Documentation for v0.36.0

Compatibility and release verification

This page records what was observed, on which machine and day, before each tag. A passing check shows that what was tested ran; it qualifies no model and grants nothing (which model to use is your choice). Each release replaces the snapshot below with the numbers observed after running the gates. Snapshots 0.16 through 0.26 live in CHANGELOG.md and in this file’s history.

Verified snapshot

0.36.0, checked on 2026-10-28 on macOS with Claude Code 2.1.312, Codex CLI 0.165.0, Python 3.15.2 and gh 2.103.3, from the release candidate.

• Install and doctor: source installed into isolated Claude, Codex, bin, data and personal-configuration directories. Static doctor: 49 OK, 1 WARN, 0 FAIL; doctor.sh --ping: 51 OK, 1 WARN, 0 FAIL. The one warning is the intentionally absent normal Codex config.toml. The installed luna-clerk/max returned PONG, then Claude sonnet/low returned PONG. Exactly two separately authorized calls were consumed, with no retry or fallback. Claude’s call disabled tools, hooks and MCP through the release wrapper. These are reachability checks, not provider identity attestation or qualification of the new builder/reviewer choices.
• Offline: 14 suites, 11 npm-wrapper tests, ruff, ShellCheck and version consistency passed. Router regressions cover suite-mismatched overlays, null scores, unmeasured effort, exact slugs and nonmonotonic effort scores.
• Catalog: the reviewed controller snapshot uses Intelligence Index v5.0 and unchanged official token prices. A read-only native refresh over the same 715 API rows agreed with every controller index, preserved official price/source facts, selected DeepSeek 1017 at 32.5, and kept Muse Wonder Full’s unknown price unset. The controller-generated catalog is retained so its version/source provenance and consumer sync stay reproducible.
• Installation regression: the first isolated install revealed that legacy cleanup removed the newly reintroduced terra-builder files. Cleanup now skips shipped names, and regression coverage checks the new Terra files, their model/effort, and the read-only Opus judge. The corrected installation was verified before either live call. Existing edited user profiles retain the installer’s preservation behavior.

Earlier releases

0.35.1, checked on 2026-10-27 on macOS, from this release candidate.

• Install and doctor: installed into isolated Claude, Codex, bin, data and personal-configuration directories. doctor.sh returned 50 OK, 1 WARN, 0 FAIL; doctor.sh --ping returned 52 OK, 1 WARN, 0 FAIL. The warning is the intentionally absent normal Codex config.toml in the scratch installation. The installed luna-clerk profile returned PONG, followed by Claude sonnet at effort low. Exactly two calls were authorized and observed, once each, with no retry or fallback. These reachability checks do not attest the provider’s model identity and did not dispatch Terra through this kit.
• Offline: 14 suites passed in 45 seconds, 11 npm-wrapper tests, ruff, ShellCheck and version consistency.
• Catalog: generated from controller-models, with Terra’s official pricing and six Artificial Analysis effort results verified 2026-10-27. Existing shipped presets and personal user configuration stay unchanged. A fresh read-only trial of the upstream refresh still matched deepseek-flash to the older 39.5 row instead of 1017 at 45.1, inserted a zero price for Muse Wonder Full and discarded official price sources. As for 0.35.0, the reviewed controller output is retained; the upstream refresh is not written over it. Terra’s six results agree with the fresh Artificial Analysis response.

Additional earlier releases

0.35.0, checked on 2026-10-25 on macOS 26.5.1 with Claude Code 2.1.310, Codex CLI 0.164.0, Kimi Code 2.3.1 and Muse Code 1.4.13 (1.4.13-R6996.1).

