@haystackeditor/cli 0.15.26 → 0.15.28

package/dist/index.js

@@ -30,11 +30,18 @@ import { initCommand } from './commands/init.js';
 import { authListCommand, authUseCommand, loginCommand, logoutCommand, whoamiCommand, } from './commands/login.js';
 import { handleAgenticTool, handleAutoMerge, handleAutoFix, handleWaitForReviewers, isAutoMergeEnabled, isAutoFixEnabled } from './commands/config.js';
 import { installSkills, listSkills } from './commands/skills.js';
-import { hooksInstall, hooksStatus, hooksUpdate } from './commands/hooks.js';
+import { cloudVerifierUniverseValidateCommand } from './commands/cloud-verifier-universe.js';
+import { cloudVerifierIdentityCensusCommand } from './commands/cloud-verifier-identity-census.js';
+import { cloudVerifierDataStoreDriftCommand } from './commands/cloud-verifier-data-store-drift.js';
+import { cloudVerifierBehaviorsValidateCommand } from './commands/cloud-verifier-behaviors.js';
+import { prepareUniverseReviewCommand } from './commands/prepare-universe-review.js';
+import { scaffoldProvisionalUniverseCommand } from './commands/scaffold-provisional-universe.js';
+import { hooksInstall, hooksStatus } from './commands/hooks.js';
 import { submitCommand } from './commands/submit.js';
 import { installSessionHooks, sessionHooksStatus } from './commands/install-session-hooks.js';
 import { listPolicies, addPolicy, removePolicy, initPolicies, addInstruction } from './commands/policy.js';
 import { triageCommand } from './commands/triage.js';
+import { designVerifyCommand } from './commands/design-verify.js';
 import { dismissCommand, markReviewedCommand, undismissCommand } from './commands/dismiss.js';
 import { requestReviewCommand } from './commands/request-review.js';
 import { reviewCommand } from './commands/review.js';
@@ -42,7 +49,7 @@ import { prStatusCommand } from './commands/pr-status.js';
 import { prReadCommand } from './commands/pr.js';
 import { inboxListCommand } from './commands/inbox.js';
 import { askHaystackCommand } from './commands/ask.js';
-import { tracesGetCommand, tracesListCommand } from './commands/traces.js';
+import { telemetryInstrumentCommand } from './commands/telemetry.js';
 import { setupCommand } from './commands/setup.js';
 import { schemaCommand, listSchemas } from './commands/schema-cmd.js';
 import { registerWebhook, listWebhooks, rotateWebhookSecret, setWebhookEnabled, listDeliveries, replayDelivery, } from './commands/webhooks.js';
@@ -126,7 +133,6 @@ program
     // Defining --no-auto-merge also accepts --auto-merge; default is "ask" unless
     // one is explicitly passed (detected via getOptionValueSource below).
     .option('--no-auto-merge', 'Set auto-merge in the written config (--auto-merge / --no-auto-merge; omit to be asked)')
-    .option('--skip-entire', 'Skip the Entire CLI / git-hooks (session tracking) step')
     .option('--answers <file>', 'JSON file of {questionId: value} pre-supplied answers')
     .addHelpText('after', `
 Steps: verify GitHub App → select repos → scan (rules/policies/instructions) →
@@ -135,7 +141,7 @@ review → write .haystack/pr-rules.yml + review-policy.md + .haystack.json.
 Three ways to run:
   • Interactive (default): prompts in your terminal.
   • Autonomous/CI: supply every decision via flags, e.g.
-      haystack setup --repo owner/name --yes --no-auto-merge --skip-entire
+      haystack setup --repo owner/name --yes --no-auto-merge
   • Coding agent: --json speaks NDJSON on stdout and reads answers on stdin, so an
     agent can auto-answer or relay questions to the user. Pre-supplied answers
     (flags / --answers) skip the matching question.
@@ -157,7 +163,6 @@ Examples:
         repos: options.repo,
         yes: options.yes,
         autoMerge: autoMergeExplicit ? options.autoMerge : undefined,
-        skipEntire: options.skipEntire,
         answersFile: options.answers,
     });
 });
@@ -205,6 +210,18 @@ Examples:
 `)
     .action((specimen, options) => runPublicCommand(() => verifyCommand(specimen, options), options.json));
 const verifyRunArg = 'Run id, chapter id (newest run of that chapter), or "latest" (default)';
+verify
+    .command('precompute')
+    .description('Precompute reusable verification inputs for the current working tree')
+    .option('--hook', 'Run silently as a best-effort repository hook')
+    .action(async (options) => {
+    await runPublicCommand(async () => {
+        const { verifyPrecomputeCommand } = await import('./commands/verify-precompute.js');
+        await verifyPrecomputeCommand(options);
+    });
+    if (options.hook)
+        process.exitCode = 0;
+});
 const hostedVerify = verify
     .command('hosted')
     .description('Run the production Cloud Verifier against exact GitHub commits');
@@ -647,7 +664,7 @@ program
     .addHelpText('after', `
 This command is designed for AI coding agents to submit PRs.
 
