You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
CLI. Three days after first use, a successful command prints one dim line asking for a GitHub star, and coding.sh ui shows a small card after 5 sessions. The line prints once; the card shows until you dismiss it or click Star. Neither makes a network call. CODING_SH_NO_STAR_PROMPT=1 turns both off. Install
Checks.npm run gate runs every pre-commit check in one pass under a throwaway home: repo hygiene, branch freshness, lint, actionlint, generated files, build and tests, store plugins, the Python SDK, the site, a secret scan and adapter versions. A failing step prints the command that reruns it. Contributing
Lint. New rules: vendor-literal, site-voice, universal-transition, reduced-motion, global-install and personal-data. slop.mjs --strict-allowlist fails on allowlist entries that no longer match anything. Contributing
Doctor.coding.sh doctor counts recent model calls past a physical limit (speed, time to first token, duration) and names the session with the most. Troubleshooting
SDK.coding_sh.weights_edit() in the Python SDK records each weight edit a script makes as one notification event with payload.kind: "weights.edit": tensor name, shape, weights changed, L2 and max-abs of the change, and the base and edited sha256. It adds no event kind, and coding_sh.weights_delta() computes the counts and norms for small arrays without numpy. Docs
SDK.weightsEdit() in the TypeScript SDK records a weight edit as the same notification event the Python SDK's weights_edit() sends, and returns true only when the event was queued. weightsDelta() computes the shape, count and norms for nested arrays and typed arrays. Both SDKs' tests check one fixture of the events an edit must send. Docs
Sessions. A session that recorded weight edits gets stats.weightEdits (edits, distinct tensors, weights changed) and the weights:edited tag, and is listed even when it holds nothing else. coding.sh show, the viewer and MCP session_get print one line per edit, and --where 'weights.changed>1000' filters by size. Stored sessions are folded again. Docs
OpenTelemetry. A span or log record named weights.edit with coding_sh.weights.* attributes is read as a weight edit, and the otlp sink sends one weights.edit span per edit (so does [serve.forward], for edits posted to /v1/ingest). The invoke_agent span carries coding_sh.stats.weight_edits and coding_sh.stats.weights_changed. Metadata mode keeps only the counts and norms. Docs
Evals.coding.sh exp run --weights <SHA256> records which checkpoint a run's task ran (the SHA-256 of the weights file, which you compute). exp compare, exp show, exp list, the viewer and the MCP tools show the stamp, and a comparison warns when both sides ran the same weights or only one recorded them. Compare an edited model with its base
SDK.coding_sh.weights_checkpoint() and weightsCheckpoint() record the file a script saved its edited weights to: one notification event with payload.kind: "weights.checkpoint", the file's SHA-256 (the value exp run --weights takes) and optionally the base file's. Both SDKs' tests check one fixture. Docs
Sessions. A session that recorded a weights checkpoint lists the file hashes in stats.weightEdits.checkpoints and gets the weights:checkpoint tag. coding.sh show, the viewer and MCP session_get print checkpoint 58899f90572c ← 0496e5111097, and --where 'weights.checkpoint:58899f90572c*' finds the session. Metadata capture keeps both hashes. Stored sessions are folded again. Docs
OpenTelemetry. A span or log record named weights.checkpoint with coding_sh.weights.sha256 is read as a checkpoint. The otlp sink and [serve.forward] send one weights.checkpoint span per event. Metadata mode keeps the two file hashes and drops the note. Docs
Evals.exp show and exp compare name the session whose checkpoint recorded an experiment's --weights hash, with its edits. --json and the MCP tools carry the ids as weights.sessions. coding.sh show <session> and MCP session_get list the experiments that ran its checkpoints. Docs
Viewer. A weights stamp on the experiment and compare pages links the session that recorded that file as a checkpoint, and the session's header links each experiment that ran it. /api/compare, /api/experiments/{id} and /api/sessions/{id} carry the same ids as the CLI's --json. Docs
Evals.coding.sh exp compare A B --given <grader> compares only the cases where that grader passes in both runs, and prints how many it kept and dropped. With a terminated grader this separates "finished more answers" from "got more of the finished answers right". Compare an edited model with its base
MCP.coding.sh mcp serves four Agent Skills over the MCP skills extension (io.modelcontextprotocol/skills): improve-agent-setup, trim-agent-setup, review-agent-session and build-agent-evals. A host that supports the extension lists them and loads one when a request matches. Skills
MCP. In 2026-07-28 requests, list, read and discover results carry ttlMs and cacheScope: the tool, prompt and template lists are cached for an hour, server/discover for five minutes, and sessions are not cached. MCP server
Site. The front page is also in German, Spanish, French, Japanese, Korean, Brazilian Portuguese and Simplified Chinese, at /de, /es, /fr, /ja, /ko, /pt-br and /zh. Each page names the others with hreflang links and the sitemap lists them as alternates. A switcher in the footer moves between them. The docs, blog and changelog stay in English. 简体中文
Changed
Site. The terse-mode guide is now a blog post that leads with what we measured on one plugin (output −8%, cost +11%, same pass rate and clarity), then shows how to run the same test on your tasks. /docs/terse-mode answers with a permanent redirect to it. Does a terse mode save money?
