Judge each agent's best roles against its CV as skills, experience and career, a stage after summarise

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
WolfandClaude Opus 5.5 committed 2026-10-03 22:00:18 +02:00
1 parent 4fc474c437
commit 9960799b96
14 files changed
+866 -44

No files matched your search

+22 -6
View File
@@ -46,7 +46,7 @@ workers:
# `agent` extra.
agent:
backend: auto # auto | api | claude_code | off
model: "" # overrides extract.model and summarise.model when set
model: "" # overrides extract.model, summarise.model and judge.model when set
timeout: 300.0 # seconds for one model call, either backend
max_tokens: 8192 # the answer's cap per call
cli_path: "" # where the `claude` binary is, if not on PATH
@@ -81,6 +81,21 @@ summarise:
# runs summarise when it is on and a model is reachable (app/jobs/cascade.py).
# Only extract reads it.
# Each search agent's best roles read against its CV: skills, experience and
# career 0 to 100, a total, and 1 to 3 strengths and gaps (app/jobs/judge.py),
# for the Best CV match order on Roles. Off by default. On, each pull judges
# the best roles of every enabled agent with a CV, by that agent's judge
# limits, and only those not judged yet or judged from another CV, prompt,
# avoid or must text, or role text. The CV rides in every call, so batches are
# small. `enabled` is only the default for a fresh database: the switch under
# Settings, Sources, Runs (or `app stages judge on|off`) overrides it and needs
# no restart. on_ingest is accepted and ignored, as for summarise.
judge:
enabled: false
model: claude-haiku-4-5-20251001 # agent.model wins when set
batch_size: 5 # roles per model call
thinking: off # off | adaptive
# Reading the best roles on their own pages, for the hosts whose list carries
# no text (app/jobs/enrich.py); asks no model. A role is read when it has no
# text and an enabled agent with a prompt ranks it ungated among its best
@@ -136,14 +151,15 @@ sources: {}
# 0.2}; taste is how close a role is to the ones the agent liked minus the
# ones it passed. A country is an ISO code, filled from the board's location
# line; EU is a role open anywhere in Europe.
# limits is how deep into this agent's ranking the two model stages read,
# each a top N (the backstop) and a minimum score (the real stop): extract
# reads the best extract_top roles at extract_min or more, summarise the best
# summarise_top at summarise_min or more. Fit is a calibrated level, so a bar
# limits is how deep into this agent's ranking the model stages read, each a
# top N (the backstop) and a minimum score (the real stop): extract reads the
# best extract_top roles at extract_min or more, summarise the best
# summarise_top at summarise_min or more, and judge the best judge_top at
# judge_min or more for an agent with a CV. Fit is a calibrated level, so a bar
# means the same in every run: a fresh role with the worst fit scores about
# 0.27 on the default weights, a perfect one about 0.87. The defaults are
# {extract_top: 500, extract_min: 0.55, summarise_top: 30, summarise_min:
# 0.65}; measured for "Copenhagen engineering" on 16,841 open roles, 0.55
# 0.65, judge_top: 150, judge_min: 0.55}; measured for "Copenhagen engineering" on 16,841 open roles, 0.55
# admits about 300 roles to extract and 0.65 about 35 to summarise. A wide
# monitoring search lowers the bars and raises the counts, a narrow hunt the
# reverse. An agent without a prompt that has text has no fit and feeds