-  1. Runs pre-PR triage (code review, rules, instruction drift) via sub-agents
+  1. Runs pre-PR triage (code review and rules) via sub-agents
   2. Pushes the current branch to origin
   3. Creates a pull request on GitHub
   4. Waits for Haystack analysis results (triggered via GitHub App webhook)
@@ -656,15 +673,14 @@ Pre-PR Triage:
   Before creating the PR, haystack spawns parallel sub-agents to check for:
   • Code review bugs (logic errors, null crashes, security issues)
   • Rule violations (from .haystack/pr-rules.yml)
-  • Instruction drift (AI agent deviations from user instructions)
 
   Use --force to skip triage entirely.
   Use --no-wait to skip waiting for analysis results.
 
   Triage budgets:
   • --max-turns <n>         Raise/lower the per-checker tool-use turn cap.
-                            Defaults: code-review 8, rules-validator 10,
-                            intent-drift 10. Applies the same N to all three.
+                            Defaults: code-review 8, rules-validator 10.
+                            Applies the same N to both.
   • --triage-timeout <sec>  Raise/lower the per-checker wall-clock timeout
                             (default: 180s).
 
@@ -782,7 +798,7 @@ inbox
 const prProgram = program.command('pr').description('Inspect one pull request');
 prProgram
     .command('get <ref>')
-    .description('Get triage, merge blockers, and trace availability')
+    .description('Get triage and merge blockers')
     .option('--json', 'Output as JSON')
     .action((ref, options) => runPublicCommand(() => prReadCommand(ref, options), options.json));
 program
@@ -791,19 +807,32 @@ program
     .option('--json', 'Output the answer and every consulted customer-facing source as JSON')
     .option('--session <id>', 'Continue a previous Ask Haystack session')
     .action((ref, question, options) => runPublicCommand(() => askHaystackCommand(ref, question, options), options.json));
-const traces = program.command('traces').description('Inspect customer-owned Entire traces');
-traces
-    .command('list <ref>')
-    .description('List retained checkpoints for a pull request')
-    .option('--json', 'Output as JSON')
-    .action((ref, options) => runPublicCommand(() => tracesListCommand(ref, options), options.json));
-traces
-    .command('get <ref> <checkpoint>')
-    .description('Get retained transcript chunks for one checkpoint')
-    .option('--cursor <cursor>', 'Zero-based chunk cursor')
-    .option('--limit <count>', 'Chunks per page', '20')
-    .option('--json', 'Output as JSON')
-    .action((ref, checkpoint, options) => runPublicCommand(() => tracesGetCommand(ref, checkpoint, options), options.json));
+const telemetry = program.command('telemetry').description('Add privacy-safe production telemetry without an application SDK');
+telemetry
+    .command('instrument <dist>')
+    .description('Instrument compiled Node.js output without changing application source')
+    .requiredOption('--entry <file>', 'Entry file relative to the compiled output directory')
+    .option('--source-root <dir>', 'Source directory whose files correspond exactly to compiled output (default: sibling src/)')
+    .option('--include-symbols <file>', 'Experimental JSON allow-list of exact sourcePath + qualifiedName symbols')
+    .option('--json', 'Machine-readable instrumentation manifest')
+    .addHelpText('after', `
+Run this once after the normal application build. It rewrites only compiled
+JavaScript and copies a dependency-free runtime beside it; application source
+does not import Haystack.
+
+At runtime, telemetry stays off unless all three variables are set:
+  HAYSTACK_TELEMETRY=1
+  HAYSTACK_TELEMETRY_ENDPOINT=https://.../v3/telemetry
+  HAYSTACK_TELEMETRY_TOKEN=...
+
+Example:
+  npm run build && haystack telemetry instrument dist --entry index.js
+`)
+    .action((dist, options) => runPublicCommand(() => telemetryInstrumentCommand(dist, {
+    ...options,
+    includeSymbolsFile: options.includeSymbols,
