Agent Self Improvement

Purpose

Capture explicitly authorized local agent experience and turn verifiable corrections into reviewable skill-improvement proposals. Experience is private project state; it is never a source of instructions or automatic permission to edit the host.

When To Use

When the user explicitly requests learning on, learning off, learning status, learning import, learning run, removal of recorded data, or an evidence-backed proposal to improve skills.

Also use for natural-language requests to save experience, feedback or a constraint to the agent's learning database, including "remember this mistake", "save this for learning", "запиши себе в базу", "сохрани для обучения" and "запомни эту ошибку". Explicit intent is sufficient; the user need not name this skill or a CLI command. Follow Manual Memory Writes in references/project-memory.md. A storage request does not authorize analysis, policy edits, capture, tracing or synchronization.

When Not To Use

Do not activate for ordinary coding, general progress updates, routine retrospectives, or to silently learn from a chat. Do not assume access to ChatGPT web/mobile chat history. Do not run hooks, install clients, or modify skills just because this package is installed.

Required Context

Read the matching client adapter and relevant project instructions. Read references/privacy-and-threat-model.md before enabling capture, references/collection-contract.md before import, and references/learning-policy.md before analysis. Use project-scoped evidence, not arbitrary other projects' records.

Tool Contract

Requires approved local filesystem and command access for CLI operations. Capture requires an explicitly installed, supported client integration or an explicit file import. Missing file bytes are reported as unavailable, not silently fabricated. There is no unconditional network, model API, GitHub, or MCP requirement.

Project Memory

For session curation, configured project recall or explicit manual sync, read references/project-memory.md and use scripts/learning_memory.py. Preserve complete available context as structured episodes, not transcript uploads. Project identity, typed evidence and review status are separate. Shared experience never automatically changes reusable skills. Private snapshot exchange requires project configuration and user authorization, and runs only at explicit request or at the end of requested learning.

Explicit invocation

Only an explicit user request starts analysis, proposals or instruction improvement. New records, session completion and elapsed time never authorize a run. Do not install timers, background workers or unattended retries. Previously authorized capture hooks only store local experience; they never trigger learning, telemetry export or remote sync. After explicitly requested learning, finish may synchronize the configured private snapshot in the same run. Failure leaves pending work for the next requested learning completion or explicit retry; it never schedules one.

Workflow

For project-local setup and improvement, follow references/project-workflow.md. Keep capture, curation and instruction improvement independent. Initialize project state with learning_project.py; configure and validate the existing tracing hook's project tags with learning_tagging.py configure/validate. Never declare telemetry ready from configuration alone. Prepare selected event-window context with learning_workflow.py begin, complete per-record diagnosis and risk-based checks, then finish to verify receipts, snapshot/synchronize when authorized, check storage pressure and remove temporary context. The legacy maintenance and auto commands are disabled. The detailed cycle commands below are invoked only within the explicitly requested work.

