Files
backnotprop__plannotator/packages/ai/context.ts
Michael Ramos cb6667e991 fix(ai): drive Codex Ask AI via codex app-server (#971)
* fix(ai): drive Codex Ask AI via app-server + answer-first review prompts (#971)

Codex Ask AI previously ran via @openai/codex-sdk (codex exec), which forces
approval_policy=never and breaks in enterprise-managed Codex environments that
ban it (#971). Replace the transport with a long-lived 'codex app-server'
process over JSON-RPC.

- New provider packages/ai/providers/codex-app-server.ts (registered as
  'codex-sdk' to preserve cookie/agents.ts/UI-gate); omits approvalPolicy so
  Codex resolves the user's + managed policy, pins read-only sandbox, and
  surfaces interactive approvals through the existing PermissionCard.
- Delete codex-sdk.ts and drop the @openai/codex-sdk dependency (and its 6
  prebuilt platform binaries); gate registration on 'which codex'.
- SessionManager: additive, optional dispose?() hook to kill the spawned
  process on evict/remove — a no-op for Claude/OpenCode/Pi (they don't
  implement it).

Also rework the Ask AI prompts (all providers, separate from the transport):

- Every mode now instructs the agent to answer the user's message directly and
  not launch an unprompted review of the context.
- Code review stops pasting the whole diff for git-reproducible diff types and
  instead tells the agent how to inspect it (git diff <base>..HEAD, three-dot
  for merge-base); non-git/PR/workspace types still paste.
- Claude gains the Bash tool so it can run git (still gated by approvals).
- The UI passes diffType/base (session) and what the user is viewing (per
  question) into the context.

Verified: full typecheck, full test suite (101 ai tests), and a live
end-to-end smoke against codex app-server.

* fix(review): pin AI approval card above the input/model bar

Render pending approval cards just above the input + provider/model bar in both the document chat (DocumentAIChatPanel) and code-review AI tab, instead of at the top of the scroll, so the user sees them where they act.

* fix(ai): harden Codex abort/cancel + add Ask AI Stop button

Addresses code-review findings on the codex app-server provider:

- turn/interrupt was sent as a notification (no id) so Codex ignored it and
  abort never took effect. It's now a proper JSON-RPC request.
- Filter turn events/approvals by turnId and reject an aborted turn's
  stragglers, so a stopped turn can no longer leak output into — or
  prematurely finish — the next turn (ask-stop-ask race).
- Guard listeners by query generation and end the drain loop on the abort
  signal, so a superseded/stopped turn can't touch the live one and abort
  returns promptly instead of waiting for turn/completed.
- Handle abort during startup: once the turn id is known, interrupt it
  instead of running it in the background.
- Add a sendAndWait timeout so a stalled (alive-but-unresponsive) process
  errors instead of hanging forever.
- Drain stderr (stdio 'ignore') to avoid a pipe-buffer deadlock.

Also add a Stop button to both Ask AI surfaces (plan/annotate
DocumentAIChatPanel and code-review AITab via ReviewSidebar). It replaces
Send while streaming and calls the hook's abort -> /api/ai/abort ->
session.abort(); the hook already exposed abort but nothing surfaced it.

* feat(ai): drive Codex models + reasoning levels from model/list

- Codex provider fetches model/list at startup (throwaway app-server, like
  Pi/OpenCode) and populates the real models plus each model's actual
  supportedReasoningEfforts + defaultReasoningEffort. Replaces the hardcoded
  model list and the static AI_REASONING_EFFORTS (which mislabeled xhigh as
  'Max' and omitted minimal).
- AIProviderBar + AIConfigBar now show the selected model's real efforts and
  hide the control when a model reports none. xhigh is shown verbatim.
- Fix: the Stop button now also appears in the populated code-review chat
  state (a prior edit missed the second GeneralInput due to indentation).

* fix(ai): scope Claude Ask AI Bash to read-only git; clear stale approval cards on Stop

- Claude Ask AI no longer auto-allows bare Bash (which ran arbitrary shell
  with no Allow/Deny prompt). Replace it with scoped read-only git rules
  (Bash(git diff:*), show, log, status, rev-parse, merge-base, ls-files).
  git reads auto-run so the agent can inspect large diffs itself; any other
  command (write git, arbitrary, or injected compound) falls through to the
  permission card. Keeps the git-inspect approach (large diffs don't fit in
  the prompt) while closing the auto-exec hole.
- useAIChat.abort() now drops still-undecided permission cards: abort cancels
  them server-side, so leaving them visible was a dead Allow/Deny.