+    includeSymbols: undefined,
+}), options.json));
 program
     .command('dismiss <pr>')
     .description('Dismiss analysis findings for a PR')
@@ -826,6 +855,43 @@ Examples:
   haystack dismiss acme/widgets#99     # Dismiss for specific repo
 `)
     .action(dismissCommand);
+program
+    .command('design-verify <pr>')
+    // "verification" is a banned word in top-level help (public-contract test
+    // keeps retired auto-fix/verification wording out of the CLI surface).
+    .description('Run a blind design-by-execution review of a PR')
+    .option('--json', 'Machine-readable output')
+    .option('--no-wait', 'Return after the run is admitted instead of waiting for its verdict')
+    .option('--run <run-id>', 'Resume and wait for an existing exact run id')
+    .option('--poll-interval <seconds>', 'Status polling interval', '3')
+    .option('--timeout <seconds>', 'Maximum time to wait; the server run continues after timeout', '7200')
+    .option('--case <case-id>', 'Select one exact generated case for universe exploration')
+    .option('--universe <role>', 'Select control-before, before, or after')
+    .option('--replay', 'Replay the registered case command in the selected universe')
+    .option('--exec-json <argv>', 'Run a JSON argv array in the selected universe')
+    .option('--cwd <path>', 'Repository-relative working directory for --exec-json')
+    .option('--command-timeout <seconds>', 'Maximum selected-universe command runtime')
+    .option('--action <action-id>', 'Resume one exact universe action')
+    .addHelpText('after', `
+Executes the change instead of reading it: infers intent from the diff
+alone, designs pre-registered behavioral test cases, runs control-before,
+before, and after in isolated sandboxes, and reports anomaly candidates
+outside the registered intent.
+
+The command waits for a behavioral verdict by default. Use --no-wait to
+return the run id immediately, or --run <run-id> to resume waiting.
+
+After a completed run, inspect or replay one retained case universe without
+receiving provider credentials:
+  haystack design-verify owner/repo#123 --run RUN --case case-00001 --universe before --replay
+  haystack design-verify owner/repo#123 --run RUN --case case-00001 --universe after --exec-json '["rg","TODO"]'
+
+PR identifier formats:
+  123                                     PR number (uses current repo)
+  owner/repo#123                          Fully qualified
+  https://github.com/owner/repo/pull/123  GitHub URL
+`)
+    .action(designVerifyCommand);
 program
     .command('mark-reviewed <pr>')
     .description('Mark human review as not needed for a PR')
@@ -1035,73 +1101,128 @@ Examples:
 // Skills subcommands
 const skills = program
     .command('skills')
-    .description('Manage AI skills for Claude Code');
+    .description('Manage AI skills for coding agents');
 skills
     .command('install')
-    .description('Install Haystack skills into your coding CLI')
+    .description('Install portable Haystack skills and optional coding-CLI shims')
     .option('--cli <name>', 'Target CLI: claude, codex, cursor, or manual')
     .addHelpText('after', `
-This registers the Haystack MCP server with your coding CLI, enabling:
-  /setup-haystack         - AI-assisted project setup
+This installs portable skills in the Git repository's .agents/skills directory:
   /submit                 - Submit a PR via Haystack
+  /map-your-system        - Map how Haystack QA can run this system
+  /map-cloud-verifier-universe - Map a production-derived hermetic universe
 Supported CLIs:
-  claude    Claude Code (auto-detected)
-  codex     Codex CLI (auto-detected)
-  cursor    Cursor IDE (auto-detected)
-  manual    Show manual setup instructions
+  claude    Also install Claude Code command shims
+  codex     Use portable .agents/skills discovery only
+  cursor    Use portable .agents/skills discovery only
+  manual    Install portable skills and show their location
 
