Triumphs

Complete a full TDD red→green→refactor cycle using tdd-red, tdd-green, tdd-refactor agents

Pop the blueprint-mode super: complete a full 6-phase Blueprint Mode workflow with all gates honored

Use the workflow skill correctly in 5 consecutive sessions without scope creep

Use blueprint-mode correctly for 3 separate features — each invocation must pass all planning gates before implementation begins, with gate output documented in the PR description

Achieve 10 consecutive GitHub Actions runs that pass on the first attempt — no re-runs, no workflow_dispatch retries, no force-push resets across any repo

Observe a workflow friction point during a session and file a GitHub issue or SKILL.md update to address it before the session ends — the fix is documented, not deferred

Fix a failing GitHub Actions workflow — diagnose the failure, apply the fix, verify the run passes — using gha-expert or bluefin-ci skill

Complete the cicd-learning canonical cycle end-to-end: SEARCH → LOAD → ACT → CAPTURE → STORE → SYNC — all steps verified in a single session

Push a PR that passes all CI checks on the first attempt, with no re-runs, no force pushes, no 'fix typo' commits

Complete a session where observed friction results in a new rule added to cicd-learning/SKILL.md or workflow/SKILL.md

Diagnose and resolve a CI failure that has been open for more than 48 hours, referencing the gha-expert or bluefin-ci skill explicitly in the fix

Conduct a formal SLO review at the end of a season — compare actual performance to SLO targets, document trends, and produce at least one concrete action based on the data.

Instrument a service with structured logging (JSON or equivalent) — every log entry must include: timestamp, level, correlation ID, service name, and a message field. No freeform string concatenation

Implement request correlation IDs that propagate across at least 2 service boundaries — a single trace ID visible in logs from ingress to final handler, queryable end-to-end

For every alert firing in production: ensure a runbook link is present in the alert body. Deploy this requirement for at minimum 5 distinct alerts across one or more projects

Identify and remediate a high-cardinality metric label (e.g., user ID, request ID, or unbounded enum in a label) that is causing metric storage explosion — document before/after cardinality count

Establish a documented performance baseline for a real service — p50, p95, p99 latency and error rate captured under normal load, committed to the repo, and referenced in future runbooks

Build an on-call dashboard that answers the first 5 questions an engineer asks during an incident — it must be usable without familiarity with the system, and linked from the on-call rotation docs

For a single service in production: demonstrate all four Golden Signals (latency, traffic, errors, saturation) are queryable, graphed on a dashboard, and each has at least one associated alert threshold defined

Execute `knowledge-search` at the START of a session (before writing any plan or touching any file) in 10 consecutive non-trivial sessions. A non-trivial session is one where at least one file is modified.

In 7 consecutive code-change sessions, the TDD red phase (failing tests written FIRST) precedes any implementation. A session where implementation precedes tests resets the counter to zero.

Complete ALL 9 items in the Definition of Done checklist for 10 consecutive sessions. Every item — knowledge-search at start, scope declared, tests written before implementation, tests run and quoted, output verified, scope respected, learnings captured, knowledge-store called at end, sync dispatched.

Every PR has at least one substantive review comment before merging — 5 consecutive PRs. Not auto-merge, not 'LGTM' without reasoning. Substantive means at minimum 2 sentences of technical feedback.

Zero agent re-dispatches in 7 consecutive sessions — every agent call resolves the task on the first invocation. A re-dispatch is defined as calling the same agent type twice for the same task because the first result was wrong or insufficient.

Zero pending todos at session end for 10 consecutive sessions — every task started is either completed or explicitly deferred with a reason. A session that ends with an unaddressed pending todo resets the counter.

First tool called resolves the task (no fallback) — 5 consecutive tasks. The correct tool is identified and deployed on the first attempt; no secondary tool call is needed to complete what the first was meant to do.

AUTOLOAD completed fully every session — 20 consecutive sessions. Every session begins with the full AUTOLOAD sequence: knowledge-search, load relevant skills, declare scope. No partial completions. No skipped steps.

Touch at least 3 different skill families (bluefin, cncf, tdd, security, etc.) in a single season — each family must produce a real deliverable (PR, knowledge entry, or skill update) in that season.

Complete the knowledge loop (knowledge-search → act → knowledge-store) in 10 or more distinct sessions across the season. Each session must be a different project or domain.

Invoke `blueprint-mode` for 3 or more different project types across the season — each invocation for a genuinely different domain (not the same type of feature in the same repo three times).

Complete a cross-repo session that meaningfully touches 3 or more distinct repositories — coordinated changes, not just read-only exploration. All repos must receive a real deliverable.

Load a skill that has never been loaded before in your session history — 3 separate occurrences across the season. Each must be a genuinely new skill (not one you've previously loaded and forgotten about).

Complete a session that combines bluefin + cncf + tdd workflows in a single session — each must produce a real output (PR, test, knowledge entry). The domains must be genuinely integrated, not sequential and isolated.

CI first-attempt green for 10 consecutive sessions. Every PR in those sessions must pass all CI checks on the initial push — no re-runs, no force-pushes, no fixup commits.

Zero force-pushes for an entire season. Every push to every branch is additive. If git history needs rewriting, it is done before the first push — never after.

Zero regressions introduced across a full season — confirmed by the `qa` agent finding no bugs introduced by your changes in any session. Each session where `qa` runs must produce a clean bill of health.

`just sync` run in 15 consecutive sessions where any skill or knowledge file was updated. The sync must occur within the session window — not batched afterward.

`pl` command run (AUTOLOAD honored) in 20 consecutive sessions. AUTOLOAD must be completed fully — all four steps — in every qualifying session.

Zero failed agent dispatches across 10 consecutive sessions. A failed dispatch is one where the agent returns an error, produces no output, or is re-dispatched because the first result was unusable.

Blueprint-mode gate honored every time across 5 sessions — no implementation started before the blueprint gate passes. The gate output must be documented in each session.

Full cicd-learning cycle completed 3 consecutive times — SEARCH → LOAD → ACT → CAPTURE → STORE → SYNC — each cycle in a distinct session with a real deliverable. Three consecutive completions without a break.

Complete an entire season of work (12+ consecutive sessions) with zero CI regressions introduced — every commit leaves the pipeline green

Instrument a real service or pipeline with all four Golden Signals: latency, traffic, error rate, saturation. Each signal must produce a queryable metric or log line — not just a print statement.

Instrument a multi-step operation with distributed tracing (OpenTelemetry or equivalent) — at minimum 3 spans covering distinct stages. Use the trace to identify an actual performance bottleneck or unexpected latency.

Define and deploy a monitoring alert that fires for a real condition and NOT for normal variance. The alert must have documented: threshold rationale, runbook link, severity level, and a suppression condition for known maintenance windows.

Create an operational dashboard covering at least 3 distinct signal types for a real project. The dashboard must be usable by someone who didn't build it — include a title and description for each panel.

Define formal SLOs for a real service: at minimum one availability SLO and one latency SLO. Each must include: measurement method, target percentage, measurement window, and the consequence of breach. Store the definition in the knowledge base.

Make and document an explicit error budget decision: either freeze releases because the budget is exhausted, OR document a risk acceptance to ship despite budget pressure. The decision must reference the SLO and be stored with a `type:slo-decision` tag.

Identify a system, pipeline, or workflow that currently fails invisibly. Add at minimum: a failure log with context, an alert or check, and a runbook link. Verified by simulating the failure and confirming the signal fires.