* feat(ai): code-review Ask AI shares the agent-review prompt machine, delivered as user messages

Code-review Ask AI built its own diff description (gitInspectInstruction) in
the system prompt from just diffType+base, which was wrong for full-stack,
hide-whitespace, untracked files, and PR worktrees. Replace it with the same
machine the review jobs use, delivered on the user's messages.

- Server: buildCurrentAiReviewContext() reuses buildAgentReviewUserMessageForTarget
  (contextOnly) for the current view and ships it as aiReviewContext in every
  diff payload (/api/diff + switch/PR/scope). Mirrored in the Pi server.
- Client: review-editor latches aiReviewContext onto each question via the pure
  buildReviewContextPreamble (packages/ui/utils/aiPrompt.ts) and buildDefaultPrompt
  — full block on the first message / when the view changes (incl. after a
  provider switch via the !sessionId fresh-session check), a short reminder
  otherwise (never re-pastes a large diff).
- context.ts: delete the duplicate gitInspectInstruction; code-review system
  prompt is now role-only. Providers untouched (provider-agnostic user message).
- Tests: machine scenario gaps (plain PR, full-stack default, PR-worktree
  origin/<base> + stale-main warning, untracked mention, jj-evolog, workspace
  lines); composition (first/reminder, command/pasted, preamble ordering).

* fix(ai): agentic remote-PR context, UTF-8 stream decode, real Stop on supersede/disconnect

Addresses review findings on the Ask AI prompt work:

- Remote PR without a confirmed local checkout no longer gets URL-only. The
  agent is told it's in a PR worktree that's being prepared, to verify the PR
  files exist before relying on them, and to diff with git diff origin/<base>...
  HEAD (URL fallback). Inform + trust the agent rather than pasting. Shared
  machine, so review jobs get the same framing.
- Decode Codex stdout with a streaming TextDecoder instead of per-chunk
  toString(), so multi-byte UTF-8 split across chunks no longer corrupts into
  U+FFFD (matches the Pi provider).
- Stop now actually stops the server turn: ask() awaits /api/ai/abort when a
  new question supersedes a streaming one (awaiting avoids racing the new query
  into session_busy), and the /api/ai/query SSE stream gains a cancel handler
  so tab-close/navigation aborts the turn too. Both reuse the existing
  per-provider session.abort(); factored a shared postServerAbort helper.

* fix(ai): per-model reasoning effort, shared PR-checkout readiness, Stop-then-ask race, type hole

- Reasoning effort is now tracked per model (a map keyed by model) instead of
  one global value, so switching to a model that doesn't support the prior
  level (e.g. xhigh) no longer posts a stale/unsupported effort that Codex
  rejects. Each model keeps its own level; nothing leaks across.
- Extract resolvePoolCwd into packages/shared/worktree-pool.ts (ready/pending/
  absent) and use it from both servers' resolvePRLocalCwd so the readiness rule
  can't drift. Fix the Pi Ask AI helper to ready-check like Bun, so a warming
  PR checkout no longer claims 'checked out at PR head' and misdirects the diff.
- Stop-then-ask no longer races into session_busy: the abort promise (from Stop
  or a superseding question) is stashed and the next ask() awaits it before
  sending. Stop still kills the turn instantly; a follow-up just waits the one
  abort round-trip.
- Declare aiReviewContext on the initial /api/diff response type.

* fix(ai): surface real Codex error messages, skip transient retries, auto-deny permission escalations

Validated against codex-rs:

- Codex's ErrorNotification nests the text at params.error.message
  (TurnError.message); we read params.message → always 'Unknown error'. Read
  the nested field (top-level fallback) so auth/usage-limit/stream failures show
  their real, actionable text.
- Skip transient error notifications (willRetry=true): Codex retries on its own
  and the turn continues, so surfacing them flashed a spurious failure before
  the real answer.
- Handle item/permissions/requestApproval: respond {permissions:{}, scope:'turn'}
  — byte-for-byte Codex's own cancel response (codex_delegate.rs) — instead of
  'Unsupported request'. Ask AI is read-only, so denying the escalation lets the
  turn continue sandboxed rather than failing. Interactive grant deferred.

* refactor(ai): share AI provider/model config in one hook; aggregate paginated model/list

- Extract useAIProviderConfig (packages/ui/hooks): one home for provider/model/
  reasoning-effort selection — initial state, auto-resolve on capabilities load,
  per-model effort (no leak across models), and persistence. Both the plan and
  code-review apps now call it and only compose the session reset (the hook can't
  own reset without a cycle through useAIChat). Plan editor adapted to the
  effect-based resolve (adds aiDefaultProvider state). This fixes the plan
  editor's stale-effort-on-model-switch bug by construction and stops the two
  apps' copies from drifting again.
- fetchModels now follows model/list's nextCursor and aggregates every page into
  one list, so larger model catalogs aren't silently truncated (with a page
  guard against a misbehaving cursor).