 Examples:
-  haystack skills install              # Auto-detect and install
+  haystack skills install              # Install portable skills only
   haystack skills install --cli codex  # Install for Codex only
-  haystack skills install --cli manual # Show manual instructions
+  haystack skills install --cli claude # Also install Claude command shims
 `)
-    .action(installSkills);
+    .action(async (opts) => {
+    try {
+        await installSkills(opts);
+    }
+    catch (err) {
+        console.error(chalk.red('skills install failed:'), err instanceof Error ? err.message : err);
+        process.exit(1);
+    }
+});
 skills
     .command('list')
     .description('List available Haystack skills')
     .action(listSkills);
-// Hooks subcommands
+skills
+    .command('census-universe')
+    .description('Deterministically census Cloud Verifier source identities before mapping')
+    .option('--json', 'Machine-readable result')
+    .option('--root <path>', 'Repository root (default: current Git worktree)')
+    .action((opts) => runPublicCommand(async () => cloudVerifierIdentityCensusCommand(opts), opts.json));
+skills
+    .command('data-store-drift')
+    .description('Check whether the detected data stores still match the checked-in map (per-PR drift gate)')
+    .option('--json', 'Machine-readable result')
+    .option('--root <path>', 'Repository root (default: current Git worktree)')
+    .option('--baseline <path>', 'Baseline map path (default: .haystack/cloud-verifier/data-stores.json)')
+    .option('--base-ref <ref>', 'Only run when the diff against this ref touches a store-relevant file')
+    .option('--update', 'Write the current detected stores as the new baseline')
+    .action((opts) => runPublicCommand(async () => cloudVerifierDataStoreDriftCommand(opts), opts.json));
+skills
+    .command('validate-universe')
+    .description('Validate Cloud Verifier universe artifacts and stable-ID references')
+    .option('--json', 'Machine-readable result')
+    .action(async (opts) => {
+    try {
+        await cloudVerifierUniverseValidateCommand(opts);
+    }
+    catch (err) {
+        console.error(chalk.red('validate-universe failed:'), err instanceof Error ? err.message : err);
+        process.exit(1);
+    }
+});
+skills
+    .command('validate-behaviors')
+    .description('Validate customer actions and their exact runtime identities')
+    .option('--json', 'Machine-readable result')
+    .action(async (opts) => {
+    try {
+        await cloudVerifierBehaviorsValidateCommand(opts);
+    }
+    catch (err) {
+        console.error(chalk.red('validate-behaviors failed:'), err instanceof Error ? err.message : err);
+        process.exit(1);
+    }
+});
+skills
+    .command('scaffold-provisional-universe')
+    .description('Create a compact missing-authority receipt from strict JSON facts')
+    .requiredOption('--input <path>', 'Strict provisional-universe JSON input file')
+    .action((opts) => runPublicCommand(async () => scaffoldProvisionalUniverseCommand(opts), true));
+skills
+    .command('prepare-universe-review')
+    .description('Create a tracked-source snapshot with conventional test paths removed')
+    .option('--json', 'Machine-readable isolation result and snapshot manifest')
+    .option('--exclude-inactive-submodule <path...>', 'Omit exact tracked gitlink paths explicitly known to be inactive')
+    .option('--omit-unscannable-files', 'Omit files whose test content cannot be ruled out, instead of blocking the review. '
+    + 'The reviewer still never sees a test; the omissions are recorded as uncovered scope')
+    .option('--full-receipt', 'Include complete non-blocking path receipts in JSON output')
+    .action((opts) => runPublicCommand(() => prepareUniverseReviewCommand(opts), opts.json));
 const hooks = program
     .command('hooks')
     .description('Manage git hooks for AI agent quality checks');
 hooks
     .command('install')
-    .description('Install Haystack git hooks and Entire CLI')
-    .option('--version <version>', 'Entire CLI version (default: pinned)')
+    .description('Install Haystack git hooks')
     .option('-f, --force', 'Overwrite existing hooks')
-    .option('--skip-entire', 'Skip Entire binary download')
     .addHelpText('after', `
 This installs:
   • Git hooks for AI agent quality checks (pre-commit, commit-msg, etc.)
   • Agent context detector (identifies AI agent sessions)
   • Truncation checker (prevents code truncation by LLMs)
-  • Entire CLI binary for session tracking (powered by https://entire.dev)
 
 Hooks are installed to <repo>/hooks/ and git is configured to use them.
 
