Roadmap#
What is planned next for blyte, what is waiting on something, and what is parked. Dated items are
decisions; the evidence behind them is in docs/evidence.md and the design plans.
In progress#
Nothing at the moment; see Next.
Next#
- The Claude usage check (about 2026-10-06): regenerate the usage report from the dispatch ledger and compare Claude usage with and without blyte across repositories and sessions.
- The decider shadow trial (about 2026-10-20): read its summary and keep it or remove it.
- The open-source decision (about 2026-10-27): if yes, the items under Parked come next.
Done#
Kept for about the last four weeks; older items are in the git history and the evidence page.
- 2026-10-02
- Decided:
blyte.tomlstays per user and untracked; each person runsblyte initin their checkout. Gates differ by machine (offline uv caches, a pinned uv, local pre-commit environments), and the gate-drift warning tells each person when theirs falls behind CI. - Dispatch time checked after the v0.7.6 rule change: tools time stayed low, and the blyte-only canary on v0.13.0 matched 27 September on tokens and cost (51.2M / $8.51 against 51.9M / $8.81), with the pass rate the same or better; the extra wall time was Gemini's latency.
- A run's work counts as merged once it is on the branch it was dispatched from (
--base, now recorded in the start file), not only the default branch, so runs on a long-lived feature branch are cleaned up when merged into it.blyte cleanreports other worktrees (sessions' siblingrepo--xworktrees) whose work is merged and that are clean, with thegit worktree removecommand; it never removes them. - Plans view: a plan file dispatched as the task is the run's plan, wherever it lives (sibling
worktrees included), read from git history once that worktree is gone, and matched by the run id
when there is no
--plan-task(v0.14.0 and after). - v0.16.0:
blyte dispatchwarns when the gate is missing checks the repository's CI runs, matched by tool;blyte initalso recognisesmake typecheck/validate, eslint, vitest, jest, tsc;skill install/updatewarn about a supersededblyte-researchskill; skill guidance on real-data checks, shared registration files and plan files as task files. - v0.17.0: follow-up chains and squash-merged runs count as merged (cherry-pick record), and the
dispatch-time clean is cheap; the first
blyte clean --allon the author's machine freed 21 GB. - 2026-10-01
blyte clean(also run quietly by everyblyte dispatch): a run's worktree goes once its work is on the default branch (local ororigin/, including squash merges of several runs), it was redone or made no changes, its research report is saved, or 14 days have passed; never with uncommitted changes, and the run's record files stay. Branches go only when merged. 29 worktrees held 15 GB on the author's machine, 21 of them for reviewed runs.- Honest outcomes: a gate that fails the same way before the change ends the run as
blocked(exit 5) instead of burning the budget; the deadline watchdog says what was running when it stopped a run (v0.11.0). Gate slots default to a quarter of the cores (at least 2), time waiting for one no longer counts against the run's deadline, and research skips them; a long failed gate keeps the test runner's summary lines (v0.11.1). - Needs attention lists only runs nobody has dealt with (no verdict, not superseded by a later
run, worktree still there, under three days old); the Plans view lists plans with a running
task, keeps each node's state colour when selected, shows a task's merged run before its
latest, and draws follow-up tasks dispatched with
--planbut not written in the plan as "added during the work" (v0.11.2). The two watchdog tests take a second instead of 32 s. - The dashboard's Plans view: each plan as a swimlane flow of its tasks (lanes by area of code,
columns by dependency), coloured by state (merged, verified, running, failed, redone, ready,
waiting), with the selected task's details, the plan's activity and decisions, all from files
blyte already writes (
/api/plans).blyte dispatch --plan/--plan-tasklinks a dispatch to its plan task; older dispatches are matched by tag, task text and theirPlan:line. - Research runs read the repository with the
read,grepandglobtools, default to 300 turns, and keep their report current; every run is warned ten turns before its turn budget ends. blyte dispatch --research: read-only research as a dispatch (a scratch workspace under.wt/, gate "report.md written", the report copied beside the run), so it gets the registry, status, dashboard and review like any dispatch; the skill says when to use it. Run reports record the model's effort level (effort), whichmodel_idhides.- The dashboard names a dispatch after its repository, not its worktree's folder (v0.7.8).
- Dispatch rules tell the model not to run the full suite itself (v0.7.6): tools time had risen
from a median of 44 s to 198 s after the switch to plain
blyte dispatch. - The running dashboard sees repositories registered after it started;
blyte initrecognisespython -m pytest/unittest,toxandnox(v0.7.5), and keeps one of two pytest steps that test the same files. - The sandbox lets a worktree write the tool caches of an enclosing
node_modules(vitest's.vite-temp), never its packages. - Linked worktrees use the main checkout's
blyte.toml;initfalls back to the repository's own checks when CI is stale, and calls a cold uv cache an environment problem (v0.7.4). - The repository registry writes one whole line per dispatch and skips temp directories (v0.7.3).
