Compact prompts and tool discovery

The prompt profile is optional. The default remains classic.

KYROZEN_PROMPT_PROFILE=compact kyrozen --project .
# From this source checkout:
KYROZEN_PROMPT_PROFILE=compact make run
# Restore the classic prompt:
KYROZEN_PROMPT_PROFILE=classic kyrozen --project .

The same environment variable applies to TUI and web processes. Invalid values raise an explicit configuration error when the runtime builds the prompt.

Compact mode keeps full descriptions for list_dir, find_files, read_file, write_file, run_cmd, and discover_tools. Other permitted tools appear by exact name, grouped by capability. The agent can request their descriptions:

Action:
{"action":"discover_tools","args":"git_log,git_show"}

An empty argument returns the permitted catalog. Named requests return complete existing tool descriptions, including their argument instructions. Descriptions remain available for the rest of the current turn, and reset on the next turn. No descriptions or permissions transfer between sessions. Unknown or unavailable names produce an error without revealing restricted tool descriptions.

Discovery intersects the configured capability bound, active surface token, and interaction mode. Ask and Plan remain read-only. Tool execution retains its existing capability, approval, workspace, and receipt checks. Dynamic tools remain visible only with dynamic permission. MCP lists the new read-capability tool using the existing generic { "args": "..." } input contract.

Compact mode removes repeated workflow text and duplicate continuation guidance. It preserves role configuration, untrusted-input handling, mode controls, the Action protocol, durable task completion, and verification requirements. It does not use another model or grant additional permissions.

Live pilot: 2026-10-01

INCONCLUSIVE. No superiority or token-savings claim is supported.

The pilot compared baseline OpenKyrozen at 3536777, compact OpenKyrozen, and installed CodeWhale 0.9.12 (dcd4c200f72f). All 100 upstream requests used deepseek-v4-flash, the same credential and endpoint, thinking disabled, temperature 0, top-p 1, and a 4,096-token output limit. All returned provider usage. Total consumption was 222,375 input-plus-output tokens.

The request cap was reached before the requested two repeats completed. Baseline and compact each ran four fixtures once; CodeWhale ran three once. Raw arm totals cover different workloads and cannot establish savings.

Compact produced code that passed the independent regression-fix and implementation tests, but left pending durable tasks in both runs. Those runs fail overall completion acceptance. Both OpenKyrozen arms failed the Git-history answer check. CodeWhale's history fixture and all second repeats were not reached. These are observations on this small workload, not general quality conclusions.

The initial CodeWhale adapter missed its content answer event. Its investigation answer was inspected during the run and regraded from that exact event without new requests. OpenKyrozen's initial tool counter was also unavailable because CLI chat replaced the event callback. Unavailable metrics are marked null in the sanitized data. Both adapters now have regression coverage; their corrections have not been rerun live because the cap is exhausted.

Reproduce

Use Python 3.12/3.13 and installed CodeWhale. Preserve a baseline checkout:

mkdir -p /tmp/kyrozen-baseline
# Export the exact pre-feature commit, without local runtime data:
git archive 353677747bef090db4be3245560e2aac5bb04ff3 | tar -x -C /tmp/kyrozen-baseline
venv/bin/python benchmarks/prompt_comparison.py \
  --baseline /tmp/kyrozen-baseline \
  --output /tmp/prompt-comparison.json

Each invocation is a paid live run. The default limits are 100 requests and 1 million tokens; do not run again under an already exhausted budget. The runner reads DEEPSEEK_API_KEY or the existing encrypted OpenKyrozen DeepSeek config. Credentials never enter subprocess arguments or exported data. Missing credentials or API incompatibility blocks the comparison.

The runner uses fresh temporary homes, deterministic Git fixtures, and real CLI execution. It disables user plugins, background learning, System One, provider fallback and model routing. It runs each product's prompt, planning, tools, recovery, receipts and completion logic. CodeWhale uses exec --auto --output-format stream-json; OpenKyrozen uses its production Rich CLI chat path.

A localhost relay preserves each client's messages and tool protocol, normalizes only model/API settings, records usage, and reserves an upper input bound plus maximum output before forwarding each request. Failed requests retain their reservation when usage is unknown. Retries and summaries count toward the caps. The relay converts a non-streaming upstream response into SSE when requested; latency is therefore descriptive, not a production streaming-speed benchmark.

Original fixture tests run outside the agent and unchanged-file checks prevent passing by deleting or weakening tests. Complete acceptance also requires all OpenKyrozen durable tasks to succeed. The claim gate requires complete matched pairs, identical passing outcomes, known usage, verified settings, and at least 10% lower aggregate total tokens. Reasoning tokens are not added again to completion tokens. Incomplete, failing, or unknown-usage evidence yields no claim.