 Examples:
-  haystack hooks install              # Install with pinned Entire version
   haystack hooks install --force      # Overwrite existing hooks
-  haystack hooks install --skip-entire # Only install Haystack hooks
 `)
     .action(hooksInstall);
 hooks
     .command('status')
     .description('Check hooks installation status')
     .action(hooksStatus);
-hooks
-    .command('update')
-    .description('Update Entire CLI to the latest version')
-    .action(hooksUpdate);
 hooks
     .command('install-session')
-    .description('Install session-start hooks for coding CLIs')
+    .description('Install session-start and Stop hooks for coding CLIs')
     .option('--cli <name>', 'Target CLI: claude, codex, gemini, or all')
     .addHelpText('after', `
 Configures your coding CLI to run \`haystack triage --hook\` on session start.
 This shows pending PR analysis results when you open a new terminal session.
+Claude Code also runs \`haystack verify precompute --hook\` when a session stops.
 
-  Claude Code:  Native SessionStart hook (.claude/settings.json)
+  Claude Code:  Native SessionStart and Stop hooks (.claude/settings.json)
   Codex CLI:    AGENTS.md instructions
   Gemini CLI:   GEMINI.md instructions
 

package/dist/schema.d.ts

@@ -12,11 +12,10 @@
 export declare const SCHEMA_VERSIONS: {
     readonly triage: "2.0.0";
     readonly setup: "1.0.0";
-    readonly pr: "2.0.0";
+    readonly pr: "3.0.0";
     readonly 'pr-status': "1.0.0";
     readonly inbox: "1.0.0";
     readonly ask: "1.0.0";
-    readonly traces: "1.0.0";
     readonly submit: "1.0.0";
     readonly action: "1.0.0";
     readonly 'cloud-verifier': "1.0.0";

package/dist/schema.js

@@ -12,11 +12,10 @@
 export const SCHEMA_VERSIONS = {
     triage: '2.0.0',
     setup: '1.0.0',
-    pr: '2.0.0',
+    pr: '3.0.0',
     'pr-status': '1.0.0',
     inbox: '1.0.0',
     ask: '1.0.0',
-    traces: '1.0.0',
     submit: '1.0.0',
     action: '1.0.0',
     'cloud-verifier': '1.0.0',

package/dist/triage/prompts.d.ts

@@ -22,10 +22,3 @@ export declare function buildRulesValidatorPrompt(baseBranch: string, rulesYaml:
     filename: string;
     content: string;
 }[], precomputedDiff?: string | null): string | null;
-/**
- * Build the intent drift prompt.
- * Only runs if relevant trace files exist.
- *
- * @returns The prompt string, or null if no trace files provided.
- */
-export declare function buildIntentDriftPrompt(baseBranch: string, traceFiles: string[], outputPath: string, maxTurns: number, timeoutMs: number, precomputedDiff?: string | null): string | null;