Skipped per discussion: legacy v1 approval methods (we're a v2 client).

* feat(agents): per-agent review default; bound Codex model discovery so it can't stall the AI panel

- The selected review profile is now tracked per review engine (claude/codex/
  cursor/opencode) instead of one flat global value, so each agent keeps its own
  review default. Public hook API (reviewProfileId/setReviewProfileId) is
  unchanged — getter derives the current engine's value, setter writes it — so
  AgentsTab needs no changes. One-shot migration seeds every engine with any
  existing flat pick. Adds parseReviewProfileByEngine + tests.
- fetchModels (Codex model discovery) now uses a short 6s timeout for its
  initialize + model/list RPCs instead of the 30s default. /api/ai/capabilities
  awaits model discovery, so an installed-but-unauthenticated codex could
  otherwise block the AI panel for ~30s; it now falls back to the static model
  list fast. Authed codex is unaffected (discovery completes in tens of ms).
2026-06-28 09:58:09 -07:00

249 lines
8.0 KiB
TypeScript

/**
* Context builders — translate Plannotator review state into system prompts
* that give the AI session the right background for answering questions.
*
* These are provider-agnostic: any AIProvider implementation can use them
* to build the system prompt it needs.
*/
import type { AIContext } from "./types.ts";
// ---------------------------------------------------------------------------
// Public API
// ---------------------------------------------------------------------------
/**
* Build a system prompt from the given context.
*
* The prompt tells the AI:
* - What role it plays (plan reviewer, code reviewer, etc.)
* - The content it should reference (plan markdown, diff patch, file)
* - Any annotations the user has already made
* - That it's operating inside Plannotator (not a general coding session)
*/
export function buildSystemPrompt(ctx: AIContext): string {
switch (ctx.mode) {
case "plan-review":
return buildPlanReviewPrompt(ctx);
case "code-review":
return buildCodeReviewPrompt();
case "annotate":
return buildAnnotatePrompt(ctx);
}
}
/**
* Build a compact context summary suitable for injecting into a fork prompt.
*
* When forking from a parent session, we don't need a full system prompt
* (the parent's history already provides context). Instead, we inject a
* short "you are now in Plannotator" preamble with the relevant content.
*/
export function buildForkPreamble(ctx: AIContext): string {
const lines: string[] = [
"The user is now reviewing your work in Plannotator and has a question.",
"Answer the user's message directly and concisely based on the conversation " +
"history and the context below. Do not re-review or summarize the work unless they ask.",
"",
];
switch (ctx.mode) {
case "plan-review": {
lines.push("## Current Plan Under Review");
if (ctx.plan.version) {
const total = ctx.plan.totalVersions ? ` of ${ctx.plan.totalVersions}` : "";
lines.push(`Plan version: ${ctx.plan.version}${total}`);
}
if (ctx.plan.project) {
lines.push(`Project: ${ctx.plan.project}`);
}
lines.push("");
lines.push(truncate(ctx.plan.plan, MAX_PLAN_CHARS));
if (ctx.plan.annotations) {
lines.push("");
lines.push("## User Annotations So Far");
lines.push(ctx.plan.annotations);
}
break;
}
case "code-review": {
if (ctx.review.filePath) {
lines.push(`## Reviewing: ${ctx.review.filePath}`);
}
if (ctx.review.selectedCode) {
lines.push("");
lines.push("### Selected Code");
lines.push("```");
lines.push(ctx.review.selectedCode);
lines.push("```");
}
if (ctx.review.lineRange) {
const { start, end, side } = ctx.review.lineRange;
lines.push(`Lines ${start}-${end} (${side} side)`);
}
lines.push("");
lines.push("## Diff Patch");
lines.push("```diff");
lines.push(truncate(ctx.review.patch, MAX_DIFF_CHARS));
lines.push("```");
if (ctx.review.annotations) {
lines.push("");
lines.push("## User Annotations So Far");
lines.push(ctx.review.annotations);
}
break;
}
case "annotate": {
lines.push(`## Annotating: ${ctx.annotate.filePath}`);
if (ctx.annotate.sourceInfo) {
lines.push(`Source: ${ctx.annotate.sourceInfo}`);
}
if (ctx.annotate.renderAs) {
lines.push(`Render mode: ${ctx.annotate.renderAs}`);
}
if (ctx.annotate.sourceConverted) {
