
• Then promote to memory: when the conclusion is stable, cross-task/cross-skill applicable, and can be expressed in one sentence.

When to Promote to Memory (hard criteria)

Only consider memory_write when any of the following is met:

1. Can be condensed into a "next time X happens, do Y" rule, independent of specific project details.
2. Recurs >= 3 times in 30 days, or appears in at least 2 different tasks/domains.
3. Explicitly belongs to your long-term preference/convention (e.g., tool usage constraints, path conventions, output format rules), and you explicitly say "remember/do this going forward".

Promotion Action Suggestions

• First use promote to promote to skill public area data/public/ (higher visibility, easier to review).
• Then extract 1-3 short rules from the public entry and use memory_write to write to daily memory.

Usage Examples

# Default write to skill's own data directory
sh /var/minis/skills/self-improving-agent/scripts/minis_auto_log.sh init

# Record skill-level learning
sh /var/minis/skills/self-improving-agent/scripts/minis_auto_log.sh learning "Fixed download timeout" "Use chunking and retry"

# To explicitly write at project level, pass --project
sh /var/minis/skills/self-improving-agent/scripts/minis_auto_log.sh --project /var/minis/workspace/my-project error "curl request failed" "HTTP 429"

# To directly write to skill public area, pass --public
sh /var/minis/skills/self-improving-agent/scripts/minis_auto_log.sh --public feature "Support batch export" "Operations need daily report"

# Search skill area + project area + skill public area
sh /var/minis/skills/self-improving-agent/scripts/minis_auto_log.sh search timeout

# Promote entry to skill public area
sh /var/minis/skills/self-improving-agent/scripts/minis_auto_log.sh promote LRN-20260317-ABC

# View current scope
sh /var/minis/skills/self-improving-agent/scripts/minis_auto_log.sh status

Logging Format

Learning Record

Appended to .learnings/LEARNINGS.md:

## [LRN-YYYYMMDD-XXX] category

**Recorded At**: ISO-8601 timestamp
**Priority**: low | medium | high | critical
**Status**: pending
**Domain**: frontend | backend | infra | tests | docs | config

### Summary
One-line description of what was learned

### Details
Full context: what happened, what went wrong, correct approach

### Suggested Action
Concrete executable improvement or fix

### Metadata
- Source: conversation | error | user_feedback
- Related File: path/to/file.ext
- Tags: tag1, tag2
- Related Entries: LRN-20250110-001 (if related)
- Pattern Key: simplify.dead_code | harden.input_validation (optional, for recurrence pattern tracking)
- Recurrence Count: 1 (optional)
- First Seen: 2025-01-15 (optional)
- Last Seen: 2025-01-15 (optional)

---

Error Record

Appended to .learnings/ERRORS.md:

## [ERR-YYYYMMDD-XXX] skill_or_command_name

**Recorded At**: ISO-8601 timestamp
**Priority**: high
**Status**: pending
**Domain**: frontend | backend | infra | tests | docs | config

### Summary
Brief description of the failure

### Error

Actual error message or output

### Context
- Command/operation attempted
- Input or parameters
- Environment details (if relevant)

### Suggested Fix
If identifiable, give possible solution

### Metadata
- Reproducible: yes | no | unknown
- Related File: path/to/file.ext
- Related Entries: ERR-20250110-001 (if recurring)

---

Feature Request Record

Appended to .learnings/FEATURE_REQUESTS.md:

## [FEAT-YYYYMMDD-XXX] capability_name

**Recorded At**: ISO-8601 timestamp
**Priority**: medium
**Status**: pending
**Domain**: frontend | backend | infra | tests | docs | config

### Required Capability
What the user wants to achieve

### User Context
Why needed, what problem it solves

### Complexity Assessment
simple | medium | complex

### Suggested Implementation
Possible implementation approach and extension points

### Metadata
- Frequency: first_time | recurring
- Related Feature: existing_feature_name

---

ID Generation Rules

Format: TYPE-YYYYMMDD-XXX

• TYPE: LRN (learning), ERR (error), FEAT (feature)
• YYYYMMDD: current date
• XXX: sequence number or random 3 chars (e.g., 001, A7B)

Examples: LRN-20250115-001, ERR-20250115-A3F, FEAT-20250115-002

Entry Resolution

When a problem is fixed, update the entry:

1. Change **Status**: pending -> **Status**: resolved
2. Add resolution block after metadata:

### Resolution
- **Resolved At**: 2025-01-16T09:00:00Z
- **Commit/PR**: abc123 or #42
- **Note**: Brief description of what was done

Other statuses:

• in_progress - being worked on
• wont_fix - decided not to fix (write reason in resolution)
• promoted - promoted to Minis memory

Promote to Minis Memory

When a learning has broad applicability (not just a one-time fix), it should be promoted to the Minis memory system.

When to Promote

• Learning applies across multiple files/features
• Any contributor (human/AI) should know this
• Prevents repeated mistakes
• Records project conventions

Promotion Target (Minis)

• Daily Memory: /var/minis/memory/YYYY-MM-DD.md (written via memory_write)
• Global Memory: /var/minis/memory/GLOBAL.md (read-only, user maintains in settings)
• Project Notes: recommended to write to /var/minis/workspace/PROJECT_NOTES.md

How to Promote

1. Refine: condense learning into concise rules or facts
2. Write: use memory_write to write to daily memory, sync to project notes if necessary
3. Back-write: update the original entry:
   • **Status**: pending -> **Status**: promoted
   • Add **Promoted to**: YYYY-MM-DD.md or PROJECT_NOTES.md

Recurrence Pattern Detection

If the recorded content is similar to an existing entry:

1. Search first: grep -r "keyword" /var/minis/workspace/.learnings/
2. Establish relation: add **See Also**: ERR-20250110-001 in metadata
3. Increase priority: if the problem recurs
4. Consider systemic fix: recurrence usually means:
   • Missing documentation (-> write to PROJECT_NOTES.md or daily memory)
   • Missing automation (-> add to scripts or toolchain)
   • Architecture issue (-> create tech debt task)

Simplify & Harden Feed

Used to ingest recurring patterns from the simplify-and-harden skill and convert them into persistent prompt rules.

Ingestion Workflow

1. Read simplify_and_harden.learning_loop.candidates from task summary.
2. Use pattern_key as the stable dedup key for each candidate.
3. Search .learnings/LEARNINGS.md for existing entries:
   • grep -n "Pattern-Key: <pattern_key>" /var/minis/workspace/.learnings/LEARNINGS.md
4. If exists:
   • Increment Recurrence-Count
   • Update Last-Seen
   • Add See Also relation
5. If not exists:
   • Create new LRN-... entry
   • Set Source: simplify-and-harden
   • Set Pattern-Key, Recurrence-Count: 1 and First-Seen/Last-Seen

Promotion Rules (System Prompt Feedback)

Promote recurring patterns to Minis memory when:

• Recurrence-Count >= 3
• Appears in at least 2 different tasks
• Occurs within 30 days

Promoted rules should be short, clear preventive rules (actions to take before/during work), not lengthy postmortems.

Periodic Review

Review .learnings/ at natural checkpoints:

When to Review

• Before starting a new important task
• After completing a feature
• When entering a domain that has past learnings
• Weekly during active development

Quick Status Check

# Count pending items
grep -h "Pending" /var/minis/workspace/.learnings/*.md | wc -l

# List pending high-priority items
grep -B5 "Priority: high" /var/minis/workspace/.learnings/*.md | grep "^## \["

# Find learnings for a specific area
grep -l "Domain: backend" /var/minis/workspace/.learnings/*.md