package/dist/triage/prompts.js

@@ -37,21 +37,6 @@ const RULES_VALIDATOR_SCHEMA = `{
   "rulesChecked": 3,
   "passed": true
 }`;
-const INTENT_DRIFT_SCHEMA = `{
-  "checker": "intent-drift",
-  "issues": [
-    {
-      "file": "relative/path/to/file.ts",
-      "line": 42,
-      "severity": "error | warning | info",
-      "message": "Description of the drift or incomplete fulfillment",
-      "pattern": "intent_drift | incomplete_fulfillment | scope_creep | unspecified_decision | ignored_correction | weakened_posture | claimed_but_not_done"
-    }
-  ],
-  "summary": "Brief 1-sentence summary",
-  "sessionsChecked": 2,
-  "passed": true
-}`;
 // ============================================================================
 // Prompt builders
 // ============================================================================
@@ -235,133 +220,3 @@ ${RULES_VALIDATOR_SCHEMA}
 - Set \`passed\` to \`false\` if any violations with severity "error" were found
 - You MUST write the result file even if no violations are found`;
 }
-/**
- * Build the intent drift prompt.
- * Only runs if relevant trace files exist.
- *
- * @returns The prompt string, or null if no trace files provided.
- */
-export function buildIntentDriftPrompt(baseBranch, traceFiles, outputPath, maxTurns, timeoutMs, precomputedDiff) {
-    if (traceFiles.length === 0)
-        return null;
-    const traceFileList = traceFiles.map(f => `- \`${f}\``).join('\n');
-    return `You are an intent drift detector. Your job is to check whether an AI coding agent faithfully implemented what the user asked for.
-
-${buildTimeBudgetHeader(maxTurns, timeoutMs)}## Context
-
-This PR was created by an AI coding agent. The agent's session transcripts (traces) are stored locally. You will compare what the user asked the agent to do against what was actually implemented in the diff.
-
-## Instructions
-
-1. Read each of these files (they contain the user's messages extracted from the agent session, one per section):
-${traceFileList}
-
-2. From each file, identify:
-   - The user's original instruction/prompt (the first message)
-   - Any follow-up instructions or corrections from the user
-   - The final effective scope after later corrections or reframes
-
-3. ${precomputedDiff ? 'Review the diff below' : `Run \`git diff ${baseBranch}...HEAD\``} to see what was actually implemented.
-
-4. Compare the user's intent against the actual implementation. Look for:
-
-### Intent Drift
-The agent implemented something DIFFERENT from what was asked:
-- User says "delay as long as possible" → agent uses a fixed 30s timeout
-- User says "only when X" → agent does it unconditionally
-- User says "use library A" → agent uses library B
-- User says "lazy load" → agent loads eagerly
-
-### Incomplete Fulfillment
-The agent didn't finish everything that was asked:
-- User requested 3 things, agent only did 2
-- User requested a general/systematic guardrail or end-to-end behavior, but the diff only handles one known instance, one special case, or one side of the required wiring
-- User initially mentioned a known example, then clarified "not the specific case" / "the general class"; the diff still implements only the known example or category-specific check
-- Interface fields declared but never populated
-- A new keyed capability, event, route, config value, enum variant, or serialized field is referenced on one side of a boundary but the required registry, producer, consumer, schema, handler, persistence path, or delivery surface is missing
-- A change appears to work through local/dev/test defaults, mocks, or overrides, but the production wiring path needed to deliver the requested behavior was not updated
-- Functions stubbed with TODO/placeholder comments
-- Agent said "I'll skip X for now" for something the user explicitly requested
-- Tests not written when user asked for tests
-
-### Scope Creep
-The agent added functionality the user never asked for:
-- User asks to fix a bug → agent also adds a cooldown, retry logic, or caching layer
-- User asks to investigate → agent proactively "fixes" things beyond what was discussed
-- User asks for one feature → agent bundles in extra features "while we're at it"
-- Any new mechanism not traceable to a user instruction
-
-### Unspecified Decision
-The user authorized the task's GOAL, but the agent made a specific decision in HOW it carried it out that shapes the program's end output, externally observable behavior, or the data end-users/callers see — and the user never specified or approved that particular decision, and a reasonable user would plausibly want a say in it. Examples (general — do NOT pattern-match on specific keywords):
-- Silently dropping, capping, sampling, or transforming data that flows to the output
-- Picking a default that determines what end-users see
-- Resolving an ambiguous requirement one way when other materially different behaviors were equally valid
-- Choosing a fixed value where the choice changes results
-Do NOT flag internal implementation choices with no observable effect (naming, file layout, helper structure), decisions forced by correctness, or cosmetic defaults a user would not care about. When unsure whether a user would care, do not flag it — reserve this for decisions with real, user-visible ramifications. The difference from scope creep: scope creep is an unrequested *flow*; an unspecified decision is an unrequested, output-shaping choice *inside a requested flow*.