Docs. The docs home, the home page FAQ and llms.txt say what coding.sh does: it records prompts, model calls, tool calls, file changes and cost, shows each run in a local viewer, and turns failed runs into evals. Other products are named only on the compare pages and where they are an integration, and a site test enforces it. Docs
Docs.--help, the package descriptions, the README, the site's description and llms.txt say what coding.sh records instead of "open-source, local-first observability and evals for any AI agent", and the site's status line shows the current release (it said v0.1.4). A site test fails on stacked category words in published text. Docs
Node. The minimum Node is 22.16, because 22.15 has no SQLite FTS5. A store that lacks FTS5 now fails with reason no_fts5 and names Node 22.16. CI runs the packed CLI on Node 22.16.0. Install
CI. New jobs: the static gate, the Postgres store, the Python SDK on 3.9 and 3.13, and the release smoke test. The site job fails instead of skipping when its build is missing. Contributing
Stats.stats rolls each subagent's cost, tokens and tool calls into its top-level session. Concepts
Experiments.exp compare --gate says "Don't ship" only when B's losses beat chance, by an exact one-sided binomial test on the cases that changed outcome. Losses within that noise hold. Each group under --by keeps the old rule, as refine's guard set does. Evals
CLI. A JSON error's next is always a label and a command. A suggestion that needs the user's approval is labelled "With their approval:". Next-step hints shell-quote the names you typed and print full session ids. API
Evals.judge grade exits 2 with reason grader_error when the grader could not run, so a verifier can tell it apart from a fail verdict (exit 1). Troubleshooting
Docs. The docs are regrouped by task, with a quickstart for your agent or a coding agent, a first eval that ends with a fix, TypeScript and Python tabs, viewer screenshots, and Copy page and View as Markdown on every page. New pages cover cost, the query language, LLM judges, eval methods, deploying the server, privacy and the roadmap. Docs
Docs. The README has a quickstart in three steps, a table of what you get with the detail folded under each area, and an agents table with what each status means. The CLI README lists every command, including statusline, tree, replay, blame and sql, and each package README links its docs. See the README.
Site. Docs pages no longer jump when the web fonts load on a phone: layout shift on /docs/cli fell from 0.208 to 0.0003 and on /docs/quickstart from 0.009 to 0.0002. Desktop screens get home screenshots sized to their slot, 1.6 kB less. Site speed
Site. The edge answers moved URLs with permanent redirects, keeps a page prefetched on hover for the click, caches hashed images and search files for a year, compresses the changelog feed, and names the page as canonical on each Markdown copy. Site speed
Agent text. llms.txt lists the blog and links every CLI command. llms-full.txt leaves out the per-version changelog pages and the per-command CLI pages (647 kB, was 1,046 kB), and each Markdown copy's updated: is its page's last change. llms.txt
Site. The front page is new: the headline and install command, a moving row of the agents it works with, a diagram of agents feeding one record, three steps to start with a drawing of one session, and four short answers. Its footer lists the docs by section. coding.sh
Site. The card shown when a link to coding.sh is shared has the front page's headline. coding.sh
Site. The line under the front page's headline reads "Record, measure, compare and benchmark every agent run." Under the install command, a line links to the notes for the latest release. coding.sh
Fixed
Site. The install script is served with X-Robots-Tag: noindex at both of its paths, /install and /install.sh. The two paths return the same file, and neither named a canonical. Install
Sessions.coding.sh show, the viewer and MCP session_get print a weight edit's or a checkpoint's note at the end of its line. Before, the note was stored but never shown. Docs
Sessions. A session with no prompt, such as an SDK trace() that only records weight edits, is titled by its trace's name in coding.sh sessions and the viewer. Before, its title was empty. Stored sessions are folded again. Docs
Evals.exp metrics prints a count or score below 0.0005 with two significant digits (0.0001) instead of 0.000, so a small value such as a gain per edited weight no longer reads as zero. Metric A/B
SDK.coding_sh.weights_edit() reads a 0-d tensor or array as the number it holds, so l2=delta.norm() records the edit. Before, a torch scalar was refused and the edit was dropped with a warning, though the docs said torch scalars count as numbers. Docs
SDK.coding_sh.weights_edit() warns about a dropped edit once per process, as its docs say and as the TypeScript SDK does. Before, every init() warned again. Docs
MCP. A request cancelled with notifications/cancelled gets no reply (a cancel for initialize is ignored), a cursor on tools/list, prompts/list or resources/templates/list is an error, a request with a null id gets an error instead of no answer (a null-id notifications/ message is still a notification), and a 2025-03-26 session accepts batches. MCP server
MCP. Tool answers stay under 25,000 characters: a longer page has fewer rows and limitUsed, and a tool with a cursor also returns nextCursor. session_replay pages its files and cuts text and lists only when an answer would not fit, saying so with truncated (a range was 5 million characters). MCP server
MCP.eval_init and score_add are marked destructive, since both can replace stored data, and score_list's error names its since argument instead of the CLI's --since. MCP server
Site. The changelog and quickstart pages no longer shift on a phone while loading: release area chips sit in a fixed grid and the quickstart tab list is laid out from first paint. Layout shift fell from 0.052 to 0 and from 0.023 to 0. Site speed
Stores. A file:///C:/... blob URL and a sqlite://user:password@/path index URL open the path they name on Windows, and the user and password are no longer read as a folder. See stores.