lines.push("Note: this content was converted before annotation, so source line numbers may not match the original document.");
}
lines.push("");
lines.push(truncate(ctx.annotate.content, MAX_PLAN_CHARS));
if (ctx.annotate.annotations) {
lines.push("");
lines.push("## User Annotations So Far");
lines.push(ctx.annotate.annotations);
}
break;
}
}
return lines.join("\n");
}
/**
* Build the effective prompt for a query, prepending a preamble on the first
* message. Used by providers that inject context via the prompt itself (Codex,
* Pi) rather than a separate system-prompt channel (Claude).
*/
export function buildEffectivePrompt(
userPrompt: string,
preamble: string | null,
firstQuerySent: boolean,
): string {
if (!firstQuerySent && preamble) {
return `${preamble}\n\n---\n\nUser question: ${userPrompt}`;
}
return userPrompt;
}
// ---------------------------------------------------------------------------
// Internals
// ---------------------------------------------------------------------------
const MAX_PLAN_CHARS = 60_000;
const MAX_DIFF_CHARS = 40_000;
/**
* Leading instruction for every Ask AI session. Ask AI is a chat assistant —
* it must respond to the user's message, not launch an unprompted review of the
* material it was given for context.
*/
const ANSWER_DIRECTLY =
"You are a helpful assistant inside Plannotator. Respond to the user's message directly and concisely. " +
"The material below is context for what the user is looking at — do NOT review, summarize, or critique it unless the user's message asks you to. " +
"Only investigate further (read files, run git) if the user's question actually requires it.";
function truncate(text: string, max: number): string {
if (text.length <= max) return text;
return `${text.slice(0, max)}\n\n... [truncated for context window]`;
}
function buildPlanReviewPrompt(
ctx: Extract<AIContext, { mode: "plan-review" }>
): string {
const sections: string[] = [
ANSWER_DIRECTLY,
"",
"The user is reviewing an implementation plan in Plannotator.",
"",
"## Plan Under Review",
];
if (ctx.plan.version) {
const total = ctx.plan.totalVersions ? ` of ${ctx.plan.totalVersions}` : "";
sections.push(`Plan version: ${ctx.plan.version}${total}`);
}
if (ctx.plan.project) {
sections.push(`Project: ${ctx.plan.project}`);
}
sections.push("");
sections.push(truncate(ctx.plan.plan, MAX_PLAN_CHARS));
if (ctx.plan.previousPlan) {
sections.push("");
sections.push("## Previous Plan Version (for reference)");
sections.push(truncate(ctx.plan.previousPlan, MAX_PLAN_CHARS / 2));
}
if (ctx.plan.annotations) {
sections.push("");
sections.push("## User Annotations");
sections.push(ctx.plan.annotations);
}
return sections.join("\n");
}
/**
* Code-review system prompt: role only. The actual changeset and how to inspect
* it are NOT in the system prompt — they ride on the user's messages, where the
* review server's shared agent-review prompt machine describes the *current*
* view (a git command to run, or the diff inline for modes git can't reproduce)
* and the client latches it on. See packages/server/review.ts
* `buildCurrentAiReviewContext` and packages/ui/utils/aiPrompt.ts.
*/
function buildCodeReviewPrompt(): string {
return [
ANSWER_DIRECTLY,
"",
"The user is reviewing a set of code changes in Plannotator. Their messages " +
"describe the changeset under review and how to inspect it — either a git " +
"command to run, or the diff inline.",
].join("\n");
}
function buildAnnotatePrompt(
ctx: Extract<AIContext, { mode: "annotate" }>
): string {
const sections: string[] = [
ANSWER_DIRECTLY,
"",
"The user is annotating a markdown document in Plannotator.",
"",
`## Document: ${ctx.annotate.filePath}`,
];
if (ctx.annotate.sourceInfo) {
sections.push(`Source: ${ctx.annotate.sourceInfo}`);
}
if (ctx.annotate.renderAs) {
sections.push(`Render mode: ${ctx.annotate.renderAs}`);
}
if (ctx.annotate.sourceConverted) {
sections.push("Note: this content was converted before annotation, so source line numbers may not match the original document.");
}
sections.push("");
sections.push(truncate(ctx.annotate.content, MAX_PLAN_CHARS));
if (ctx.annotate.annotations) {
sections.push("");
sections.push("## User Annotations");
sections.push(ctx.annotate.annotations);
}
return sections.join("\n");
}