claude-tracker

A local table of the Claude Code sessions you care about, cross-referenced to Jira tickets. Replaces a notepad full of claude --resume <uuid> lines.

bun run server.ts     # http://127.0.0.1:4000

Everything configurable comes from the environment so no real value enters git:

export CT_JIRA_BASE="https://your-host.atlassian.net/browse"   # clickable tickets
export CT_TICKET_PREFIXES="ABC,XYZ"                            # defaults to NR
export CT_JIRA_EMAIL="you@example.com"                         # ticket titles and status
export CT_JIRA_TOKEN="…"                                       # id.atlassian.net API token

The credential pair can live in ~/.claude-tracker/credentials.json instead, as {"jiraEmail": "…", "jiraToken": "…"}, for when the server is started from a desktop launcher that inherits no shell. chmod 600 it; the tool warns if you have not. That file is in $HOME, never in the repo.

Without any of this the app still works. Ticket keys render as plain text, or as links with no title, and a banner says which.

What it does

• Reads ~/.claude/projects and never writes to it. You never type a UUID.
• Resolves each session's Jira ticket and title from the transcript itself, preferring the aiTitle Claude Code already writes.
• Tracks the sessions you pick: tickets with per-ticket done state, Jira-style priority, a name, and tags.
• Shows the canonical name Claude Code carries for a session (@api-worker) under the name you typed, when the session has one. Read from the transcript's last agent-name record, so a session renamed mid-run shows its current name.
• Lists the pull requests opened during a session in a second tab, read from the transcript's own pr-link records. No GitHub call, no token, no gh.
• Fills ticket chips with the Jira summary, status and assignee when credentials are set. Fetched after the page renders, so Jira being slow or down costs nothing: the page is exactly what it was before the feature existed.
• Resume opens a terminal running that session, falling back to the clipboard.
• Search covers every session on disk, not just the tracked ones, so deleting a row loses no history. It matches session ids, titles, ticket keys, and the body text of the transcript; a session that owns a ticket ranks above one that merely mentioned it, and body text ranks below both. Full-text hits show the matching phrase in context, which is how you find the sessions that never carried a ticket at all.
• Only what was said is indexed, not tool output or file dumps. That is 2% of the bytes on disk, and searching the other 98% finds the file rather than the session you were trying to remember.

Layout

File	Job
scan.ts	Find sessions, resolve real project paths from cwd
extract.ts	Transcript to ticket and name
annotations.ts	The working set, atomic writes, serialised mutations
jira.ts	Credentials, and the one network call in the tool
server.ts	Routes, ticket index, guarded resume
index.html	The UI
prs.html	The pull requests from one session
tokens.css	Colour tokens, shared so the AA pairs cannot drift
MANUAL-CHECKLIST.md	The UI test pass
TODOS.md	Deferred work, with reasons
VERSION / version.ts	Current version, semver parsing and bumping
release.ts	Cut a release: bump, changelog, commit, tag

State lives in ~/.claude-tracker/annotations.json and is safe to hand-edit.

Releasing

Write what changed under ## [Unreleased] in CHANGELOG.md as you go, then:

bun run release.ts minor --dry   # show what would happen
bun run release.ts minor         # bump, changelog, commit, tag

It refuses to run on a dirty tree, on failing tests, with an empty Unreleased section, or when the tag already exists. Each of those would record something untrue. Semver here versions the annotations.json format and the HTTP surface: major means a stored file stops loading, minor means new capability with old files still working, patch means nothing observable changed.

bun test    # 210 tests, no dependencies