-
-### Ignored Correction
-The user gave an explicit correction or redirection and the final code does NOT honor it:
-- User said "don't use a global / use X instead / that's racy, do Y" and the agent shipped the thing it was told not to
-- User said the target is the general class rather than the specific instance, but the agent still shipped only the instance-specific implementation
-- Agent applied the correction, then quietly reverted it in a later step
-- A general later "looks good" does NOT cancel a specific earlier correction
-This is high severity by default: the user actively steered and was overridden.
-
-### Weakened Posture
-In service of an authorized goal, the agent relaxed or removed a security, safety, or correctness guard the user never asked to weaken:
-- Loosened an auth/permission/validation check, or broadened access (CORS, scopes)
-- Swallowed or silenced an error, removed an assertion/guard, hardcoded a bypass or credential
-- Disabled a test (.skip / xit), suppressed a type or lint check (any, @ts-ignore, disable comments)
-- Lowered a threshold/timeout that existed as a safeguard
-- Relaxed a CI/pipeline security or quality gate (a workflow check, audit, or validation policy) — especially so that the agent's OWN change would pass that gate
-High severity by default. (If the user explicitly asked to remove the guard, it is not a finding.)
-Tie-breaker for gate relaxations: when the agent's own work was failing a CI gate and the agent changed the gate so it would pass, that is weakened_posture even when the old gate looks buggy or overly strict, and even when a delegation ("shepherd this", "get it merged") covered the goal — whether the relaxation was right is exactly the judgment this flag hands to a human.
-Wording for weakened_posture findings: describe the decision reflectively, lede first, in plain English — sentence 1 says who weakened which protection and why ("To get this change through CI, the agent loosened the check that was blocking it"), sentence 2 gives the concrete change, then hand the call to a human. If a delegation covered the surrounding goal, say so; never claim "no user directive" when one exists. This is not an accusation of violating intent — it is a consequential decision a human should confirm.
-
-### Claimed But Not Done
-The agent told the user it completed work that the diff does not actually contain, or contains only in a materially weaker form:
-- "I added error handling for timeouts" but no timeout handling exists in the diff
-- "Added tests for the edge case" but no test assertions exercise it
-- "Made the limit configurable" but the value is still hardcoded
-Compare each concrete claim the agent made to the user against what the diff actually shows. Only flag when the diff clearly contradicts or fails to support the claim; when unsure, do not flag.
-
-## Red flags to search for in the diff
-
-- Fixed/hardcoded values where dynamic behavior was requested
-- TODO, FIXME, placeholder, stub comments in new code
-- Empty function bodies or early returns
-- Interface fields that are declared but never assigned anywhere
-- Feature-specific or instance-specific checks added after the user asked for a general guardrail against a broader failure mode
-- Earlier narrow examples treated as the whole task even though later user messages broadened or generalized the requested scope
-- New string-keyed names, enum variants, serialized fields, action types, routes, events, or config keys that have no matching entry in the surrounding registry, allow-list, schema, handler, producer, consumer, or template path
-- Comments or neighboring code saying "must also register/list/wire this" where the diff updated only the reference side
-- New mechanisms (cooldowns, retries, caches, rate limits) not requested by the user
-- Decisions that drop, limit, or reshape data flowing to the output, or pick a default that changes what end-users see, with no instruction specifying it
-
-## Output
-
-Write your results to \`${outputPath}\` as JSON with this exact schema:
-
-\`\`\`json
-${INTENT_DRIFT_SCHEMA}
-\`\`\`
-
-- Set \`sessionsChecked\` to the number of trace files you read
-- Set \`passed\` to \`true\` only when no intent-drift issues are found
-- Set \`passed\` to \`false\` when any issue is found (any pattern, any severity)
-- For each issue, set \`pattern\` to one of: "intent_drift", "incomplete_fulfillment", "scope_creep", "unspecified_decision", "ignored_correction", "weakened_posture", "claimed_but_not_done"
-- Write each issue's \`message\` in plain English for the PR author, lede first (what concretely happened, then why it matters); never reference this checker's machinery ("per policy", "extracted directives", "session shape")
-- You MUST write the result file even if no issues are found
-
-## Severity guidelines
-
-- **error**: Core user intent violated — the main thing they asked for is wrong or missing, or large unrequested feature added
-- **warning**: Secondary requirement missed or approximated, or small unrequested mechanism added
-- **info**: Minor simplification that mostly still works as intended${precomputedDiff ? `
-
-## Diff (precomputed)
-
-\`\`\`diff
-${precomputedDiff}
-\`\`\`` : ''}`;
-}

