config.example.toml

# Copy this file to config.toml (gitignored) and fill in real values.
# scripts/model_fix_loop.py and scripts/parallel_model_fix_loop.py both read
# config.toml by default (override the path with --config).
#
# Each [[worker.models]]/[[reviewer.models]] entry is one pool member. One
# is picked at random for every individual model call (each fixer attempt,
# each reviewer verdict), so a run rotates across the whole pool instead of
# pinning to one model. A single-entry pool behaves exactly like the old
# fixed-model setup.
#
# An entry needs only "name" -- it inherits base_url/api_key from its
# [worker]/[reviewer] table. Add base_url/api_key directly on an entry to
# pull that one model from a different provider (see the Fireworks example
# below), so a single pool can mix providers.

[worker]
base_url = "https://pass.wafer.ai/v1"
api_key = "your-api-key-here"
max_tokens = 4096
# max_prompt_tokens: section 6 raised the default 4096 -> 8192 (not the
# critiqued 12288 -- TPM economics), paired with graduated per-section
# truncation (assemble_prompt_sections) so the extra room reaches the
# learning block instead of just more parser-file text.
max_prompt_tokens = 8192
reasoning_effort = "low"
# stream = true is the default; the reply (and, when the provider sends a
# final usage chunk via stream_options.include_usage, its cache/token
# accounting) is reassembled exactly as a non-streamed call would return.
stream = true
# prompt_cache: "auto" (default) sends plain messages and relies on the
# provider's automatic prefix caching -- build_prompt orders sections
# static-first so that prefix is maximal. "explicit" additionally wraps the
# static prefix in an Anthropic-style cache_control breakpoint (only helps
# providers that accept it; opt-in since it changes the request shape).
# "off" disables both. Either way, cached-token counts the provider reports
# are logged to manifest.log as cached_tokens=<hit>/<total>.
prompt_cache = "auto"
thinking = false
temperature = 0.7
max_prompt_tags = 40
max_prompt_file_bytes = 60000
timeout = 600
# deadline_seconds (default 120): wall-clock ceiling on ONE model call,
# checked as stream chunks arrive and NOT reset by arriving data. This is
# the knob that actually bounds a call -- `timeout` above is urlopen's
# SOCKET read timeout, which every arriving SSE chunk resets, so under
# stream = true it cannot bound a slow-drip response at all (measured on
# one provider: a single request ran 2118s against timeout=1200). On
# expiry the connection is closed and the request is replayed, counting
# against max_retries.
deadline_seconds = 120
max_request_turns = 20
max_request_repeats = 3
max_verify_turns = 10
compaction_trigger_tokens = 150000
compaction_keep_recent_turns = 10
# compaction_min_elide_tokens: once compaction triggers, only served user
# payloads whose own estimated size exceeds this floor are stubbed (default
# 3000); smaller turns are left intact.
compaction_min_elide_tokens = 3000
governor_calls_per_minute = 30
governor_burst = 5
governor_cooldown_seconds = 30
governor_max_cooldown_seconds = 300
max_cluster_tags = 6
use_sccache = true
# claim_stale_seconds: a tag claim in the shared tag-state older than this
# is treated as abandoned (owner dead) and re-claimable. Generous because a
# live owner's heartbeat thread re-stamps its claim every heartbeat_seconds
# while an attempt is in flight -- stale means dead, not slow.
claim_stale_seconds = 7200
heartbeat_seconds = 60
max_retries = 1000
retry_backoff_seconds = 2
max_retry_backoff_seconds = 120
# build_semaphore: spec section 5's cross-process cargo build/test slot
# ceiling (find_tag_gaps.build_semaphore) -- the max number of concurrent
# cargo build/check/test invocations across every worker+merger sharing
# this host, so a full round of workers all rechecking gaps at once can't
# oversubscribe the host's cores by linking concurrently. 5 is the spec's
# own default for a 10-core host; raise only with measured disk/CPU
# headroom (see scripts/check_scale_gate.py).
build_semaphore = 5

