vscode-ai-plugin

Review the conversation for threads that were started but not followed — ideas mentioned in passing, decisions deferred implicitly, tensions named but not resolved. Use when taking stock, before saving, or whenever the user wants to know what's been left on the table.

Use when drafting, updating, or reviewing a GitHub pull request title or body, including PR creation flows that need reviewer-facing prose. Follow the repo PR template when present, preserve the user's or project's voice, and route validation/session details to the right surface.

End-of-day close: the user is stepping away and won't serve as a bridge. Full collaborative triage, working-style reflection, and anticipatory framing for the next day. Invoke with /session-close.

Start a new session cycle by reconciling Project Orientation, restoring repo context, and resuming the active bet or opening vision steering when selection genuinely changed.

End-of-cycle transition: interpret the session, triage dangling threads with the user, and prepare SESSION-TRAJECTORY.md for the next cycle. The user is still around and serves as a light bridge. Invoke with /session-rest.

Keeps SESSION-TRAJECTORY.md warm as work progresses and offers checkpoints at natural boundaries. Saving is not stopping.

Use when coordinating multiple live Codex threads, maintaining a coordination heartbeat, routing work across implementation threads, inspecting cross-thread ownership or gates, or keeping a coordinated effort moving through PR, review, evidence, and human-action boundaries.

Use when running a prepare-execute-review (PER) cycle, doing a PER workflow, or coordinating work through `prepare`, `execute`, and `review` subagents with explicit gates.

Reorient a project by recovering its durable thesis and experienced outcome, reading live plans/tasks/PRs/substrate as candidate bets, and selecting one smallest meaningful end-to-end move with organic proof. Use when the user asks to steer or prioritize by vision, several technically valid gates compete without an active bet, an active bet's named steering signal has arrived, a closed milestone leaves no current bet, the user revises project purpose or outcome, or useful local work has lost its relationship to the project experience.

Use when selecting among locally valid project moves by relating them to a durable project thesis, or when new evidence should revise the route while the user's situated judgment remains authoritative about what the project is for.

Use when drafting, revising, reviewing, or promoting RFCs and design proposals, including Exo-managed RFCs, staged RFC lifecycle work, Rust/Ember-style RFC prose, and canonization or recommendation work for related RFCs.

Use when writing durable public design prose for RFCs, proposals, architecture notes, and canonization summaries where the document must serve users, implementers, maintainers, educators, and future readers.

Use when reviewing git diffs, local changes, or pull requests for merge-relevant risks: correctness regressions, security/privacy, data integrity, performance, reliability, tests, UX, maintainability, and deployability. Produce evidence-backed findings, inline comments when available, and a reviewer-facing summary focused on findings, residual risk, questions, and readiness.

Use when PR review or diff comments are supplied as actionable feedback to implement, including UI flows that send comments without an explicit typed request. Completing this work includes verifying the fix and resolving matching GitHub/Copilot review threads; observation-only requests should leave threads open.

Use when debugging, recon, review, or incident work has conflicting observations and needs to pause, restate what is known, check whether surfaces are comparable, and hand off to the next investigative move.

Use when investigating unfamiliar code, tracing data flows, mapping architecture, or answering questions that require adaptive exploration — where the agent needs to follow leads, use tools interactively, and synthesize findings rather than just search for known terms.

Use when walking through code changes with the user one chunk at a time, pausing for discussion after each chunk, linking directly to code, and checking design alignment as the review progresses.

Background guidance for all plugin work: keep language inside the work, focus on what we're doing, and use precise and meaningful language that stays attached to the work underway.

Use when observations from different surfaces appear to disagree and the agent must keep each observation in its observed form before explaining causes: UI vs behavior, logs vs state, metrics vs reports, tool output vs user-visible result.

Use when working with the user on tasks requiring their situated knowledge — priorities, intent, energy, context the agent can't see. Describes the relational structure of productive human-agent collaboration.

Use when investigating unfamiliar territory, narrowing possibilities through targeted questions, or diagnosing a problem by updating your own model of the situation.

The cognitive mode of reading for what a coherent narrative left out — threads abandoned, overtaken, or never recognized as significant. Complements interpretive-synthesis (which builds the account) by interrogating what it smoothed away.

Use when comparing expected outcomes against actual results, judging what a gap means, or evaluating whether a change achieved its intent. The review agent crystallizes this.

Use when predicting what code looks like before reading it, anticipating where a plan will encounter friction, or forming expectations specific enough to be tested. The prepare agent crystallizes this.

Use when building a coherent account from many observations, summarizing conversation state, or explaining how pieces connect. The skill is in what gets left out.

Use when walking through code, documents, or changes with the user where both perspectives contribute — the agent sees structural patterns, the user sees design intent.

Use when helping the user articulate their own thinking — when they have insight they haven't fully formed, or when the task requires the user to see something the agent can point to but not resolve.