• Offline: the 14 suites green in 53 seconds, ShellCheck at -S warning on the 30 files CI collects, ruff, the eleven Node tests, the version gate and git diff --check.
• Install and doctor: a reinstall from the release branch removed no edited profile and installed sol-builder and sol-reviewer in place of the terra-* ones. Static doctor.sh reports 57 OK, 6 WARN, 0 FAIL; doctor.sh --ping: 59 OK, 6 WARN, 0 FAIL, the Claude → Codex round-trip returned PONG.
• Catalog: written by controller-models from its registry, with that day’s Artificial Analysis index and every price re-checked on its official page on 2026-10-25. catalog refresh --shipped was not used: it matched deepseek-flash to the retired V4.1 Flash (39.5) instead of V4.1 Flash 1017 (45.1), priced Muse Wonder Full at $0 and dropped the price sources.
• External models: none was called through the kit’s runners for this release (the GLM lane has no Z.ai key on this machine). The usage key delegation-grok now reads, grok-4.7-build, was observed on 2026-10-25 in a Grok Build CLI run with grok-4.7 made through bottega-kit.

Additional earlier releases

Newest first. “Suites” means the regression suites; every live call was authorized and no key was printed.

• 0.34.1 (2026-10-24, Claude Code 2.1.310, Codex 0.164.0, Kimi Code 2.3.1, Muse Code 1.4.12-R6909.1): suites in 48 s, ShellCheck, ruff, Node tests, version gate; doctor 57 OK, 6 WARN, 0 FAIL, with --ping 59 OK. Catalog refreshed, nothing changed. The old matteoscurati/delegation-kit URL still listed the 0.34.0 tag.
• 0.34.0 (2026-10-24, Claude Code 2.1.309, Codex 0.164.0, Kimi Code 2.3.1, Muse Code 1.4.12-R6909.1): suites in 43 s (new: the Muse runner suite, openai-compatible effort, contributor_mode); doctor 57 OK, 6 WARN, 0 FAIL, with --ping --ping-muse 60 OK. Catalog: Muse Spark 1.3 (48.1, mid), Muse Wonder (52.4) and Muse Wonder Full (55.8, hand-priced at $4/$20) new. Live on the subscription, Contributor mode off: Spark PONG from an isolated home, a sandboxed read and a one-file builder edit, muse-code-wonder and muse-code-wonder-full PONG through delegation-run. API-key Muse profiles not dispatched.
• 0.33.0 (2026-10-23, Claude Code 2.1.309, Codex 0.164.0): suites in 43 s; a fresh init validated and listed minimax-m3.1-flash; doctor 55 OK, 6 WARN, 0 FAIL, with --ping 57 OK. MiniMax not dispatched. Catalog refreshed, nothing changed.
• 0.32.0 (2026-10-23, Claude Code 2.1.309, Codex 0.164.0, Kimi Code 2.3.1): suites in 44 s (two new Kimi runner cases); doctor 55 OK, 6 WARN, 0 FAIL, with --ping --ping-kimi 58 OK, Codex and Kimi K3.1 Pro both PONG. Catalog refreshed: Kimi K3.1 Pro new at 55.1 (frontier).
• 0.31.0 (2026-10-23, Claude Code 2.1.309, Codex 0.164.0): suites in 44 s (86 router tests, seven new for conversation pins, out-of-schema answers, harness blocks and native dispatch cancellation); doctor 55 OK, 6 WARN, 0 FAIL, with --ping 57 OK. Catalog refreshed, nothing changed. Live: the Jev decider, decider eval runs and 250 pool dispatches for the golden sets; see Decider measurements.
• 0.30.4 (2026-10-21, Claude Code 2.1.306, Codex 0.163.0): housekeeping; suites in 51 s (79 router tests; the installer suite at 18, one new for files a release no longer ships); a reinstall removed the two library modules earlier releases had left; doctor 55 OK, 6 WARN, 0 FAIL. Catalog refreshed, nothing changed.
• 0.30.3 (2026-10-21, Claude Code 2.1.306): documentation and one check warning; suites in 214 s (79 router tests), ShellCheck, ruff, Node tests, version gate; doctor 55 OK, 6 WARN, 0 FAIL, and without a personal configuration it named the missing keys file and reported four FAILs without stopping. Catalog refreshed, nothing changed.
