Files
orca/src/shared/commit-message-plan.ts
T
Neil 71bbab72e1 fix(commit-message): keep Windows paths intact in agent command overrides (#14984)
* fix(commit-message): keep Windows paths intact in agent command overrides

`tokenizeCustomCommandTemplate` applies POSIX backslash-escape rules on every
platform. On Windows `\` is the path separator, so a native absolute path in an
agent command override is silently destroyed:

  C:\Windows\System32\WindowsPowerShell\v1.0\powershell.exe
  -> C:WindowsSystem32WindowsPowerShellv1.0powershell.exe

which is then reported as not found on PATH. The agent *startup* path already
routes Windows shells to the Windows tokenizer, but the commit-message AI path
still calls the generic tokenizer directly, so overrides, extra CLI args and
custom commands there are all affected.

The tokenizer gains an explicit `'escape' | 'literal'` mode rather than reading
`process.platform`, because the same template can be parsed on one host and
executed on another. `'escape'` stays the default, so POSIX behaviour — where
`foo\ bar` is deliberately one token — is unchanged.

`'literal'` is selected only where the command provably runs on native Windows:
a LOCAL target, on win32, with no WSL distro. A WSL target runs a Linux binary
inside the distro, and a remote target runs on a host whose platform this
process cannot see; both keep POSIX escaping.

Fixes #11375

* test: pin the platform decision for literal-backslash parsing

commandBackslashMode is the only place that reads the platform, so it is where
this can be wrong in the direction that matters — applying Windows rules to a
command that will actually run under a POSIX shell. WSL and remote targets are
pinned explicitly; both were previously untested.
2026-08-16 21:12:14 -07:00

