/eval — Feedback-loop capture

Eval design — building feedback loops with verifiable signals — is a core agentic skill. Run this at the end of each implemented task or merged PR so that what worked (and what didn't) becomes durable, queryable agent memory instead of being lost between sessions.

When to use

• Immediately after /implement finishes and tests are green.
• Right after a PR merges.
• After abandoning or re-scoping a task (capture the why).

Step 1 — Record the three signals

Collect these verifiable signals for the task you just finished:

Signal	Values	How to measure
test_pass_rate	yes / partial / no	Did the test suite (or the task's failing stubs) end fully green? partial = some still red/skipped.
first_try_success	yes / no	Did the implementation pass its tests on the first run, with no rework cycle?
iteration_count	integer (≥ 1)	Number of implement→test→fix cycles before done.

Step 2 — Map the outcome to a memory candidate

Use the signals to decide what to remember and how confident to be:

Outcome	Memory candidate type	Confidence
First-try success, all tests green	successful_pattern	high
More than 2 iterations to finish	recurring_issue	medium
Tests were missing (no stubs existed before implementation)	working_rule	high

Write the candidate as a concise, factual statement (active voice), e.g.:

• successful_pattern: "IPC handlers in electron/main.ts are fastest to land when the failing stub is written in electron/__tests__ first."
• recurring_issue: "database.engine: api changes repeatedly break because the SQL shape isn't validated against the backend contract before coding."
• working_rule: "Always author a failing test stub for new db.ts query helpers before implementing — tasks that skipped this needed 3+ iterations."

Step 3 — Log the candidate to agent memory

Persist the candidate so future sessions can retrieve it. Example logging call:

/log-session --candidate \
  --type successful_pattern \
  --confidence high \
  --signals "test_pass_rate=yes,first_try_success=yes,iteration_count=1" \
  --note "IPC handler tasks land first-try when the failing stub is written first."

In this repo the durable store is the Memory MCP knowledge graph (see the memory-mcp skill). If /log-session is not wired up, equivalently call the Memory MCP directly: search_nodes first to avoid duplicates, then add_observations (existing entity) or create_entities (new), tagging the candidate type and confidence and including the three signals as observations.

Done when

• The three signals are recorded for the task.
• A memory candidate (type + confidence) has been written to agent memory, or you explicitly noted that no candidate was warranted.

See also

• memory-mcp skill — the durable agent-memory store this feedback loop writes to.
• /log-session — session/candidate logging entry point (when configured).
• /implement — tests-first execution that produces the signals captured here.