Repository navigation
v0.8.0
Kapso 0.8.0 — release notes
Turn coding-agent output into verified outcomes, not review debt: plug in the harness for your domain, set the goal, every step traced.
This release lands the redesign as one change from the integration branch (docs/plans/redesign/MASTER.md). It renames the verbs, replaces the modes with one configuration, makes the harness the domain layer a build loads, makes experience something a build is served explicitly, turns deployment into target skills, records every session, ships a coding-agent plugin, ports the benchmarks' knowledge into two bundled harnesses, and reorganises the docs by what the reader is doing. The mapping below is complete for the surfaces that changed; a name on the left no longer exists. Every line was checked against kapso --help, each verb's help and the packaged configuration at the release candidate.
The verbs
| Before | Now |
|---|---|
kapso evolve … |
kapso build --brief kapso-brief.md [--evaluation DIR] [--data DIR] [--repo PATH] --output DIR [--detach] [--harness REF] [--experience REFS] [--budget-minutes N] [--budget-experiments N] [--budget-usd N] [--budget-reserve-minutes N] [--resume] [--stop] |
kapso bank create, kapso bank connect, kapso learn init-bank |
kapso experience add mine, kapso experience share <url> |
kapso research "…", kapso learn knowledge …, --knowledge, --kg-index, --no-kg |
kapso experience add --research "…" --name <index> / kapso experience add --repo <url> --name <index>; served to a build with --experience <refs> or the brief's Experience line; kapso experience list, update, remove |
kapso watch --tree |
kapso trace <campaign> (the attempts, the lessons, what was voided, what was delivered); kapso watch <campaign> --follow for a live view |
(new, from main) |
kapso report <campaign> [-o FILE] [--top N]: the campaign as shareable Markdown |
kapso deploy --strategy …, --interactive |
kapso deploy <campaign> --target auto|local|docker|modal|bentoml|langgraph|<a harness's target> [--env K=V] [--option K=V] [--status] [--stop] [--start] [--logs] [--number N] |
--mode, -m MODE, the modes: block, load_mode_config |
one packaged config plus one override file passed as --config |
| (new) | kapso harness new|check|list|add|update|remove|export; the brief's Harness line |
| (new) | kapso inbox / kapso inbox reply [<campaign>] [<id>] "<note>" for a build that is waiting on you; kapso doctor with --show-config |
kapso learn <subcommands> |
kapso learn <campaign> learns from a finished campaign; kapso learn dev … is the development pipeline |
The Python API
| Before | Now |
|---|---|
Kapso(config_path=…) |
Kapso(config=…) |
kapso.evolve(...) |
kapso.build(brief, evaluation=, data=, repo=, output=, budget=Budget(minutes=, experiments=, usd=, reserve_minutes=), harness=, experience=, resume=) → Solution (.score, .branch, .diff(), .explain(), Solution.open(path)) |
kapso.research(...), learn_knowledge(...) |
kapso.add_experience(repo=|paper=|research=, name=) then kapso.load_experience(ref) |
LessonResult from learn() |
kapso.learn(solution) → the lesson with its evidence (.explain()), banked under mine |
| the deploy strategies | kapso.deploy(solution, target=) → Software (.run(), .proof, .record, .stop(), .start()) |
| (new) | kapso.load_harness(ref); Brief.read(path) |
The configuration
- One packaged
config.yamlwith one override file per project; no modes and nodefaults:layer (an unknown key is refused by name).hosts:carries each role's host and model;harness:,experience:anddeployment:are new blocks;deployment.targetnames the default target;report.leaderboard_sizesizeskapso report's leaderboard. - Retired:
serving.enabledandserving_scope(what is loaded is what is served; a source'sscoperides on its reference,mine:alllifts it),strategies.yamland the deploy strategy classes (target skills instead),budget.min_iteration_seconds_insuredand the handlers' insured floor, thefeedback_generator:andcoding_agent:blocks (routes per role underhosts:),knowledge_search.use_knowledge_graph. - From
mainsince the branch:search_strategy.params.parent_policy: weightedwithparent_sampling(temperature, offspring penalty, seed);repo_memory_update_payload_max_chars(a changed file that is binary or larger than that budget on its own is listed by name and size in the update, never sent: the bundled example's dataset had put every update past Codex's input limit); oneexpansion_lane_enventry per lane.
The records
- Every session of a campaign writes
session.json(its module, pages, skills, hooks, runs with tools, hooks fired, cost and tokens);harness.jsonis gone..kapso/launch.json,status.json,solution.jsonand the trace carry the harness and the experience a campaign ran with, every path resolved; a resume refuses a different list.
The docs
- The site is organised by what the reader is doing: Build, Harnesses, Experience, Learn, Deploy, How it works, Benchmarks, Reference. Learn is two pages (
/docs/learn/what-learn-produces,/docs/learn/development-mode); the six/docs/learning/*pages redirect./docs/agent-accessis part of/docs/coding-agent-skills;/docs/deployment/*→/docs/deploy/targets;/docs/knowledge/*→/docs/experience. - Two bundled harness pages (
predictive-modelling0.3.3,competitive-programming0.3.2); every benchmark page says who ran its numbers, on what and when; two claim lints hold every public sentence to the project's own rules.
The plugin
claude plugin marketplace add Leeroo-AI/kapso && claude plugin install kapso@kapso(Codex:codex plugin marketplace add Leeroo-AI/kapso && codex plugin add kapso@kapso): the commands/kapso:setup,/kapso:build,/kapso:watch,/kapso:inbox,/kapso:resume,/kapso:review,/kapso:learn,/kapso:deploy,/kapso:harness,/kapso:experience; six skills; thekapso-mcpserver.
