Agent Run Replay for Small Teams

Opportunity thesis

A local-first coding-agent run replay layer can capture terminal actions, prompts, tool calls, file diffs, approvals, usage estimates, sandbox policy, and review checkpoints across Claude Code, Codex, Cursor, Copilot, and custom agents. Generic LLM observability tools prove trace demand, but the wedge is last-mile developer workflow provenance: what changed, why, at what cost, under what permissions, and what should be reviewed before merge.

Supply gap

Current supply is strong for instrumented LLM apps and generic agent traces, but weaker for local coding-agent operations: black-box CLI/session capture, Git worktree timeline, approval and sandbox policy history, token/cost attribution, provider comparison, and review packets that small teams can use before merge.

Small-team wedge

Start as a local wrapper and run importer for one repo: capture shell commands, agent transcript files, Git diffs, checkpoints, model/provider metadata, approvals, and usage estimates; render a chronological replay with risk flags and a shareable review packet. The wedge depends on developer workflow integration, not frontier model tuning.

Model tailwind

As coding agents become more capable and run longer, developers will delegate more work and need better provenance, replay, budget guardrails, and selective review. Better models increase the number and stakes of autonomous edits rather than removing the need for trust and audit trails.

Monetization

Charge small teams per active builder seat or per retained run history, with a free local tier, paid team sharing/search/retention, and a premium private/self-hosted option for sensitive repositories. Cost-control and review-time savings are the strongest initial willingness-to-pay hooks.

Distribution

Distribute through coding-agent communities, HN and GitHub issue threads about agent safety/review, templates for Claude Code/Codex/Cursor workflows, and open-source local collectors that produce useful review packets before requiring a hosted team workspace.

Validation plan

Interview ten developers or small teams that use Claude Code, Codex, Cursor, or Copilot agent mode at least weekly. Ask for three recent painful runs and reconstruct them from available transcripts, shell history, and Git diffs. Prototype a local replay packet for five runs and measure whether it reduces review time, catches missed risky edits, explains cost/usage surprises, or changes merge decisions. Separately test whether teams will pay for shared retention, search, and policy reports after the local free workflow proves value.

MVP brief

Build a local-first CLI plus small web UI. The CLI wraps or imports a coding-agent run, snapshots Git before/after state, records shell commands and approvals, parses known transcript formats when available, estimates usage from visible model metadata, and emits a structured run JSON. The UI shows a timeline, diff-linked steps, risk flags, budget/cost summary, sandbox settings, and a shareable Markdown review packet. No code leaves the machine unless the user explicitly exports or enables team sync.

Build prompt

Build a local-first coding-agent run replay prototype for one Git repo. Implement a CLI command that starts or imports a run, records timestamps, commands, exit codes, transcript paths, model/provider metadata, approval decisions, sandbox policy, Git diff hunks, and test results into structured JSON. Implement a web UI that renders the run as a chronological review timeline with file diffs, risky command flags, cost/usage estimates, checkpoints, and an exportable review packet. Include redaction controls for secrets and a no-network default.