336 lines
11 KiB
TypeScript

import type { CommandTemplateBackslash } from './commit-message-prompt'
import {
getCommitMessageAgentSpec,
getCommitMessageModel,
isCustomAgentId
} from './commit-message-agent-spec'
import { planCustomCommand, tokenizeCustomCommandTemplate } from './commit-message-prompt'
import type { TuiAgent } from './tui-agent'
// Why: planning is a pure transformation from "user request + prompt text"
// into "spawn-ready binary + argv". Keeping it in shared lets both the local
// generator (main process) and the SSH provider (which delegates to the
// relay over JSON-RPC) reuse the exact same validation and arg-building
// logic without duplicating the spec/model/thinking checks.
export type CommitMessagePlanInput = {
agentId: TuiAgent | 'custom'
/** How to read `\` in the user's command override / args / custom command.
* Defaults to POSIX escaping; pass `'literal'` only when the command is known
* to run on native Windows, where `\` is the path separator (#11375). */
backslash?: CommandTemplateBackslash
model: string
thinkingLevel?: string
customAgentCommand?: string
agentCommandOverride?: string
agentArgs?: string
}
export type CommitMessagePlan = {
binary: string
args: string[]
/** Non-null when the prompt should be piped via stdin. */
stdinPayload: string | null
/** Human-readable label used in error prefixes (e.g. "Claude failed: ..."). */
label: string
}
export type CommitMessagePlanResult =
| { ok: true; plan: CommitMessagePlan }
| { ok: false; error: string }
export function planAgentBinary(
defaultBinary: string,
commandOverride: string | undefined,
backslash: CommandTemplateBackslash = 'escape'
): { ok: true; binary: string; prefixArgs: string[] } | { ok: false; error: string } {
const command = commandOverride?.trim()
if (!command) {
return { ok: true, binary: defaultBinary, prefixArgs: [] }
}
const tokenized = tokenizeCustomCommandTemplate(command, backslash)
if (!tokenized.ok) {
return { ok: false, error: `Agent command override is invalid: ${tokenized.error}` }
}
const [binary, ...prefixArgs] = tokenized.tokens
if (!binary) {
return { ok: false, error: 'Agent command override must start with a binary name.' }
}
return { ok: true, binary, prefixArgs }
}
function planAdditionalAgentArgs(
agentArgs: string | null | undefined,
backslash: CommandTemplateBackslash = 'escape'
): { ok: true; args: string[] } | { ok: false; error: string } {
const trimmed = agentArgs?.trim()
if (!trimmed) {
return { ok: true, args: [] }
}
const tokenized = tokenizeCustomCommandTemplate(trimmed, backslash)
if (!tokenized.ok) {
return { ok: false, error: `CLI arguments are invalid: ${tokenized.error}` }
}
return { ok: true, args: tokenized.tokens }
}
const DEFAULT_SINGLETON_OPTIONS: readonly (readonly string[])[] = [['--model']]
function matchesOption(token: string, aliases: readonly string[]): boolean {
return aliases.some(
(alias) =>
token === alias ||
token.startsWith(`${alias}=`) ||
(alias.startsWith('-') &&
!alias.startsWith('--') &&
token.startsWith(alias) &&
token.length > alias.length)
)
}
function findOptionOccurrence(
tokens: string[],
aliases: readonly string[],
stopAtTerminator: boolean
): { index: number; consumed: number } | null {
for (let index = 0; index < tokens.length; index += 1) {
const token = tokens[index]
if (stopAtTerminator && token === '--') {
break
}
if (!matchesOption(token, aliases)) {
continue
}
const nextToken = tokens[index + 1]
const consumesNext =
aliases.includes(token) && nextToken !== undefined && !nextToken.startsWith('-')
return { index, consumed: consumesNext ? 2 : 1 }
}
return null
}
function applyRecipeOptionOverride(args: {
generatedArgs: string[]
recipeArgs: string[]
aliases: readonly string[]
}): { generatedArgs: string[]; recipeArgs: string[] } {
const recipeOption = findOptionOccurrence(args.recipeArgs, args.aliases, true)
const generatedOption = findOptionOccurrence(args.generatedArgs, args.aliases, false)
if (!recipeOption || !generatedOption) {
return { generatedArgs: args.generatedArgs, recipeArgs: args.recipeArgs }
}
const overrideTokens = args.recipeArgs.slice(
recipeOption.index,
recipeOption.index + recipeOption.consumed
)
return {
generatedArgs: [
...args.generatedArgs.slice(0, generatedOption.index),
...overrideTokens,
...args.generatedArgs.slice(generatedOption.index + generatedOption.consumed)
],
recipeArgs: [
...args.recipeArgs.slice(0, recipeOption.index),
...args.recipeArgs.slice(recipeOption.index + recipeOption.consumed)
]
}
}
function removeAllOptionOccurrences(tokens: string[], aliases: readonly string[]): string[] {
let result = tokens
while (true) {
const found = findOptionOccurrence(result, aliases, true)
if (!found) {
return result
}
result = [...result.slice(0, found.index), ...result.slice(found.index + found.consumed)]
}
}
/** Drops every occurrence after the first, so a user who types the same singleton
* twice in one field still gets a single flag rather than a rejected argv. */
function keepFirstOptionOccurrence(tokens: string[], aliases: readonly string[]): string[] {
let result = tokens
while (true) {
const first = findOptionOccurrence(result, aliases, true)
if (!first) {
return result
}
const tail = result.slice(first.index + first.consumed)