- The default Gemini model runs at high effort, as the docs said (v0.7.2); a run never outlives
--max-secondsby more than 30 s and always leaves a report (v0.7.1). - The operator's local dispatch scripts and their sampled A/B against agy are retired (74 paired
tasks: 0.55x agy's time, 0.64x its cost). Every repository on the operator's machine now
dispatches with plain
blyte dispatch. - 2026-09-30
- blyte owns the whole dispatch workflow, so the operator's local wrapper scripts can retire:
blyte review --verdictrecords the reviewer's verdict;reviewflags commands that reached outside the worktree;blyte.tomltakes[dispatch.env]andfold_budget; dispatch saves its summary and names the gate's quick steps;blyte inittakes the gate from CI, sets what the environment needs and proves each step in a fresh sandboxed worktree; the dashboard watches every repository blyte has dispatched in; the skill has tasks declare their dependencies. blyte dashboard: a local page (127.0.0.1) of running and recent jobs across repositories and run directories, with time, cost, turns, the task's section of its plan, and clickable tags (repo, plan, harness, A/B) that group related jobs.blyte dispatch --tagadds tags.- A release on every merge (commitizen, semantic versioning, #157/#159/#162): a release PR, a
vX.Y.Ztag, the GitHub release with its changelog, and static binaries for macOS and Linux. blyte updateinstalls the latest release in place, with--checkand--rollback(#169).- blyte loads the repository's rules for agents itself on every run (#160):
.blyte/rules.md, else.claude/agy-rules.md, elseAGENTS.md, however it is started. - Codex CLI run end to end (#164): the skill names the escalation Codex must request; the "(untested)" labels are gone.
- Apache-2.0 only (#166), and a dependency licence check in CI (
cargo deny, #168): every dependency is permissive. - Fewer, larger dispatches as the default guidance (docs, skill and the eval's tools).
blyte skill install --no-defaultremoves an earlier default block, so switching from on to "only when asked" needs no uninstall first.- Every blyte run is recorded for reporting, including ones started with
blyte dispatchdirectly, with a check of whether each one's work reached main (local tooling). - 2026-09-29
- More evidence on katex: v2 repeated (Claude usage down 86%), and a parallel test that showed splitting one feature does not pay (evidence page, #154; blyte-eval #88).
- Provider keys from
~/.config/blyte/envwhen they are not exported (#151): a session started without the key failed every dispatch at turn 0. AWS_PROFILEandAWS_DEFAULT_PROFILEwithheld with the credentials (#150): blyte had edited a repository'sconftest.pyto work around a profile the sandbox cannot read.- Research mode (local dispatch scripts):
dispatch-auto.sh --research, read-only, with a "report exists" gate, the sampled A/B and ledger records. - The eval's dispatch tool runs dispatches in parallel, each in its own copy of the repository (blyte-eval #83), with start-then-wait tools (#85), because Claude Code runs an MCP server's calls one after another. Results on scc: blyte-eval #86.
Information gathering (2026-09-29 to about 2026-10-27)#
- blyte is the implementer in every local repository (the A/B against agy ended 2026-09-30). Check usage across repositories and sessions around 2026-10-06.
- The decider shadow trial (a local model classifying failed gates) runs alongside, testing only. Decide whether it earns a place around 2026-10-20 to 10-27.
Parked#
- A shared Cargo build directory for worktrees. Measured 2026-10-01: a warm build is 15 s of a gate of about two minutes, which the test run dominates (its two slowest tests were then cut from 32 s to 1 s each). Worth it only if builds come to dominate; it would need a sandbox allowance, and parallel dispatches would queue on cargo's lock.
-
Codex as the planner, step 2. The blocker (Codex's 5-minute MCP tool timeout) is fixed in blyte-eval #93; rerunning the Codex + blyte against Claude Code + blyte comparison is parked (2026-10-02): the failures were likely down to the Codex subscription level.
-
Open-sourcing blyte: done. The repository is public at
github.com/dnf0/blyteand every release publishes the crates to crates.io.test_public_namesstill fails on any internal name in the files. - Docs hosting: done. GitHub Pages serves the docs at https://dnf0.github.io/blyte/, deployed by a
workflow on merge to
main. Still open: a custom domain (about USD 10-50 a year), which would keep links stable if the repository moves. - Still open now that the repository is public:
blyte updateupdates the way blyte was installed:cargo install blyte --locked(orcargo binstall blyte) for a cargo install, detected from~/.cargo/.crates2.json, and the release binary over plain HTTPS otherwise, with no GitHub CLI needed. Today it always downloads the release binary throughgh.- Contributions: a DCO sign-off or a CLA, decided before accepting outside pull requests.