perf(agents): hand the agent amico-run's absolute path — 28x faster on the task that needed it

[Rchari1]

The finding

AGENTS.md said bare amico-run and relied on PATH. When PATH missed, the agent went looking for the binary — and looking is expensive. Profiling one session's history:

• 128 unbounded filesystem searches, 417 seconds total, individual calls at 123s
• e.g. which amico-run || find /Users/raghavchari -type f -name "amico-run", and find ~/.julia -name "Piccolo" -type d at 126s

The fix

A new {{AMICO_RUN}} substitution (alongside the existing {{JULIA_PROJECT}} / {{TEMPLATE_PATH}}) resolves to the launcher's absolute path at session prep, and the guidance forbids which plus the find/pip list fallbacks.

Degrades safely: with no resolvable bin dir it substitutes the bare command — exactly today's behaviour.

Measured, A/B

Same model (amazon-bedrock/moonshotai.kimi-k2.5), same prompt, only AGENTS.md differing. Task: "Run amico-run to print its usage/help output." 4 baseline vs 7 treatment trials.

	turns	tool calls	wall (median)	which calls	timeouts
baseline	6	10	226.6s	4	2 of 4
treatment	2	1	8.0s	0	0 of 7

~28× on median wall, 3× fewer turns, 10× fewer tool calls — and the baseline failed to finish at all in half its trials (363s+, one of them 11 consecutive single-tool turns).

The mechanism is visible in the transcripts. Baseline:

amico-run --help
which amico-run || which amico || echo "Not in PATH"
ls -la /usr/local/bin/amico* …
ls -la ~/.amico/bin/amico* …
find /usr -name "amico*" -type f …
find /opt -name "amico*" -type f …
pip list | grep -i amico …

Treatment: the absolute path, once.

What did NOT work — and is trimmed accordingly

An earlier draft of this branch carried a long batching instruction quoting the profiling numbers. I A/B'd it on a six-independent-reads task and it showed no measurable effect, because both models already batch:

• kimi-k2.5: per_turn = [6, 0] in all six trials, both arms — already optimal
• deepseek-v4-flash-free: turns 3,3,2 baseline vs 4,2,3 treatment — noise

Since it ships on every turn, I cut it to one sentence: +160 tok/turn instead of +313. The intent stays; the prose that bought nothing measurable doesn't.

This also corrects an over-claim I made while profiling. The 212 "collapsible" read pairs are ~5% of turns, not a 2–3× lever, and the models mostly batch already.

Honest caveat on effect size

Under the trimmed wording, one of three re-runs regressed to 5 turns (from a consistent 2). The substitution is what does the work; the prose is a backstop. which calls stayed at 0 across all 7 treatment trials under both wordings, so the mechanism is robust even where turn count varies.

Tests

3 new: the substitution, the safe degradation to a bare command, and the two instructions. A pre-existing guard — "leaves no unknown {{...}} placeholder after session-prep substitution" — caught the new placeholder immediately; it now covers it.

853 passed, 6 skipped, tsc --noEmit clean, built.

Not addressed here

Context is still ~43k tokens per call (base AGENTS.md ~8.1k, a 37-skill index ~1.8k injected every turn regardless of need, plus 25 tool schemas and history). Trimming the always-on surface is the next lever — #244's two-tier skill work already points at it — but it wants its own measurement rather than being bundled in here.