taskctx — Task Context Packs for LLM Sessions

Every AI coding session starts blank. For a real task, the context is scattered across a GitHub issue, its comments, several pull requests, and feature branches in multiple repositories. Re-assembling that by hand for every session is slow and lossy — so I built taskctx to do it in one command.

# What it does

• taskctx collect <N> — fetches issue #N, its comments, PR states, and per-repo branch state (commits, merge-base, per-commit diffs), then renders a single context-pack.md.
• taskctx pack <N> — re-renders the pack from the local cache, fully offline.
• taskctx status <N> — prints only the delta since the last collect, writing nothing.
• taskctx init — creates the config and an instruction.md that tells any coding agent how to consume packs.

# What goes into a pack

• Delta first: what changed since the last collect — new commits per repo, PR state transitions (web#8636 state: open → merged), new issue comments.
• Issue snapshot: title, state, labels, body, and comments — plus automatically extracted open checkboxes (- [ ]) from the body and every comment.
• Repository map: each repo's branch for the issue (matched by configurable patterns like issue-{N}), commit list with file stats, and paths to full per-commit diffs on disk.
• PR states: number, title, state, merged flag, and review decision for every linked PR.

# Design decisions

• Cache everything: raw GitHub JSON, manifests, and per-commit diffs live in a .taskctx/<N>/ store. Re-rendering never touches the network, and --offline recomputes from local refs only.
• Deterministic output: the same store renders the same pack, so packs are diffable and safe to commit.
• Agent-native consumption: taskctx init writes an instruction file that points the agent to the pack, the diff directory, and the repo map — the agent reads files instead of running git commands itself.
• Multi-repo aware: repos can be local checkouts or remote-only; one repo can be refreshed while the rest come from the previous snapshot.

# Stack

# Commands

taskctx init                 # seed config + agent instruction.md
taskctx collect 123          # fetch + render .taskctx/123/context-pack.md
taskctx pack 123             # re-render from cache, no network
taskctx status 123           # delta only, to stdout
taskctx collect 123 --repo web --offline

The story behind it is in the blog article: Context Packs for AI Coding Agents: Building taskctx.