Capture. Model calls no longer show impossible speeds such as 3750 tok/s. A hook reading a transcript from mid-response, or late Codex lines, gave a call a start time inside the response; such a call now has no duration, time to first token or speed. A Claude Code call gets them back when the import reads the whole transcript. Adapters
Tests. The viewer's browser tests no longer hang when Chrome drops its DevTools connection: a waiting call fails at once, and any call fails after 30 s. The helper lets go of Chrome's pipe after closing it, and the CI runner ends each test file once its tests finish, so an open handle cannot hold a Windows job. Contributing
Capture. A blob store outage no longer drops spool events while the cursor moves on, and a full disk no longer leaves a torn spool line or loses the whole line hook.sh appends after one. Ingest redacts short payloads inside the spool file once their day has passed. A session run with CODING_SH_DISABLE=1 no longer comes back through backfill. Privacy
Doctor.doctor reports a missing hook.sh, a spool file hooks cannot write and recent failed hook writes. doctor and status mask the password in a store URL. init exits 0 when only an agent's setup failed, and init --dry-run lists the bin files it would write. Troubleshooting
Init.init registers no MCP server for an agent whose hooks it could not install, so a failed agent is left as it was. Install
Stores.error: matches an error event's message, a quoted * is a wildcard and \* a literal star, and the Postgres and ClickHouse stores compile --where as SQLite does. Concurrent dataset add runs merge instead of losing a version. An empty dataset import exits 1. Query
Stores. The index, its WAL files, blob packs and backups are written owner-only (0600), and new store folders 0700. On the SQLite store, a search of only emoji or symbols finds matches. upgrade-store on a plugin store says there is nothing to upgrade. Privacy
Stores.tree, cache doctor and diff-calls on a --store file that does not exist stop with store_missing instead of creating an empty index. Troubleshooting
Sessions. A prefix of a parent session opens the parent, not one of its subagents. A bare agent id or an SDK id without its project opens that session. latest skips empty sessions and resolves an agent run under coding.sh run to its wrapper. Terminal tables keep session ids whole. CLI
CLI. Impossible dates such as 2026-11-31 are refused, diff prints call times in local time, show prints the end of a failed tool's full output, and sessions counts what it hid. config get and config set know every documented key, and config set prints the old value. CLI
Cache.cache doctor keeps its priced total and counts breaks on an unpriced model apart. CLI
Evals.toolevals drivers honor ANTHROPIC_BASE_URL and OPENAI_BASE_URL, judge calibrate counts labels on case targets, eval export and eval import work across two homes, and eval init --cases works from a subdirectory. Task sandboxes no longer inherit type: module from a parent package.json. Evals
Scores.score add keeps an earlier comment when you leave it out and the value is unchanged. It refuses a run that does not exist and experiment targets, each with a next step. Evals
Experiments.exp ab gives each arm only the files the session's agent loads, and an edit that agent does not load stops with "No arm difference" before any model call. exp show latest works, and an HTTP 529 reads "overloaded". Evals
Bench. A pair with over 10% of cases excluded says "no verdict" and why, and an interrupted run names the exact --resume command. Benchmarks
Improve. Fix lines go to the file the flagged sessions' agent reads and name the test command it uses. Improve
Workflows. A run folder copied under a second session is imported once, and agents silent for 15 minutes show as stalled. Workflows
Serve.serve honors the [proxy] section of --config, the collector refuses a host other than loopback, and the proxy rejects bad upstream URLs. Production
SDK.require() of the TypeScript SDK works, and both SDKs warn once on a bad collector URL. SDK
Viewer. The live header refreshes in place, empty fields no longer show "null", server-rendered rows match the page's, and an unknown session id costs one request. Explore
Viewer. Enter presses the focused button, filter chip or link, and Space the focused button, also inside an open list row. Before, they opened the selected session, folded the row or started the replay. The arrow keys scroll the focused step detail, and a redraw moves focus to the rebuilt control or its pane instead of dropping it. Explore
Packaging. The published package.json no longer has source export conditions that point at files the tarball does not hold. smoke.sh runs under a throwaway home, and the GitHub Action has an annotate input. Release
Site. The docs sidebar and the phone menu open on the current page, and the phone menu fits the screen. A terminal block no longer touches the paragraph above it, wide code blocks scroll by keyboard, code copy selects the code when the browser blocks the clipboard, and search results open without a redirect. Docs
Site. Changelog notes link docs to their page on the site instead of showing raw Markdown or going to GitHub, and the changelog counts only shipped releases. Changelog
Site. Search snippets end on a whole word, structured data resolves on every page, sitemap dates follow each page's last change, and pages no longer carry hidden text for agents. The site build now fails on each of these. Site
Site. The skip link opens below the header, one Tab press crosses the agent logo row, the footer links sit in rows of four and three on a phone, and a long word wraps instead of widening the page. Site