# --- Phase 1 knowledge layer / section 6 knobs (spec K1-K5) ---------------
# learning_budget_tokens: flat, reserved token budget for build_prompt's
# learning block (GLOBAL-PITFALLS.md excerpt + module playbook + sweep
# reviews + lessons tail) -- never squeezed further by the graduated
# per-section truncation, and never dropped entirely.
learning_budget_tokens = 1200
# parser_floor_tokens: the parser-files section is elastic under overflow
# but never shrinks below this -- "elastic with a floor", never squeezed
# to zero (the inverted-starvation critique this resolves).
parser_floor_tokens = 2000
# lessons_tail_kb: how far back build_prompt seeks into the tail of
# logs/lessons.jsonl (bounded, no full scan of a ledger that only grows)
# when rendering the learning block's "recent lessons" entries.
lessons_tail_kb = 256

# Tiered routing: explore-phase turns (investigation, REQUEST/VERIFY
# round-trips) use the cheap fast model; patch-phase turns (final diffs,
# repairs, retries) use the strongest model. Omit "phase" on an entry to
# use it for both.
[[worker.models]]
name = "gpt-5.6-terra"
phase = "explore"
reasoning_effort = "medium"

[[worker.models]]
name = "gpt-5.6-sol"
phase = "patch"
reasoning_effort = "max"

# Mixed-provider example: an entry with its own base_url/api_key pulls that
# one model from a different provider (see the header comment above).
# [[worker.models]]
# name = "accounts/fireworks/routers/kimi-k2p7-code-fast"
# base_url = "https://api.fireworks.ai/inference/v1"
# api_key = "your-fireworks-api-key-here"

# Optional: omit this table entirely to have the reviewer reuse [worker]'s
# pool/config exactly. Add it to review with a different model pool and/or
# provider than the fixer.
[reviewer]
base_url = "https://pass.wafer.ai/v1"
api_key = "your-api-key-here"
max_tokens = 256
reasoning_effort = "max"
stream = true
thinking = true
temperature = 0
timeout = 600
# deadline_seconds (default 120): wall-clock ceiling on ONE model call,
# checked as stream chunks arrive and NOT reset by arriving data. This is
# the knob that actually bounds a call -- `timeout` above is urlopen's
# SOCKET read timeout, which every arriving SSE chunk resets, so under
# stream = true it cannot bound a slow-drip response at all (measured on
# one provider: a single request ran 2118s against timeout=1200). On
# expiry the connection is closed and the request is replayed, counting
# against max_retries.
deadline_seconds = 120
max_retries = 1000
retry_backoff_seconds = 2
max_retry_backoff_seconds = 120
# reviewer_max_prompt_tokens: independent of the fixer's max_prompt_tokens
# -- the reviewer prompt now carries the Perl reference, live post-fix
# evidence, and a scoped emission scan (spec K5) alongside the C1-C5
# checklist, so it gets its own generous budget.
reviewer_max_prompt_tokens = 8192

[[reviewer.models]]
name = "glm5.2-fast"

[[reviewer.models]]
name = "MiniMax-M3"

[[reviewer.models]]
name = "Kimi-K2.6"

[[reviewer.models]]
name = "accounts/fireworks/routers/kimi-k2p7-code-fast"
base_url = "https://api.fireworks.ai/inference/v1"
api_key = "your-fireworks-api-key-here"

# Optional: configures scripts/parallel_tag_fix_loop.py. Omit entirely to
# use its own defaults (workers = 4, max_tags_per_process = 1).
[parallel]
workers = 4
max_tags_per_process = 1

# Optional: spec S3's T3 TABLE-PORT / T4 FOUNDATION-UNLOCK job knobs --
# scripts/model_fix_loop.py's attempt_table_port/attempt_foundation_job.
# Omit entirely to use the spec's own defaults (16384 / 8). These tiers
# are opt-in (no job runs unless explicitly dispatched); this table only
# tunes them when they do.
[table_job]
max_prompt_tokens = 16384
max_repair_rounds = 8
# model: the strongest configured model, per spec S3 -- tag the
# strongest [[worker.models]]/[[reviewer.models]] entry(ies) with
# phase = "table" (alongside the existing "explore"/"patch" tags) rather
# than repeating a model name here; attempt_table_port/
# attempt_foundation_job select via models_for_phase(pool, "table").