The examples
full-loop(build, the review, learn, build again served the lesson, deploy),churn-first-build,fraud-no-evaluator,pipeline-dbt,trading(on the trading harness underexamples/harnesses/),prompt-harness,agent-scaffold,gpu-kernels,torch-performance;examples/README.mdsays which have a recorded result.
What was measured, and what was not
- The semantic suite from the release wheel on fresh machines, run by us from 2026-10-30 04:36Z to 2026-11-01 00:09Z (the reports
tests/semantic/reports/rel-*.md, each launch read to its cause indocs/plans/redesign/wp11-release.md; forty-four launches, each on a fresh machine of its own: five of S1 and four of S4 and S5 on Claude Code, five of S1, three of S4 and six of S5 on Codex, eight of S2 on Claude Code and nine on Codex): seven of eight reports passing and the eighth, S2 on Codex, at 22 of 24 with two findings for the owner — S1 on both hosts (17 of 17 steps, every bar met, the judge's three dimensions at 1); S4 on both hosts (every reading step passing, the campaign's integrity void held and read, the rejection recorded rather than required); S5 on both hosts (8 of 8 injections diagnosed truthfully with their class and fix named; the Codex record read again under the current vocabulary, said in the record); S2 on Claude Code PASS (25 of 25 steps, the judge's bar met at 5, 4 and 4: its eighth launch, the record read again under the current readers for a closing line's ask and a batch-shaped deployment answer); S2 on Codex FAIL at 22 of 24 on its ninth launch (its probe clean, the churn build delivered and deployed at parity, the baseline measured, the learn banked, the judge's proactive and clear at the bar): two misses that are findings for the owner — a Codex session served the bank found no campaign verb while the same server started standalone listed them in six seconds, and a second engineer's attempt that scored 0.8494 lost its delivery to a verification cut bybudget.verification_minutes— with the judge's return on attention at 1 on the campaign-tools answer alone. The gate found, and this release fixes: the repo-memory update and its first inference bounded to the update's payload budget, a binary or over-budget file withheld by name and size (a campaign's 1 MB dataset sent whole had put every call past Codex's input limit); the Python API's detached launch,build(detach=True)andlearn(detach=True)handing back aLaunch(every S1 engineer on both hosts had reinvented it); a bare resume running on the launch's budget rather than the brief's; an answered inbox continuation admitted whatever the remainder (the orchestrator had refused a resume on the paused attempt's own elapsed time, and the reply never reached the session);kapso deploy --statusanswering when a campaign has no deployment; the trace header's voided and rejected counted apart; the deployment's proof reproducing the delivered prediction (a batch-rank model keeps its batch); and in the plugin, the harness written where the repository loads it, the answers shaped in counts with their detail in the fence and their findings never dropped, and the Codex MCP registration unconditional and verified. It raised for the owner, in the sub-plan and DECISIONS: an implementation session restoring the protected labels from a copy it could reach, against the brief's rule, with the judge letting the score stand;rejected_candidatescounting integrity voids; an attempt failed on the provider's 429s before any implementation turn counting against the experiment allowance; the feedback quoting the brief's budget where the launch's differs; the ideation handoff storing the session's last reply, a cleanup line after background-task notifications, in place of its<solution>; a learn on a one-attempt campaign taking two to two and a half hours on the Opus accommodation model;budget.verification_minutes(10) short of a multi-model blend's reproduction on a fresh machine, a scored attempt then lost to the finishing cut; a Codex session served the bank seeing the bank tools and not the campaign verbs, which the same server started standalone lists in six seconds; the bundled harnesses' requirements under a tool install; the S4 rejection bar changed to the void alone, since no provided-evaluator campaign records a judge rejection; the learn's banked cards a recorded outcome, not a bar; and the plugin's two advisory judge dimensions. S6 (experience pays) from the suite's build; S3 (the harness ports) from the arms, each a quick run at the budget its record states beside the recorded one: the MLE quick arm (spooky-author-identification, three hours on one GPU) graded at a log loss of 0.2271 on the private answers, silver; the ALE quick arm (ahc008, three hours) accepted on every private case at a performance of 1363 against the recorded 1221; the RelBench six-hour arm (rel-trial/study-outcome, one A100, four lanes) at a test AUROC of 0.7555 on the lanes' registered finals, inside the recorded band, with no candidate delivered (an evaluator change request mid-iteration, which benchmark runs no longer file); every arm and swap on a pinned harness tree; the RelBench swap (rel-event/user-attendance, a second family and a minimised metric) at a test MAE of 0.2298; the ALE swap (ahc011) 38th of 699 on the private cases; the Kaggle swap (spooky-author-identification under the modelling harness) at a best public log loss of 0.30161 over four ledger submissions, with no candidate delivered (the provided evaluation read the CLI's status under the wrong spelling, fixed since). - Not measured: a cold start on the IOAI practice task (the one arm replayed the account's own earlier kernel, no candidate delivered; a cold start needs an account that never entered the competition); a maximised-metric RelBench swap and a tabular Kaggle swap (the tasks run were a minimised-metric regression and a text classification); the MLE arm ran with the web on in implementation, which the modelling harness's configuration has since turned off, and its in-loop score is the candidate's own hold-out number, checked by the verifier and the final private grade; no arm reran for the review's prose fixes to the harnesses (0.3.3, 0.3.2).