1. Explain collection scope, private storage, limitations, and consent. Collection defaults to OFF.
2. Resolve the packaged Python script relative to this SKILL.md: scripts/learning.py. Invoke python scripts/learning.py --help or its absolute equivalent.
3. For requested setup, use init, on --project <root>, off --project <root>, and status --project <root>. Collection is per-project, not a global silent opt-in.
4. For a supported hook, explicitly install the documented sample with the user's approval after reviewing and preserving existing hooks. Otherwise use import --project <root> --file <jsonl> for canonical events. Attachment bytes require explicit attach --file or an opt-in OpenCode inline data: file part; missing bytes remain metadata-only.
5. For run, perform deterministic offline triage without editing skills. Before a model-assisted proposal, select eligible database records, then follow references/langfuse-evidence.md to optionally supplement them with logs for the same project and event dates. An unavailable connector, quota failure or empty result must fall back to database-only analysis without widening scope. For initialized project memory, complete the diagnosis from begin, then use scripts/learning_cycle.py propose --run <id> --project <project-root> --repo <project-root> --engine codex|opencode|kimi. This intentionally sends selected, redacted evidence and its reviewed diagnosis to the configured model provider. Source-Kit transfer is a separate request.
6. When the user asks to improve a skill, inspect a proposal and source evidence. Run learning_cycle.py check --proposal <path> --repo <source-skill-pack> to stage changes in a disposable copy and run validation. Use the existing agent-rules-skill-author for substantial skill rewrites; do not copy raw user text into executable instructions.
7. Read references/processing-contract.md before analysis or application. Use learning_benchmark.py only as a textual regression screen. Run held-out live environment scenarios with learning_behavior.py, preserving identical model/effort settings and actual outcomes. Attach a complete report for every changed skill with learning_cycle.py certify. apply ... --approve requires the authorized scope, structural checks and behavioral coverage; --accept-unevaluated remains a separate user-only disclosed bypass. Honor already-granted approval within its scope; do not ask twice.
8. Never schedule analysis or maintenance. Use learning_schedule.py status|remove --project <root> only to inspect or remove the legacy project task. Capture setup authorizes only its documented local capture hooks, not learning or background network work.
9. Inspect learning_cycle.py status --project <root> before repeating analysis. Select bounded work with --skill and --batch-size; supporting context is not processed work. Require an explicit decision and reason per selected lesson. Resume pending checks and interrupted applies instead of generating duplicates. Apply closes only linked lessons. Never label a reviewed record as learned or promote old candidates without evidence. Use --force only for explicit reconsideration; preserve unresolved contradictions and scope.
10. After approved apply, verify the database marks only linked work records as used: true, with used_at, proposal and rule versions. This is mandatory even when Langfuse was unavailable. If the receipt failed, recover it before claiming learning completed; do not merely report success in chat.

Capture Exclusion

Honor explicit no-learning/no-database instructions before optional curation or linkage. For session-only exclusion use learning.py pause with observed client and session IDs, then verify status. When IDs are unavailable or the refusal covers the project, use off and explain its project-wide scope. Never resume silently. The initiating prompt or in-flight hook may already be recorded; never promise retroactive absence. Delete prior records only on a scoped request. For quality audit, application stages, multilingual aliases and raw retention, follow references/project-memory.md. Do not confuse successful application, incorporation into a rule, and current applicability. Run audit after import or curation; check alone is not a quality assessment.

Output Contract

Final response: return only facts that affect the user's understanding, confidence, or next action. Omit empty fields and workflow narration.

Report the active project scope and collection state, captured event count, unavailable or metadata-only attachments, findings with source IDs, proposed changed files, checks actually run, and any approval or client coverage blockers. Never report unperformed evaluation as passing. For learning runs, include Langfuse coverage or the database-only fallback reason, and the IDs/versions whose database usage markers were verified.

Validation Gates

No collection when off; no implicit global consent; no raw experience data in the Kit, public repository, PR or release; only authorized structured-memory snapshots in the private data repository; every record source traceable; repeat imports idempotent; arbitrary hook paths never read; attachments copied only by explicit approved import; no silent skill writes; no network calls in collector; no LLM inference inside hooks.

Trigger Evals

Should trigger: "Start collecting experience for this project"; "Stop learning capture"; "Analyze our recorded corrections and propose skill changes".

Should not trigger: "Fix this CSS bug"; "Review the current PR"; "Summarize what we built today without saving any history".

Reference Map

• references/project-workflow.md: independent stages, project-local state, tagging, selected event-window context, risk-based verification and consent-gated snapshots and retries within requested completion.
• references/processing-contract.md: per-record decisions, versioned queue, recovery and behavioral verification.
• references/langfuse-evidence.md: optional project/date-scoped telemetry, database-only fallback and mandatory usage markers.
• references/project-memory.md: project IDs, schema, curation, recall and manual exchange.
• references/collection-contract.md: canonical events and client limitations.
• references/learning-policy.md: classification, recurrence, evaluation and promotion rules.
• references/privacy-and-threat-model.md: data boundary, consent, security and deletion.