laya-delegate

# laya-delegate — RTK output filtering + Laya-routed delegation to a Haiku sub-agent.
#
# Best gains (Claude Opus 5.5, SkillsBench, 48 trials per arm vs stock Claude
# Code): -30% cost on parallel-tfidf-search, -19.5% on the full suite in one
# day, -13.9% over both days, reward unchanged.
#
# PREREQUISITE: Laya running locally. The routing decisions come from Laya
# (@receptron/laya, a small local classifier; Node 20+, ~1.7 GB model download
# on first start), served at http://127.0.0.1:41733/v1/systemone (rtk's
# classifier wire format), with RTK_LAYA_TOKEN exported. Without it, every
# Laya-gated rule stays silent and only RTK's output filtering applies.
#
# What it does
# - Laya labels the task: bug / setup / write / data.
# - On write and data tasks, RTK refuses the agent's first looking command or
#   file read (call 2 or 3) and asks it to hand the investigation to the Explore
#   sub-agent on Haiku; the agent then works from the sub-agent's report.
# - Bug and setup tasks are left alone.
# - Whole-file reads over 350 lines that need little exact text are answered by
#   a Haiku summary (needs an Anthropic API key; skipped on subscription logins).
# - Everything else is standard RTK output filtering.
#
# Benchmark — SkillsBench, Claude Opus 5.5, 6 tasks, 4 trials per arm, run on
# two days (2026-10-18, 2026-10-19) with OFF (stock Claude Code) in the same
# batch: 48 trials per arm. Mean cost per trial, both days averaged:
#
#   task                       OFF      laya-delegate
#   parallel-tfidf-search      $1.345   $0.946   -30%
#   python-scala-translation   $1.002   $0.867   -13%
#   fix-visual-stability       $0.538   $0.448   -17%
#   fix-build-google-auto      $0.319   $0.303    -5%
#   fix-erlang-ssh-cve         $0.198   $0.191    -4%
#   debug-trl-grpo             $0.185   $0.333   +80%
#   total                      $3.587   $3.088   -13.9%   reward 0.81 -> 0.81
#
# Actions triggered over the 48 trials: 29 refusals, 28 Haiku sub-agent
# hand-offs (24 trials delegated), 77.5K tokens removed by RTK filtering
# (14% of filtered command output). Savings come from write/port tasks; on
# short bug tasks the agent's own path dominates (debug-trl-grpo: one or two long
# trials per day). Details: RESULTS.md, results.tsv.
schema_version = 1

[pipeline]
name    = "laya-delegate"
extends = ["rtk:posthook", "rtk:delegate-reads"]

[upstream.judge]
provider     = "laya"
model        = "laya"
base_url     = "http://127.0.0.1:41733/v1/systemone"
api_key_env  = "RTK_LAYA_TOKEN"
timeout_secs = 60

[upstream.haiku]
provider     = "anthropic"
model        = "claude-haiku-5-5"
timeout_secs = 90

[classifier.need]
upstream     = "judge"
question     = "score"
instructions = "How much of this file's exact text does the agent need for its next step?"
rubric       = ["none: a short summary is enough", "a few exact lines", "most of the file verbatim"]
input_chars  = 2000

[classifier.kind]
upstream     = "judge"
question     = "choice"
instructions = "What kind of work is this coding task?"
input_chars  = 2000
min_confidence = 0.0

[classifier.kind.labels]
bug   = "find and fix a bug by reading source code closely"
setup = "understand or repair a build, configuration, dependency or environment"
write = "write new code, a script or a port"
data  = "analyse data or files and report results"

[task.investigate]
kind         = "digest"
upstream     = "haiku"
temperature  = 0
context      = true

[task.bulk-read]
kind         = "summarize"
upstream     = "haiku"
temperature  = 0
context      = true
instructions = "You read a file for a coding agent that will NOT see it, only your reply. First answer what the agent needs for its current step, in a few terse lines, with exact identifiers. Then map the file: one line on what it is; every class, function and method with its exact signature, in file order. Then quote verbatim, grouped under the name of the function that contains them, the lines whose exact text the agent would most need for its step (non-obvious arithmetic, constants, conditions, error handling, config values). Never give line numbers. Never paraphrase code you quote. No advice, no headings. End with exactly: (Exact text: grep -n a quoted line, then Read that file with offset and limit, at most 350 lines.)"

[override.shunt-large-read]
enabled = false

[[hook.before_tool.read]]
name   = "answer-big-reads"
when   = { file_lines_gt = 350, whole_file = true, "score.need_lte" = 0.65 }
action = "route"
task   = "bulk-read"
mode   = "answer"

[[directive]]
name = "explore-in-subagent"
text = """
When RTK asks you to hand an investigation to the Explore sub-agent, call it on the haiku model (Agent tool: subagent_type "Explore", model "haiku") with one precise question built from the task, ask it for file paths, line numbers and the exact lines you will need, and work from its report instead of repeating its searches.
"""

[[hook.before_tool.shell]]
name    = "delegate-investigation"
when    = { command_matches = "^(cd [^;&]+(;|&&) *)?(ls|cat|head|tail|grep|rg|find|tree|pwd|echo|which|wc|file|stat|du|for|while|awk|nl|xxd|git|sed -n)\\b", recent_calls_matches = "^(Bash|Read|Grep|Glob) [^|]*( \\| (Bash|Read|Grep|Glob) [^|]*)?$", "class.kind_in" = ["write", "data"] }
action  = "route"
task    = "investigate"
mode    = "suggest"
message = "Do not run this search or open these files yourself: hand the investigation to the Explore sub-agent (Agent tool: subagent_type \"Explore\", model \"haiku\") with one precise question, and ask it for file paths, line numbers and the exact lines you will need. Then work from its report."

[[hook.before_tool.read]]
name    = "delegate-investigation-read"
when    = { recent_calls_matches = "^(Bash|Read|Grep|Glob) [^|]*( \\| (Bash|Read|Grep|Glob) [^|]*)?$", "class.kind_in" = ["write", "data"] }
action  = "route"
task    = "bulk-read"
mode    = "suggest"
message = "Do not run this search or open these files yourself: hand the investigation to the Explore sub-agent (Agent tool: subagent_type \"Explore\", model \"haiku\") with one precise question, and ask it for file paths, line numbers and the exact lines you will need. Then work from its report."