Skip to content

Roadmap: October 2026 to September 2027

Chalk is maintained by one person, part-time. This roadmap is a plan, not a promise: each quarter starts by fixing whatever real runs broke, and items move when the numbers say they should. Quarters are calendar quarters, so Q4 2026 means October to December.

Each item links to its design notes in the detailed roadmap. Progress is tracked on the GitHub project board, and the pinned roadmap issue is the place to comment on priorities.

Principles that hold for every item:

  • Real runs come before new features.
  • Rules come before models. Anything that can change a run ships in shadow mode first: it records what it would have done and changes nothing. It is turned on only when the report card (chalk dashboard) shows it pays.
  • Optional services never block a run.
  • Chalk collects no usage data. Your code goes only to the model provider you configure, and anything else leaves your machine only when you choose to send it.

Already shipped in v0.6.0: stuck-loop verdicts in shadow mode (CHALK_FP_RULES), JUnit XML test reports (CHALK_TEST_REPORT), and Postgres 17 with pgvector and chalk db upgrade.

Shipped since: v0.7.0 brought the docs site, go test -json and jest --json reports, the auto mode check and a next-step line for every detention, and lesson recall by fingerprint. v0.8.0 brought PRIVACY.md and chalk share, and, ahead of Q2, the decider: CHALK_DECIDER, the local reference service, the calibration gate and the hosted-decider disclosure. Items below marked (shipped) are done.

Q4 2026: Make it trustworthy (v0.7)

  • Real runs. Use Chalk on itself and on Node, Python and Go repositories, on both GitHub and GitLab. Fix what breaks.
  • Hardening. Never lose a run's work to a crash, a signal or a database restart, show the agent's own error, keep the evidence a detention points to, and work behind TLS-intercepting proxies. The same design plans fleet scheduling and a harness that tunes itself to each repository.
  • (shipped) Docs site with a glossary of the school terms (rubric, detention, office hours, report card) and a recorded demo.
  • (shipped) More test report formats: go test -json and jest --json.
  • (shipped) chalk doctor and clearer detentions. Detention is where a run stops when it cannot make progress; each reason will say what to do next.
  • (shipped) PRIVACY.md and chalk share. See "Data" below.

Q1 2027: Cheaper loops, open models (v0.8)

  • (shipped) Lesson recall by the fingerprint of a failure, replacing the optional Hindsight memory server. Done in #55, released in v0.7.0.
  • Verdicts on: let the stuck-loop verdicts stop runs early, where the shadow numbers show it saves money.
  • Resume after a crash, sleep or Ctrl-C without losing paid work.
  • Bring your own model inside Claude Code: a model per kind of call; Bedrock, Vertex and gateways such as LiteLLM; and a price table for providers that do not report cost, so the report card stays accurate and a run stops once it goes over budget.

Q2 2027: Easier start, your own decider (v0.9)

  • Sandbox images for Python, Go, Rust and the JVM, and chalk init that detects your stack.
  • chalk spec draft to write checkpoints from a ticket.
  • GitHub and GitLab issues as ticket sources for chalk fleet.
  • Bring your own decider. A decider is a small, fast model that answers bounded questions such as "is this loop stuck?" so Chalk does not have to ask Claude. Mostly shipped early, in v0.8.0: you can run the local reference service or point Chalk at a hosted one. Every decider starts in shadow mode and has to show its confidence is trustworthy before it may act. Hosted deciders are never the default, and Chalk says exactly what they receive. Still to come: the adapter for providers that offer only a chat API.

Q3 2027: Beyond Claude Code, then 1.0 (v1.0)

  • Harness adapters. Claude Code becomes one adapter; Codex CLI is the first other one. The report card compares cost per finished checkpoint across harnesses.
  • Run in CI. Label an issue and get a pull request. Detention becomes a comment, and replying to it does what chalk office-hours does on your machine.
  • 1.0 stability: stable config keys, interfaces, schemas and flags, with a deprecation policy and a migration guide.

Items marked stretch in the detailed roadmap move to the backlog before any quarter slips.

Data

Chalk collects no usage data today, and that stays the default.

  1. You send it (Q4). chalk share builds an anonymised summary of your report card: version, OS, harness, models, stack, and counts and rates. It never includes repository names, paths, ticket IDs, code, prompts, test output or lesson text. It prints the exact JSON, and you decide whether to post it to a GitHub Discussion.
  2. Opt-in telemetry (Q3). Only if too few people post chalk share reports to tune the defaults. chalk init asks once, and the default is no. chalk telemetry show prints the next payload before it is sent; CHALK_TELEMETRY=off and DO_NOT_TRACK=1 always win; CI runs never report. The ingest code lives in this repository.

How to help

  • Run Chalk on a real repository and file what breaks with the real-run label.
  • Comment on any roadmap issue.
  • From Q4, post your chalk share report card in Discussions.

Not this year

  • A hosted service or paid tier.
  • A web UI beyond the report card.
  • Giving extra loops to runs that look like they are converging.
  • A central lessons server.

Backlog

See the backlog for what comes next when there is time: shared lessons, hooks and extra gates, Linux and Podman, budget alerts, and more.