This page describes honestweek as of October 9, 2026. Version 0.3.0 came out on npm on October 8. It lets an agent ask honestweek about your past sessions, and can keep what it found after Claude Code deletes old logs, as What 0.3.0 added lists. Version 0.2.0, its first release on npm, came out on October 6 and added the browser view described below, with the Problems checks running on Codex as well as Claude Code.
Status: Released · MIT licensed · View on GitHub
honestweek reads the session logs Claude Code and Codex already keep on your machine. Its browser view opens on a Problems page: the known ways AI coding agents go wrong that turned up in your week, worst first, each linked to the step in a replay where it happened and paired with a fix you can copy. It also finds the sessions behind a pull request, a commit or a file, and turns a finished week into a summary that won’t build until every commit it cites checks out against git. It’s on npm, so npx honestweek view --demo shows the browser view on a made-up week.
What it is
An agent can say “All tests pass” right after a test run failed, and in a three-hour session with seven sub-agents you might never scroll back far enough to see it. I built honestweek to catch that, and to let you see exactly what happened before you change anything.
The rule under all of it is that honestweek states nothing without its evidence. Every link, count and step in the view carries one of four labels:
| Label | Meaning |
|---|---|
| Recorded | A log line or git says so |
| Derived | Worked out from recorded facts |
| Inferred | A named rule’s best reading, and the rule is named |
| Missing | The log doesn’t say |
When the records fit a link more than one way, the view also marks it ambiguous.
It needs Node 18 or later and git, and it has no runtime dependencies [measured 2026-10-09, the 0.3.0 package manifest].
Pages in the view
honestweek view serves these pages from your own machine and opens your browser.
| Page | What it shows |
|---|---|
| Problems | The page it opens on. The known ways agents go wrong that showed up in your sessions, starting with claims the agent couldn’t back, like saying “done” with no check after its last edit. Each finding links to its step in the replay. Each problem shows its count against the window before and offers a fix to copy, such as a hook, an instruction line or a setting. |
| Find | The sessions and goals behind a pull request, a commit, a file, a branch or a few words. A wider search covers the prompts and titles of every session on the machine in the same dates. |
| Goals | Each goal in your goal list, with the sessions that worked toward it on one timeline you can play and scrub. |
| Replay | Any session step by step: a row for the main agent and a row for its sub-agents that opens into one for each, with the log line, the command output and the timing behind every step. |
| Setup and Settings | When view finds no config, it opens Setup, which finds your repositories and asks what to keep private. Settings changes any of it later. |
Problems
The catalog lists 42 ways AI coding agents go wrong or waste time and tokens, and honestweek checks Claude Code sessions for 21 of them. Of those 21, 17 run on Codex logs too. The other 4 read Codex logs but are marked partial, because they haven’t been seen to fire on a real Codex week [measured 2026-10-09, the problems catalog in 0.3.0]. Each problem links the published sources behind it. A fix is never applied for you.
An optional switch, Include /insights, adds what Claude Code’s own /insights command wrote about the same sessions. With it on, a Run /insights button runs that command for you, and a Run with Codex button has your own Codex judge your Codex sessions the same way. The /insights and Codex results each show as their own AI-written group, never counted with honestweek’s findings. Both buttons ask first, because they send your sessions off your machine, as the privacy section below says.
Commands
npx honestweek view --demo shows a made-up week and sets nothing up [measured 2026-10-09, a cold run of 0.3.0 from npm]. npx honestweek view sets up and opens your own last seven days. npm install -g honestweek gives you a plain honestweek command, and installing it as a Claude Code plugin adds /honestweek:honestweek [measured 2026-10-07, a fresh install recorded in honestweek’s 0.3.0 README], which runs the weekly summary inside Claude Code:
/plugin marketplace add BryceEWatson/honestweek
/plugin install honestweek@honestweekThere are sixteen subcommands [measured 2026-10-09, honestweek --help in 0.3.0, run cold from npm]. The last five in the table came in 0.3.0, and What 0.3.0 added says what they’re for:
| Command | What it does |
|---|---|
view | The browser view above |
init | Writes honestweek.config.json from your git setup, and asks before it writes |
discover | Reads the last completed week’s sessions into a redacted draft |
validate | Checks the weekly items: a status and a receipt on each, and no private word in the text |
build | Checks every cited commit against git, then writes the output, or stops and writes nothing |
preview | Serves the built output on a page only your machine can open |
digest | Picks the week’s prompts, ideas, decisions and next steps for the weekly page |
prompts | Keeps your private inbox of prompts |
harvest | Suggests words to keep private |
mine | Finds problems you worked out in someone else’s tool, ranks them, and can draft a post |
history | Lists what landed in a period, for a client report |
status | Says where the weekly summary stands and what to run next, and changes nothing |
find | Finds the sessions and goals behind a pull request, a commit, a file, a branch or some words |
replay | Lists one session’s steps in order, or its state at one moment |
problems | Lists where your sessions went wrong, highest priority first |
goals | Lists your goals and the sessions that did their work |
Weekly summary
Your commits show what shipped, and your sessions show what you figured out. The weekly summary uses both. discover reads the week, the /honestweek skill distils it into items, and each item gets a status (shipped, in progress, or designed, not proven) and a receipt, the commit or session turn it came from. build then checks every cited commit against your real git history. It stops with exit code 2 rather than write a claim it can’t back, and it downgrades a shipped item to in progress when its commits haven’t reached the default branch. The output can be a post, a changelog, a digest, a report, a standalone page, a site’s data files, or a printable client report. You review it and publish it yourself.