• 0.30.2 (2026-10-20, Claude Code 2.1.304): suites in 49 s (79 router tests), ShellCheck, ruff, Node tests, version gate; doctor 55 OK, 6 WARN, 0 FAIL, and without a personal configuration it no longer stopped on an unset variable. Catalog refreshed: MiniMax-M3.1-Flash 42.4 → 42.7, still mid.
• 0.30.1 (2026-10-20): docs and guard wording only; suites and version gate pass. MiniMax-M3.1-Flash dispatched once with its key only in the personal .env: PONG in 3 s, provider_reported_model MiniMax-M3.1-Flash (provider-reported), 214 input / 14 output tokens.
• 0.30.0 (2026-10-20): suites (new personal .env tests), ShellCheck, ruff, version gate. The reinstall created ~/.config/delegation-kit/.env from .env.example, mode 600; an openai-compatible MiniMax profile with its key only there was reachable in check. Doctor 55 OK, 6 WARN, 0 FAIL (new line: the keys file’s mode and variable count).
• 0.29.1 (2026-10-19): catalog only. MiniMax-M3.1-Flash replaced MiniMax-M3, matched by slug at 42.4 (mid, $0.15/$0.60); 24 of 25 placed; everything else unchanged since 2026-10-17. Suites and version gate pass.
• 0.29.0 (2026-10-19): suites (75 router tests), ShellCheck, ruff, Node tests, version gate. Eight-profile pool all reachable, runner checks 0.04–3 s (Kimi longest); doctor 54 OK, 6 WARN, 0 FAIL. From a source checkout the GLM runner’s key file was not found and check reported it unreachable. Live: two decide --lane builder calls to the DeepSeek Flash decider, nothing dispatched: a rename rated 0 → glm53-flash-max-builder (cheapest of five builders), a Go deadlock 2.5 (minimum 43.6) → kimi-k3, the three profiles without the role blocked; 0.8–1.1 s each.
• 0.28.0 (2026-10-19): suites (70 router tests), ShellCheck, ruff, Node tests, version gate; doctor 54 OK, 2 WARN, 0 FAIL (luna-clerk, opus-reviewer: native limits). Live: four decide --raw calls to the DeepSeek Flash decider, no pool profile dispatched: a rename 0 (code), a contract summary 1, a Go worker-pool deadlock 2.5 (code), a four-colour proof 3.5 (math, rare); 0.8–1.1 s, about $0.00014 each; minimums 0, 0, 43.6, 49.6. README routes checked against choose() with the shipped presets.
• 0.27.2 (2026-10-19, Claude Code 2.1.303): suites (installer PATH warning asserted both ways), ShellCheck, ruff, Node tests, version gate; doctor 53 OK, 0 WARN, 0 FAIL with opus-builder and opus-reviewer pinned to claude-opus-5-5/max. Live: delegation-run --profile opus-reviewer --lane routine-review answered PONG in 3 s; receipt requested_model: claude-opus-5-5, model_identity_source: requested-only (Claude Code’s JSON result names no model). README routing example checked against choose().
• 0.27.1 (2026-10-17): npm wrapper only; suites, ShellCheck, ruff and the eleven Node tests pass, three of them new for --ref.
• 0.27.0 (2026-10-15, re-checked 2026-10-17; Codex CLI 0.163.0, Claude Code 2.1.302): suites in 46 s (about 80 s on macOS CI), 64 router tests, ShellCheck, ruff, eight Node tests, ten-check version gate; reinstall and doctor.sh --ping 55 OK, 0 WARN, 0 FAIL, router not configured — optional. catalog refresh: 700 rows, 24 of 25 models matched by slug or alias; re-run with --shipped on 2026-10-17, placements and prices as on 2026-10-15. Deciders measured (below). A no-spend probe of serve against the real DeepSeek and Qwen endpoints with fake keys returned 401 provider-key-rejected naming the default profile, with no key in the response or in the single log line. Dipylon was not probed (no local service running); no pool model was dispatched.

