Buddy review

Buddy Review Control

Translate the user's request into one allowlisted action. With no explicit action, use toggle for review mode because /buddy-review is intentionally a one-step toggle.

Review-mode actions are toggle, enable, disable, or status:

1. Walk upward from this file until you find one directory containing both .codex-plugin/plugin.json and scripts/buddy-review.mjs; use that directory as the plugin root. Stop if that exact root cannot be found.

2. Resolve the current Git repository root. If the current task is outside Git, explain that automatic code review needs a Git workspace and stop.

3. Run exactly this command, passing every value as a separately quoted argument:

   node "<plugin-root>/scripts/buddy-review.mjs" mode <action> --cwd "<repo-root>"
   

   For the no-action /buddy-review toggle only, append --continuous-review. This keeps the command one step while making the user's invocation the explicit authorization for bounded intermediate evidence if the resulting state is ON. A toggle to OFF clears that authorization.

   When the user explicitly requests reviewer configuration, append only the documented flags under Supported optional configuration as separate quoted arguments before running the command.

4. Report the resulting ON/OFF state, ordered primary and optional secondary provider/model connections, workspace scope, and continuous-review state. Explain that enabling continuous review authorizes privacy-filtered intermediate evidence for at most two speculative generations plus one exact final fallback per configured reviewer during a turn. Two lanes run independently and concurrently for a generation, one success produces an attributed partial result, and Buddy never retries or falls back to an unconfigured provider. When the resulting state is ON, also state that enable does not capture a baseline: the next Codex turn establishes the private start snapshot at prompt submit, review covers only changes after that baseline, and enabling mid-turn does not invent a missing baseline.

5. For the animated companion, tell the user to run the first-party /pet command once and choose a built-in or installed Buddy pet in Codex Settings. The host persists its open state, selection, position, and animation surface.

Pet actions are pets/pet list, pet status, pet install <buddy-pet-id>, pet update <buddy-pet-id>, pet remove <buddy-pet-id>, pet restore <backup-id>, or pet reconcile. Run exactly:

node "<plugin-root>/scripts/buddy-review.mjs" pet <action> [identifier]

For install/update/remove, only accept pet IDs printed by pet list. For restore, only accept a backup ID printed by a prior pet remove result or pet status; never synthesize a path or identifier. Every pet mutation and reconciliation must be an explicit user request. Report the command result and the native-host step honestly: Settings → Pets → Refresh, select the appearance, then enter /pet once. Buddy cannot select or wake it programmatically.

Presentation actions are presentation status or presentation set. set requires at least one closed value:

node "<plugin-root>/scripts/buddy-review.mjs" presentation set \
  [--pet-id <native:selected|id-printed-by-pet-list>] \
  [--personality <precise|warm|wry>] \
  --cwd "<repo-root>"

Presentation is local and cosmetic. It does not select the native host pet, change review authorization, increment review-mode revision, alter findings, or cause provider egress. Mood is derived from review state. XP grants exactly 10 points per unique durable completed review, regardless of findings, no-findings, or defensible abstention; failures earn no completion credit.

The worker-summary claim advisory is disabled by default and has a distinct egress consent. Status or disable may be run directly. Enable only after the user explicitly asks to send the bounded final worker summary to the current primary provider, and always include the confirmation flag:

node "<plugin-root>/scripts/buddy-review.mjs" summary-guard status --cwd "<repo-root>"
node "<plugin-root>/scripts/buddy-review.mjs" summary-guard enable \
  --cwd "<repo-root>" --confirm-summary-egress
node "<plugin-root>/scripts/buddy-review.mjs" summary-guard disable --cwd "<repo-root>"

The advisory shares the primary review call, is separately labeled, and its closed fields can never be promoted into a code finding. A secondary reviewer always receives technical evidence only. Changing the primary provider/model makes prior summary consent stale and fail closed; changing only the secondary does not widen summary egress. Because the summary does not exist during implementation, enabling this guard disables speculative background review for that turn and uses the exact final Stop review. Do not claim that the optional packet cannot influence the primary model's technical output; the guarantee is structural separation plus independent validation. Do not describe it as hidden chain-of-thought, reasoning extraction, or transcript access.

Diagnostics are read-only and make no provider call by default:

node "<plugin-root>/scripts/buddy-review.mjs" doctor --cwd "<repo-root>"

Only add --provider-check after explicit user authorization for one bounded network/model health call per configured reviewer, with a maximum of two calls. The checks run concurrently, send no repository evidence, and return exact per-role pass/fail plus aggregate pass, warn, or fail status. A partial result is warn; no connection is substituted. This tests connectivity and strict one-field output handling, not review quality. Doctor cannot prove host hook trust, command-menu discovery, native pet selection, or visual Running/Ready animation.

