The workflow gap
Agent CLIs (Claude Code, Codex) increasingly write spec/plan markdown files (SPEC.md, PLAN.md, design docs) before touching code, and the human's job is to review those specs and give targeted feedback. Today the diff viewer is perfect for seeing what the agent wrote — but reacting to it means either typing a prose reply into the terminal ("in section 3, the retry logic should…") or bouncing out to GitHub to get line-anchored comments. On a phone, both are painful; the terminal route especially loses the line anchoring that makes spec review precise.
The ask
Inline commenting on files in the diff viewer (specs/markdown first, but any file works):
- Long-press a line / selection in the diff or file view → add a comment
- Collect comments into a small review batch
- On submit, Moshi injects them into the paired agent session as a structured message, e.g.:
Review feedback on .planning/SPEC.md:
L42-48 ("Retry policy"): use exponential backoff, not fixed 5s — see how we do it in worker.ts
L103: this section contradicts the auth decision in L67, pick one
Since Moshi already owns the session channel (send-keys / gateway), delivery needs no new infrastructure — the comments land exactly where the agent is already listening, with file+line context it can act on immediately. No GitHub round-trip, no PR needed for pre-commit specs that only exist in the working tree.
Why this fits Moshi
This would close the loop on the app's core story — "review and steer agents from your phone": approvals cover permission, the inbox covers awareness, and inline spec comments would cover direction. For agent-driven development the spec review IS the main human touchpoint, and right now it's the only part of the loop that still needs a desktop.
(Context: Pro user, several machines paired, using the diff viewer daily for Claude Code sessions in tmux. Filing here as moshi-hooks is archived — happy to move wherever you prefer.)
The workflow gap
Agent CLIs (Claude Code, Codex) increasingly write spec/plan markdown files (SPEC.md, PLAN.md, design docs) before touching code, and the human's job is to review those specs and give targeted feedback. Today the diff viewer is perfect for seeing what the agent wrote — but reacting to it means either typing a prose reply into the terminal ("in section 3, the retry logic should…") or bouncing out to GitHub to get line-anchored comments. On a phone, both are painful; the terminal route especially loses the line anchoring that makes spec review precise.
The ask
Inline commenting on files in the diff viewer (specs/markdown first, but any file works):
Since Moshi already owns the session channel (send-keys / gateway), delivery needs no new infrastructure — the comments land exactly where the agent is already listening, with file+line context it can act on immediately. No GitHub round-trip, no PR needed for pre-commit specs that only exist in the working tree.
Why this fits Moshi
This would close the loop on the app's core story — "review and steer agents from your phone": approvals cover permission, the inbox covers awareness, and inline spec comments would cover direction. For agent-driven development the spec review IS the main human touchpoint, and right now it's the only part of the loop that still needs a desktop.
(Context: Pro user, several machines paired, using the diff viewer daily for Claude Code sessions in tmux. Filing here as moshi-hooks is archived — happy to move wherever you prefer.)