Decider measurements

Live, 2026-10-15 (0.27.0), each call authorized.

What they rated: a rename 0.7–1.0 with code; a race-condition design question 1.6–1.7 (Sage, Jev; Sage adds rare 0.98) and 2.0 (DeepSeek, Qwen: “classic SELECT-then-UPDATE race, SKIP LOCKED”); a percentage question 1 with math. “Needs a frontier model” was ≈ 0.01–0.09 everywhere. The deciders agree with each other and rate the race condition lower than a human would, which is why eval on a golden set exists.

Live, 2026-10-22 and 2026-10-23 (0.31.0), each call authorized:

• Jev answered as jev-1.14.0 (it was jev-1.13.0 a week earlier) in 235–376 ms for ~$0.00002, its score always carrying probabilities keyed "0"–"4". Its own confidence was kept: Sage’s band mass in its place, replayed on 20 recorded answers, changed no answer and sent two more requests to the default.
• Harness blocks: 20 items, each also wrapped in OpenCode’s, Roo Code’s and Claude Code’s blocks, decided with 0.30.4 and with 0.31.0 by the DeepSeek Flash llm decider and by Jev. The tier never changed across variants (robustness 1.0 in all four runs); with the blocks dropped the decisions cost 25 % (DeepSeek) and 23 % (Jev) less. DeepSeek p50 1.1 s, p95 1.3 s; Jev p50 0.23 s, p95 0.33 s.
• Golden sets: the 20 items were answered correctly by all eight pool models, so they measure over-spending only. Of 30 harder candidates (program tracing, counting, search), GPT-5.6 Luna solved 29 and MiniMax-M3.1-Flash 25; on the five either missed, DeepSeek Flash, Opus 5.5 and GPT-5.6 Terra solved all five, GLM 5.3 Flash four, Kimi K3 and Qwen3.8 Max three (GLM and Qwen ran out of 30 minutes on one). Both deciders rated all 30 between 1 and 2.7, never at the frontier; the llm decider sent one of them to Kimi, which got it wrong, and Jev’s needs-frontier answer sent it to the default, which got it right. Kimi scores above DeepSeek on the index and did worse here: five items are a hint about the index, not a measurement of it.

Host-specific boundaries

Codex profiles

Use the shipped ephemeral profiles with:

codex exec --ephemeral -p <profile> "<self-contained prompt>" </dev/null

Do not add --ignore-user-config: Codex then ignores the selected -p profile as well and can silently use the ambient base model. Treat the model, effort, and sandbox shown in the invocation banner as part of the smoke assertion. Ephemeral profiles carry no role prose, so the prompt must be bounded and self-contained; native agent profiles carry their role through developer_instructions.

Claude Code agents

The clerk and scout agents expose Bash, to inspect Git state and run read-only search or test commands; the reviewers and the judge do not. For those two, read-only is role discipline plus the parent’s permissions, not Codex’s OS-enforced read-only sandbox: run them on a clean worktree and check git diff and git status afterwards.

For headless checks outside the current repository, pass the canonical fixture path through --add-dir; macOS /var and /private/var aliases can otherwise make an allowed temporary fixture look like a different path. --output-format json does not settle model identity: its result names no model (0.27.2 above).

External runners: see Executors for check --json and exit codes.

Reproduce the integration gates

./run-tests.sh          # every suite under tests/, in parallel
./doctor.sh --ping
git diff --check

For semantic verification, use disposable Git repositories and role-appropriate prompts: deterministic extraction for clerks, flow tracing for scouts, one bounded edit for builders, concrete bug detection for reviewers, and a reversible architecture decision for judges. Verify exact model identity, scope, output meaning, and worktree changes—not merely process exit.