[IMP] common: install the skills a run declares, resume the session it asked for, review its prompt

[sea-odoo]

Description

Five things a command could ask the AI plugin for and not get. They share mixins.py and one call path, so they land together, plus one isolated fix.

--resume never resumed anything, for three silent reasons stacked on top of each other:

• The latest session was looked up in ~/.claude/sessions.json — a file Claude Code does not write, and that ClaudeHandler.cleanup_junk here creates as an empty {"sessions": []}. The lookup was reading odev's own placeholder, so every run ended on "No previous session found to resume". Sessions live in ~/.claude/projects/<slug>/<session-id>.jsonl.
• The id was then passed as --session-id, which names a new session and refuses one that already exists. It is --resume.
• latest was resolved before the working directory was known, though sessions are per-directory. Resolving it after is what makes odev scaffold in a folder and odev ai --resume in that same folder the same conversation; failing that, the most recent session anywhere is resumed and said so.

Where an agent keeps its sessions is now the handler's question rather than an if-chain on the CLI name — agy resolves latest itself, and opencode keeps its sessions somewhere other than the Claude layout it inherits, where answering would have resumed the wrong agent's conversation. A bare --resume now means the latest session.

Skills were suggested, never installed. _ensure_skills printed a skills add line for the developer to copy and ran the agent anyway — without the method its prompt sends it to, which looks exactly like an agent that chose not to read a skill. It installs now, and verifies by listing again since the installer exits 0 even where it failed. The suggested command was wrong too: --skills is not an option the CLI knows, and an unknown option is ignored rather than refused, so it installed every skill in the repository.

Which skills a run needs was decided by if self._name == "test" inside the shared mixin; each command declares required_skills instead. Every command now gets odoo_coding_guidelines: each of them puts an agent in front of a client's checkout, and the rule that a dev reformats nothing it was not asked to is not specific to writing code.

The skill store was never refreshed. Only missing skills were installed, so a skill stayed at the revision it was installed at for good — a rule added to it reached the agents that had never loaded it and no one else. skills update compares the upstream folder hash against the recorded one and refetches on a mismatch, rate-limited by a new skills.interval (1 day).

--edit / -E opens the assembled prompt in $VISUAL/$EDITOR and hands the agent what comes back, the way a commit message is written: an emptied buffer aborts the run, and so does an editor exiting in error. No CLI odev drives takes a prompt without answering it, so reviewing one before it is sent had no other way of happening. The editor opens before the sandbox and the ephemeral cluster, so a run decided against leaves nothing to clean up.

Also in passing: the (Environment: ...) note was assembled after the command carrying the prompt was built, so it reached no agent at all.

sandbox_repository lets a command name the checkout the run is about, preferred over the directory odev was called from. odev-plugin-ai-scaffold on beta already sets it — the attribute lived only on an unpushed branch, so it was set, never read, and the agent worked wherever the command happened to be typed.

[FIX] sandbox is independent: the Chrome wrapper was produced by assigning to odev.home_path, a read-only property. That raised on every run, so the wrapper was never written and ODOO_BROWSER_BIN never set — visible only as a Failed to setup Chrome wrapper warning.

Testing

Exercised end to end against a live odev ai: bare --resume resolving to the last session of the working directory, the warning and global fallback when that directory has none, --edit aborting before any sandbox is created on an emptied buffer, the edited text reaching the agent (verified in the resulting session transcript), and a stale skill store being refreshed by an ordinary odev ai run while a fresh one is left alone.

Note on ruff-required

The hook does not pass on beta today. On the files this PR touches it goes from 53 findings to 43 — none introduced, ten removed (BLE001, PLC0415, S110, SIM102, S607, PLR0912). The remainder is pre-existing and deliberately left alone rather than reformatted as a side effect of this change.

Linked Issues

• none

Compliance

• I have read the contribution guide
• I made sure the documentation is up-to-date both in doctrings and the docs directory
• I have added or modified unit tests where necessary
• I have added new libraries to the requirements.txt file, if any
• I have incremented the version number according the versioning guide
• The PR contains my changes only and no other external commit

Assisted by Claude Opus 5