Skip to main content

Changelog

v0.24.1 goals that name files get those files​

[fix] Files a goal names reach the model whatever the wording or language: .mjs, .ts, .py, .go, .md, .toml and other common extensions, plus root documents named in capitals such as README.

[fix] A goal that asks to add or write a test no longer passes when the diff changes no test file.

[add] Workers see a short list of the repository's files and can ask once for up to four of them with NEED_FILES: before writing the diff.

[fix] A model still thinking when the request times out is reported as a timeout, not as "error decoding response body". Only files inside the repository are ever shown to the model. See the v0.24.1 release.

v0.24.0 a first local run that reads well​

[change] phonton models setup MODEL prints a summary in a terminal: the runtime, the model, each calibration probe as pass or FAIL with its time, and the edit format Phonton will ask the model for. --json (or piping the output) still prints the full record. Download progress stays on one line.

[change] The local plan review is shorter: Goal, Repo, Edits, Checks as an exact command line, Model and Budget, then a few plain notes.

[change] When no context fits in memory, calibration says which memory is short and what to do. See the v0.24.0 release.

v0.23.1 receipts that agree​

[fix] A finished run's TUI receipt is the one place it shows tokens and cost. A second line under it repeated the cost and counted tokens from a different source, so the two could disagree. The frontier-pricing estimate now sits in the receipt, for priced models only.

[change] "No application launch command was inferred" is no longer listed as a known gap; libraries have none. See the v0.23.1 release.

v0.23.0 local models on Linux and macOS​

[new] phonton models setup installs, starts and verifies the local runtime on Linux and macOS, as it already did on Windows. No Ollama install and no --allow-unverified-runtime needed. Tested on fresh Ubuntu and Intel Mac runners (goals accepted) and an Apple Silicon runner (runtime verified).

[fix] Local edits that rewrite a whole function or class from its first line now replace that block instead of duplicating it. Acceptance on the local benchmark is unchanged; the edits now parse. See the v0.23.0 release.

v0.22.3 honest receipts​

[fix] The TUI no longer shows "saved N% vs naive" or a "best" streak. It opened every goal with "saved 100%" against a fixed 35,000-token guess. Priced models show estimated cost and the saving against frontier pricing; local and unpriced models show tokens only.

[change] Subtasks are named by their headline ("Implement method remove") in plan, goal, review, why-tokens and the TUI receipt, and the receipt drops repeated or internal lines. See the v0.22.3 release.

v0.22.2 half the cost, fewer false passes​

[change] Test-writing subtasks are planned only when a goal asks for tests; the project's own tests still run on every edit. Same 36/36 on the smoke suite at about half the cost.

[change] phonton goal asks before running the project's checks, or stops before any model call in scripts without --allow-host-checks. phonton models and phonton review print readable summaries.

[fix] Local goals that change a value ("from 120 to 80") edit the file that holds it. A 3B model used to change a different file and still pass the existing tests.

[fix] A local candidate that does not add a function or name the goal asks for is rejected, and the model is told why.

[new] PHONTON_LOG=debug writes worker attempts and verification rejections to phonton.log in the Phonton home.

[new] The local-model benchmark runs in CI on a CPU-only runner. qwen2.5-coder:3b: 15/36 accepted with 2 false passes, from 11/36 with 8. See Benchmarks and the v0.22.2 release.

v0.22.1 diffs that land​

[fix] Worker diffs with miscounted @@ ranges are recounted and moved to where their lines match, instead of failing and retrying. Same acceptance, about half the tokens on mid-file edits.

[fix] Files over 24 KB reach workers as excerpts around the names a goal mentions instead of signatures only.

[fix] A goal containing "release" (as in "release the stock") is no longer planned as broad multi-agent work.

[new] phonton models setup MODEL installs the runtime and model, calibrates it, and selects it in one command.

[change] Removed plan special cases tuned to old benchmark fixtures. Smoke suite on 0.22.1: 36/36 accepted. See Benchmarks and the v0.22.1 release.

v0.22.0 local models​

[new] Local models: phonton models setup, install, calibrate, select. With a calibrated model, the TUI and phonton goal run on your GPU, edits stay in a copy until you apply them, and nothing falls back to a cloud model. Managed runtime on Windows x64; on macOS and Linux, connect an Ollama you installed.

[change] phonton review reject restores the files the task changed. Workers may add tests; edits to existing tests count only if the originals still pass.

[change] Memory is scoped to the repository it came from.

[fix] Escalation uses model ids that exist; reasoning models get a 16k output cap; dropped connections retry; token totals no longer double-count.

[note] Install npm install -g phonton-cli, then phonton doctor. Full notes: v0.22.0 release.

v0.21.2 pixel marks​

[change] README and npm logos use the pixel CLI mark. TUI splash stays ANSI Shadow.

[note] Desktop 0.3.4 uses the pixel phi icon. Install npm install -g phonton-cli.

v0.21.1 cost receipt​

[new] CostReceipt on GlobalState and HandoffPacket: actual spend, labeled frontier estimate, route hops.

[new] phonton serve JSON-RPC (goal.status, review.get) includes the receipt fields Desktop reads.

[change] Missing provider keys fail closed. Escalation uses per-tier model ids.

[note] Install with npm install -g phonton-cli then phonton doctor --provider. Desktop 0.3.4 talks to this CLI over phonton serve.

v0.21.0 desktop serve​

[new] JSON-RPC sidecar for Phonton Desktop on port 47831.

v0.19.0 alpha slice​

[new] Added typed swarm planning metadata to plan preview and run evidence: PlanGraph, subtask assignments, conflict groups, swarm mode, and swarm reason.

[new] Added conflict-aware orchestration so overlapping expected touch scopes are serialized while isolated subtasks can still run concurrently.

[new] Added index backend selection with default local-hnsw and optional Qdrant HTTP code retrieval for larger codebases.

[new] Added phonton mcp capabilities <server-id> [--json] [--yes] for MCP initialize metadata, tool descriptors, and preview-only permission proposals.

[new] Added doctor output for the configured index backend, including Qdrant connectivity checks when Qdrant is selected.

[new] Updated website release docs to describe shipped v0.19.0 commands only.

[note] Qdrant is an optional external code retrieval backend. Phonton does not start containers and does not migrate SQLite decision memory into Qdrant.

[note] MCP capability preview does not call tools, grant permissions, or write config silently.

[note] Public benchmark and efficiency language remains conservative unless backed by complete reproducible artifacts.

Earlier public alpha​

[new] Published Phonton CLI to npm so users can run npx phonton-cli or install it globally.

[new] Added memory commands for listing, editing, deleting, pinning, and unpinning local decision memory.

[new] Improved plan previews so original goal terms are preserved more reliably.

[new] Expanded review actions for approve, reject, and rollback decisions.

[new] Added provider diagnostics with JSON output for automated setup checks.

[fix] Cleaned release packaging so npm includes only the README, licenses, package metadata, and installer scripts.

[fix] Kept public release publishing isolated from private workspace and website files.