Guided setup is a deliberate plan/apply workflow. A fresh or disabled workspace plans final-only review unless the user explicitly requests --continuous-review. With no continuous-review flag, an already-enabled workspace preserves its current continuous or final-only choice:

node "<plugin-root>/scripts/buddy-review.mjs" setup plan --cwd "<repo-root>" [closed options]
node "<plugin-root>/scripts/buddy-review.mjs" setup apply --cwd "<repo-root>" \
  --plan-id "<exact-id>" --plan-digest "<exact-sha256>"
node "<plugin-root>/scripts/buddy-review.mjs" setup rollback --cwd "<repo-root>" \
  --plan-id "<exact-id>" --plan-digest "<exact-sha256>"

Always show the plan before apply. Apply and rollback require explicit user requests and the exact generated ID/digest. Never invent either value. Setup installs or updates a hash-pinned pet first and enables review last; rollback restores mode first and the pet second. Its immutable receipts resume recognized process-crash states, including the update-rollback midpoint, while refusing unexpected hashes, ownership, backups, or revisions. It never modifies global AGENTS.md, trusts hooks, selects a host pet, or wakes /pet.

Workspace data actions are data status and data purge. Status is read-only:

node "<plugin-root>/scripts/buddy-review.mjs" data status --cwd "<repo-root>"

Purge is an explicit destructive action. Run it only when the user asks to purge Buddy-owned local data for the exact current Git workspace. First disable automatic review for that workspace, then run:

node "<plugin-root>/scripts/buddy-review.mjs" data purge \
  --cwd "<repo-root>" --confirm-purge

Pass --include-settings only when the user explicitly asks to remove the workspace's Buddy mode, connection, presentation, summary-consent, and circuit settings too. Purge refuses while review mode is enabled or an egress capability is live. It must never delete provider CLI authentication, arbitrary plugin data, another workspace's data, or any path supplied by the user.

Headless renderer actions are renderer register, next, ack, status, unregister, and prune. Pass only documented closed flags from renderer --help. Register, unregister, and ack change local renderer state. Bare renderer prune is read-only; only renderer prune --apply deletes. New events never persist worker summaries. --include-worker-summary can expose only a still-retained legacy v1 summary and requires an explicit user request. Apply deletes eligible v1 or v2 events only through the minimum acknowledged sequence of every active consumer and after the effective minimum age. No active consumer means nothing is eligible. The requested minimum age can shorten retention but cannot extend the 24-hour privacy ceiling. Run the dry run first unless the user explicitly asks for immediate apply. Lifecycle pruning hard-expires aged v1 and v2 content even without a registered renderer and compacts the legacy v1 index under the outbox lock.

Supported optional configuration

Only pass these flags when the user explicitly requests them:

• --provider ollama --model glm-5.2:cloud
• --provider claude --model claude-opus-4-8
• --provider opencode --model openai/gpt-5.6
• --also-provider <ollama|claude|opencode> with optional --also-model <id> and --also-effort <level>
• --single-reviewer to clear the configured secondary connection
• --continuous-review to authorize bounded intermediate evidence and enable at most two speculative generations per turn; the no-action toggle is the documented exception that always appends this positive flag
• --no-continuous-review to use final-only Stop review
• --effort <level>
• --confidence <0..1>
• --max-patch-bytes <integer>
• --timeout-seconds <1..1800>

The currently used adapter IDs for this release lane are claude, ollama, and opencode. Claude Pro or Max uses only the direct Claude adapter with Claude Code. Ollama local and Ollama Cloud can use the direct Ollama adapter, while an exact configured OpenCode Ollama Cloud route is also allowed. ChatGPT Plus or Pro uses opencode with an OpenCode ChatGPT OAuth connection. Kimi or Moonshot is allowed only through opencode with the exact API-backed provider/model identifier already configured in the user's OpenCode connection. These routed connections are not native Buddy adapters. Direct codex and direct kimi adapters are unsupported because strict no-tools and no-inherited-context isolation has not been proven for those subscription routes. OpenCode projects only the selected provider auth entry; never copy, print, or request the user's auth store or ambient credentials.

Claude and OpenCode accept effort values low, medium, high, xhigh, and max. Ollama accepts only low, medium, and high. Reject an unsupported provider/effort combination rather than changing it or contacting the model.

Never interpolate unvalidated free-form text into the command. Provider and model values must come from the closed adapter choices above or an explicit user-supplied OpenCode provider/model token accepted by the CLI. Reject unknown actions or flags. --also-provider, --also-model, and --also-effort must describe one complete secondary reviewer and cannot be combined with --single-reviewer. --continuous-review and --no-continuous-review cannot be combined. Do not modify Codex configuration, arbitrary pet files, global instruction files, or the reviewed repository. Pet-file writes are allowed only through the catalog/setup CLI above, only after an explicit request, and only for allowlisted buddy-* packages and private registry/backups.

Product boundary

The command-menu entry is a skill invocation selected by typing /buddy-review; it is not a third-party native slash-command registration. Continuous review uses exact private Git checkpoints, never screenshots or UI state. The plugin cannot programmatically wake, select, or make the native pet speak. Review text is delivered as one paragraph of at most three sentences and 700 characters through the audited Codex task transcript, while the full result remains in the private local receipt and the native pet supplies persistent animation and task-state signaling. If an exact final result is pending, the local event detail is Code review and suggestions are in progress.