imprint teach deep-dive

Use this to answer "where did the ~N-minute teach go, turn by turn, and what's fixable."

Two data sources

1. Macro (phase/span timing + cost) — Phoenix. Requires IMPRINT_TRACE=1 + a running Phoenix collector (the teach must have been run with tracing on; scripts/teach-from-scratch.sh sets this). Project name imprint, root span cli.teach.
2. Micro (turn-by-turn per tool) — the per-tool compile log at ~/.imprint/<site>/<tool>/.compile-log.json (a Codex-cli stream-json event array: {type:'assistant'|'user'|..., message.content[], timestamp}).

Procedure

# 1. Macro: list recent teach traces, pick the site/run, get the span tree
bun run scripts/analyze-phoenix.ts --kind teach --last 8
#    Span tree under cli.teach: teach.detect_tool_candidates, teach.plan_prereqs
#    (build-plan), build_shared_module[_shared/*], plan_tool[*],
#    compile.generate[*] (the dominant phase), compile.playbook[*].
#    Cost is on llm.cost.* attrs; sum is the trace total.

# 2. Micro: turn-by-turn for one tool (or sweep the whole site)
bun run scripts/analyze-compile-log.ts ~/.imprint/<site>/<tool>/.compile-log.json --full
bun run scripts/analyze-compile-log.ts --site <site>          # all tools

analyze-compile-log.ts prints: tool-call breakdown, exploration %, done attempts (and how many were rejected), live-integration/pacing time, top time sinks, and the full per-turn timeline with wall durations (delta between consecutive tool_result timestamps).

What the numbers usually show (and the general levers)

• compile.generate of heavy RPC tools dominates (e.g. Google batchexecute search/booking ≈ 20–25m each). Most of that is per-tool field-index discovery on a schema-less positional array — the tool walks the decoded payload to find where its fields live. This is inherent, NOT a missing helper: the generic decode/round-trip already exists as a _shared/ module and is imported (verify with grep -n "_shared" ~/.imprint/<site>/<tool>/parser.ts). Do not "recommend adding a decode helper" — that double-counts the shared-modules phase.
• 25s live-request pacing on every integration/chain run, serialized inside each compile (anti-flag throttle to the real site). Large, but only partly controllable — a shared token-bucket across lanes can overlap the waits without raising the aggregate request rate.
• done rejections each cost a full verify cycle. Two recurring causes: (a) missing param:<name> coverage tests, (b) noUncheckedIndexedAccess TS errors that the agent's own run_tests never surfaces (it runs unit tests, not tsc). General fixes: run tsc inside run_tests; add a pre-done per-param lint.
• Concurrency is capped at 2. The critical path = sequential prefix (triage + build-plan + the shared-module build, gated by the slowest module) + the single slowest tool chain. More lanes can't beat that floor.
• A /tmp/dbg.ts scratch script can't import '../_shared/...' (relative path doesn't resolve from /tmp) → a wasted iteration. Scratch scripts belong in the tool dir.

Gotchas

• analyze-compile-log.ts handles both the current stream-json shape (type + message.content) and the legacy role/content shape. Tool names are MCP-prefixed (mcp__imprint-compile__run_bash → run_bash).
• Per-turn durations come from the user-entry timestamps; assistant entries have no timestamp, so a turn's duration is the time from the previous result to this one (think + tool exec).
• Phoenix compute sum > wall is normal — shared modules build in parallel and plan/playbook overlap adjacent compiles under concurrency-2.