const duplicate = findOptionOccurrence(tail, aliases, true)
if (!duplicate) {
return result
}
const offset = first.index + first.consumed
result = [
...result.slice(0, offset + duplicate.index),
...result.slice(offset + duplicate.index + duplicate.consumed)
]
}
}
/** Removes generated singleton options shadowed by user input. Recipe args
* outrank a command-override prefix, which outranks Orca's generated value. */
function applySingletonOptionOverrides(args: {
generatedArgs: string[]
prefixArgs: string[]
recipeArgs: string[]
singletonOptions: readonly (readonly string[])[]
}): { generatedArgs: string[]; prefixArgs: string[]; recipeArgs: string[] } {
let generatedArgs = args.generatedArgs
let prefixArgs = args.prefixArgs
let recipeArgs = args.recipeArgs
for (const aliases of args.singletonOptions) {
recipeArgs = keepFirstOptionOccurrence(recipeArgs, aliases)
prefixArgs = keepFirstOptionOccurrence(prefixArgs, aliases)
const recipeOption = findOptionOccurrence(recipeArgs, aliases, true)
const prefixOption = findOptionOccurrence(prefixArgs, aliases, true)
const prefixHasTerminator = prefixArgs.includes('--')
if (recipeOption && !prefixHasTerminator) {
prefixArgs = removeAllOptionOccurrences(prefixArgs, aliases)
} else if (prefixOption && !prefixHasTerminator) {
const generatedOption = findOptionOccurrence(generatedArgs, aliases, false)
if (generatedOption) {
generatedArgs = [
...generatedArgs.slice(0, generatedOption.index),
...generatedArgs.slice(generatedOption.index + generatedOption.consumed)
]
}
continue
}
const withRecipe = applyRecipeOptionOverride({ generatedArgs, recipeArgs, aliases })
generatedArgs = withRecipe.generatedArgs
recipeArgs = withRecipe.recipeArgs
}
return { generatedArgs, prefixArgs, recipeArgs }
}
function insertAdditionalAgentArgs(args: {
baseArgs: string[]
agentArgs: string[]
promptDelivery: 'argv' | 'stdin'
prompt: string
}): string[] {
if (!args.agentArgs.length) {
return args.baseArgs
}
const promptPlaceholderIndex = args.baseArgs.lastIndexOf('{prompt}')
if (promptPlaceholderIndex !== -1) {
const merged = [...args.baseArgs]
merged.splice(promptPlaceholderIndex, 0, ...args.agentArgs)
return merged
}
if (
args.promptDelivery === 'argv' &&
args.prompt.length > 0 &&
args.baseArgs.at(-1) === args.prompt
) {
return [...args.baseArgs.slice(0, -1), ...args.agentArgs, args.prompt]
}
return [...args.baseArgs, ...args.agentArgs]
}
export function planCommitMessageGeneration(
input: CommitMessagePlanInput,
prompt: string
): CommitMessagePlanResult {
if (isCustomAgentId(input.agentId)) {
const command = input.customAgentCommand?.trim() ?? ''
if (!command) {
return {
ok: false,
error: 'Custom command is empty. Add one in Settings → Git → AI Commit Messages.'
}
}
const planned = planCustomCommand(command, prompt, input.backslash)
if (!planned.ok) {
return { ok: false, error: planned.error }
}
const agentArgs = planAdditionalAgentArgs(input.agentArgs, input.backslash)
if (!agentArgs.ok) {
return agentArgs
}
return {
ok: true,
plan: {
binary: planned.binary,
args: insertAdditionalAgentArgs({
baseArgs: planned.args,
agentArgs: agentArgs.args,
promptDelivery: planned.stdinPayload === null ? 'argv' : 'stdin',
prompt
}),
stdinPayload: planned.stdinPayload,
// Why: a custom command has no friendly name, so the binary doubles
// as the label in error prefixes ("ollama failed: ...").
label: planned.binary
}
}
}
const spec = getCommitMessageAgentSpec(input.agentId)
if (!spec) {
return { ok: false, error: `Agent "${input.agentId}" does not support AI commit messages.` }
}
const model = getCommitMessageModel(input.agentId, input.model)
if (!model) {
return { ok: false, error: `Model "${input.model}" is not available for ${spec.label}.` }
}
if (input.thinkingLevel) {
if (!model.thinkingLevels && spec.modelSource !== 'dynamic') {
return {
ok: false,
error: `Model "${model.label}" does not support a thinking effort level.`
}
}
if (model.thinkingLevels && !model.thinkingLevels.some((l) => l.id === input.thinkingLevel)) {
return {
ok: false,
error: `Thinking level "${input.thinkingLevel}" is not valid for ${model.label}.`
}
}
}
const argvPrompt = spec.promptDelivery === 'argv' ? prompt : ''
const baseArgs = spec.buildArgs({
prompt: argvPrompt,
model: input.model,
thinkingLevel: input.thinkingLevel
})
const agentArgs = planAdditionalAgentArgs(input.agentArgs, input.backslash)
if (!agentArgs.ok) {
return agentArgs
}
const command = planAgentBinary(spec.binary, input.agentCommandOverride, input.backslash)
if (!command.ok) {
return { ok: false, error: command.error }
}
// Why: repeating a singleton flag makes yargs-based CLIs parse it as an array and
// crash (OpenCode's `model.split('/')`). User values replace Orca's, never stack.
const merged = applySingletonOptionOverrides({
generatedArgs: baseArgs,
prefixArgs: command.prefixArgs,
recipeArgs: agentArgs.args,
singletonOptions: spec.singletonOptions ?? DEFAULT_SINGLETON_OPTIONS
})
const args = insertAdditionalAgentArgs({
baseArgs: merged.generatedArgs,
agentArgs: merged.recipeArgs,
promptDelivery: spec.promptDelivery,
prompt: argvPrompt
})
return {
ok: true,
plan: {
binary: command.binary,
args: [...merged.prefixArgs, ...args],
stdinPayload: spec.promptDelivery === 'stdin' ? prompt : null,
label: spec.label
}
}
}