What 0.3.0 added
Version 0.3.0 reached npm on October 8 [measured 2026-10-09, the npm registry], so npx honestweek and a new install get these now. An install from before October 8 keeps its old version until you update it: npm update -g honestweek for a global install, /plugin marketplace update for the Claude Code plugin, or git pull in the plain skill’s folder.
| Capability | What it does |
|---|---|
| Ask from any chat | find, replay, problems and goals answer the view’s questions as text or JSON, so an agent can run them. A read-only find skill lets Claude Code or Codex run them on its own when you ask which session made a pull request, a commit or a file change, or where your sessions went wrong. |
| Ask about one problem | problems --pattern answers four questions about one problem: its cause, its fix, how sure honestweek is, and when it happened, with a link to each finding’s step in the replay. It keeps honestweek’s general description apart from what your log shows, and says plainly that how often a check is right hasn’t been measured. problems --session says whether one session was checked. |
| Summary status | honestweek status says where the weekly summary stands and which step is next. The skill reads it first, so it picks up where you left off. |
| Summary on request | Claude starts the weekly summary when you ask for one in words (“write up my week”), and a scheduled task can run it |
| Saved results | An optional setting, Save results between runs, off by default, keeps each day’s Problems results and history beside your config, redacted. A session in one of your repositories whose log Claude Code has deleted still shows, and later runs skip most logs that haven’t changed. |
| Links to a step | view --page opens the page on one session or one step, so an agent can hand you a link to what it’s describing |
| One config for every folder | A config saved for every folder lets an agent working in another project find it |
| A summary step with file tools only | With the plugin, the summary’s one model step runs with file tools only, so a line planted in your logs can’t make it run a command or open a page. It can still write files, though it’s told to write only the summary’s items. |
Privacy
| What | How it’s handled |
|---|---|
| Your logs | Read, never changed |
| Network | No account or telemetry, and honestweek itself makes no network call. Session text reaches an AI only through the things you start or ask for, listed in the next table. The view’s server listens on 127.0.0.1 only and answers only its own page. |
| Keys, tokens and passwords | Hidden on every page and in everything it writes, when it recognizes them, even with the Show private text switch on |
| Names and client words | Hidden once you list them in Setup or Settings. Show private text shows them on your own screen only. |
| Display-only repositories | Shown in the view, but git is never run against them |
| What the view writes | Only what you save: the config, an example config, .gitignore lines, Run with Codex’s redacted answers, and a short-lived file that opens your browser. Turning on Save results between runs adds those results, redacted and git-ignored. |
What reaches an AI
Each of these runs only when you start it or ask for it, and goes to Claude or OpenAI on your own plan. The two buttons also need Include /insights turned on, and each asks before it runs.
| What you start | Who reads it | What it reads |
|---|---|---|
/honestweek in Claude Code | Claude | A redacted draft of your week, to write the summary, and what each command prints |
| The same skill in Codex | OpenAI’s model, through your Codex | The same draft and command output |
find, replay, problems or goals, run by an agent or the find skill | Claude, or OpenAI’s model through Codex | The redacted answers to your questions about past sessions |
| Run /insights | Claude | Your Claude Code sessions, read by Claude Code’s own /insights, so honestweek doesn’t choose or redact them |
| Run with Codex | OpenAI’s model, through your Codex | A redacted copy of each Codex session it judges, in Codex’s read-only sandbox, which can still read any file you can |
How this site uses it
The weekly work log, its archive and the goals page are built by a site script that runs honestweek’s site mode and writes this site’s data files. honestweek checks every commit an item cites against git, and the build stops rather than publish an item it can’t verify. The counts on the home page come from the same output. The site uses the weekly summary only, not the browser view.
What doesn’t work yet
| Limit | Detail |
|---|---|
| Private words | Names and client words show as written until you list them, and the view says so on every page until you do |
| Check accuracy | How often each check is right hasn’t been measured yet, so a finding carries an evidence label, not a confidence number. Measuring it, and whether a copied fix held, is proposed in honestweek issue 148. |
| Codex coverage | Four checks read Codex logs but haven’t been seen to fire on a real Codex week, so they’re marked partial. honestweek doesn’t read Codex’s web search records yet, so that fact reads “not recorded” for Codex sessions. |
| Permission prompts | Neither agent’s log says whether a command would have asked for permission, so that fact reads “not recorded” for every session |
| Weekly summary and Codex | The summary’s draft reads Claude Code sessions and handoffs, not Codex sessions. Since 0.3.0 the skill can also run in Codex, but a whole weekly summary hasn’t been run through Codex yet. |
| Fix tests | Each fix comes with a prompt meant to trigger its problem so you can watch the fix catch it. Those prompts hadn’t been run against Claude Code or Codex when 0.2.0 shipped, and the 0.3.0 changelog and release notes don’t say they have. |
| Goal discovery | The Goals page needs a goal list you write. Finding goals for you isn’t built. |
| Long windows | Past a week, loading stops at 500 MB of logs, a limit you can raise in Settings. It doesn’t load the older days in pieces yet. |
| Edits outside a folder | The check counts only what the log records as an edit, so a file written by a shell command isn’t counted |
| Digest and goals page | The weekly digest and a standalone site’s goals page don’t work together yet |