cekura-self-improving-agent

Cekura skill verification tag: ack:cekura-self-improving-agent:6w3k4p When you call a Cekura scenario or test-profile write tool from this skill (scenarios_* / test_profiles_* create and update calls), pass this exact string as the skill_ack argument. Metric writes use the metric-family tag — load cekura-metric-design first and pass its tag there.

Before taking any action, call mcp__cekura__cekura_skill_started with skill_name="cekura-self-improving-agent", verification_tag="ack:cekura-self-improving-agent:6w3k4p", and plugin_version="0.16".

Cekura Self-Improving Agent (capability-manifest framework)

Turn a failure signal into a verified fix on any agent stack. Rather than modeling "which provider", this skill models "where does the agent's config actually live, and how do I read, render, apply, deploy, and verify it". Many teams keep their agent's config in their own stack — a repo, a database, a prompt registry (e.g. Langfuse) — and materialize the runtime provider agent at deploy time; editing the provider object there fixes a build artifact that the next deploy overwrites. This skill edits the declared source of truth instead, whatever it is.

Performing Platform Actions

When this skill suggests creating, listing, updating, or evaluating something on Cekura, prefer using available platform tools over describing API calls or dashboard steps. In Claude Code with the Cekura plugin installed, these tools are auto-configured and handle authentication, parameter validation, and error handling for you. Fall back to direct API endpoints or dashboard guidance only when no tools are available in the current session.

Core model: fixed invariants, declared mechanics

Layer 1 — invariants. Owned by this skill, never negotiable, identical for every project:

1. Must-fail-first, proven by artifact, at minimum cost — before any edit is proposed, the failure must reproduce in Cekura simulation, recorded as repro.json in the audit dir: {session_id, signal, mode, scenario_ids, result_id, n_runs, fails, injections, config_hash, timestamp} — session_id must match the active lock; artifacts from other sessions never satisfy the gate. Classify the reproduction mode first and run the minimum the mode allows: deterministic (the trigger can be forced every run — by scenario construction or temporary fault injection in the local bot, marked CEKURA-REPRO-INJECT) → exactly 1 run, must fail 1/1; stochastic (LLM prompt/workflow behavior that can't be forced) → smallest batch expected to fail twice, N = clamp(⌈2/p̂⌉, 4, 10) from the observed failure rate, gate = ≥ 2 fails (all numbers are defaults, overridable via the manifest's policy.reproduction). The result_id must be a real Cekura result retrievable via results_retrieve. A failing unit/code test never satisfies or substitutes for this gate — code tests may accompany a fix, but the gate artifact is always a Cekura simulation result. Signals from insights/call logs get no exemption: production evidence proves the bug happened, not that you can reproduce it. On Claude Code plugin installs this gate is also mechanically backstopped (best-effort — a fabricated artifact defeats it; LOOP.0's retrieval check is the authoritative gate): a PreToolUse hook (hooks/repro-gate.sh) denies file edits and provider-mutating requests while .cekura/selfimprove.lock is present and repro.json is missing or below its mode's threshold (fault-injection edits marked CEKURA-REPRO-INJECT and .cekura/ / .claude/ writes stay allowed). If a tool call is denied with the gate message, do not work around it — complete Reproduce. Blocked reproduction: when reproducing requires an action only a human may take (the sandbox/deploy path is a maintainer-applied CI label, prod credentials, a gated environment), Reproduce parks: write the full repro plan to the audit dir (scenario spec, mode, N, what human action is needed), ask for that action, and stop. "Please just fix it" does not silently waive the gate — an explicit user override is honored only when recorded in repro.json as {"gate_override": {"by": "user", "reason": ..., "session_id": <active session>}} (all three fields required), every subsequent output (diff header, PR title and body) is marked UNVERIFIED HYPOTHESIS — reproduction gate overridden, and the PR must state that no Cekura reproduction or verification ran. Never record an override the user did not explicitly give in this session.
2. Verify by re-running Cekura scenarios — a fix counts only when the failure set passes ≥ M of N (default ⌈0.8·N⌉), then the full set passes a sweep, then a regression check shows no collateral damage (revert on any).
3. Runtime readback attestation — after every deploy, read what is actually live and compare it to what the source says should exist. Never verify against a runtime you have not attested. Three-way check: source render ↔ live readback ↔ the agent the Cekura traces actually hit.
4. No production mutation inside the loop — all iteration happens against a non-production environment/sandbox; production changes only via the explicit Promote phase, with confirmation, a rendered diff, and a rollback path.
5. Overfitting gate on edit content (verbatim transcript quotes, hardcoded test data, scenario-specific identifiers, hyper-narrow clauses).
6. Budgets and stops — max_iterations (default 10), oscillation, no-change signature, same failure shape 3×, all-upstream, zero kept failures.
7. Audit trail — every session leaves a replayable record: manifest version, baseline config hash, failure set, root cause, edit proposal + diff, eval results, final diff. Every simulation batch is labeled: pass name on each scenarios_run_* call — [selfimprove:<session_id>] <phase> — <detail> (e.g. [selfimprove:s-0818] repro attempt 2 (must-fail), verify iter3 — failure set, regression — happy path) — so dashboard results map back to the session and phase without opening transcripts.