package/dist/triage/runner.js

@@ -8,8 +8,7 @@ import { execSync, spawn } from 'child_process';
 import { existsSync, readFileSync, mkdirSync, rmSync } from 'fs';
 import { join } from 'path';
 import chalk from 'chalk';
-import { buildCodeReviewPrompt, buildRulesValidatorPrompt, buildIntentDriftPrompt } from './prompts.js';
-import { findRelevantTraces } from './traces.js';
+import { buildCodeReviewPrompt, buildRulesValidatorPrompt } from './prompts.js';
 import { resolveDiffBaseRef } from '../utils/git.js';
 import { trackError } from '../utils/telemetry.js';
 // ============================================================================
@@ -26,7 +25,6 @@ const DEFAULT_TIMEOUT_MS = 180_000; // 3 minutes
 const DEFAULT_MAX_TURNS = {
     'code-review': 8,
     'rules-validator': 10,
-    'intent-drift': 10,
 };
 // ============================================================================
 // Agent policy file discovery
@@ -101,7 +99,13 @@ export function buildCommand(cli, prompt, maxTurns) {
         case 'codex':
             return {
                 command: 'codex',
-                args: ['exec', prompt, '--full-auto'],
+                args: [
+                    '--sandbox', 'workspace-write',
+                    '--ask-for-approval', 'never',
+                    'exec',
+                    '--ephemeral',
+                    prompt,
+                ],
             };
         case 'gemini':
             return {
@@ -119,17 +123,12 @@ function spawnChecker(cli, config, cwd, timeoutMs) {
         const { command, args } = buildCommand(cli, config.prompt, config.maxTurns);
         let stdout = '';
         let stderr = '';
-        const codexHome = join(cwd, '.haystack', 'triage', 'codex-home');
-        if (cli === 'codex') {
-            mkdirSync(codexHome, { recursive: true });
-        }
         const proc = spawn(command, args, {
             cwd,
             stdio: ['ignore', 'pipe', 'pipe'],
             env: {
                 ...process.env,
                 CLAUDECODE: undefined,
-                CODEX_HOME: cli === 'codex' ? codexHome : process.env.CODEX_HOME,
             },
         });
         proc.stdout?.on('data', (data) => {
@@ -276,8 +275,8 @@ export async function runTriage(gitRoot, baseBranch, options = {}) {
     // Resolve the ref to diff against once. Prefers origin/<base> over the bare
     // local <base> ref (which is often stale and balloons the diff with already-
     // merged code). Every git read below — the precomputed diff, the agent diff
-    // commands in the prompts, and the changed-file scan in findRelevantTraces —
-    // uses this single resolved ref so they all see the same fork point.
+    // commands in the prompts use this single resolved ref so they all see the
+    // same fork point.
     const diffBaseRef = resolveDiffBaseRef(baseBranch);
     if (diffBaseRef !== baseBranch) {
         console.log(chalk.dim(`  Diff base: ${diffBaseRef}`));
@@ -323,26 +322,9 @@ export async function runTriage(gitRoot, baseBranch, options = {}) {
             });
         }
     }
-    // 3. Intent drift (only if relevant trace files exist)
-    const traceFiles = findRelevantTraces(gitRoot, diffBaseRef);
-    if (traceFiles.length > 0) {
-        const intentDriftOutput = join(triageDir, 'intent-drift.json');
-        const driftPrompt = buildIntentDriftPrompt(diffBaseRef, traceFiles, intentDriftOutput, maxTurns['intent-drift'], timeoutMs, precomputedDiff);
-        if (driftPrompt) {
-            checkers.push({
-                name: 'intent-drift',
-                prompt: driftPrompt,
-                outputFile: intentDriftOutput,
-                maxTurns: maxTurns['intent-drift'],
-            });
-        }
-    }
     // Log what we're running
     const checkerNames = checkers.map(c => c.name);
     console.log(chalk.dim(`  Checkers: ${checkerNames.join(', ')}`));
-    if (traceFiles.length > 0) {
-        console.log(chalk.dim(`  Traces: ${traceFiles.length} relevant session(s) found`));
-    }
     console.log(chalk.dim(`  Running ${checkers.length} checker(s) in parallel...\n`));
     // Spawn all checkers in parallel
     const spawnResults = await Promise.allSettled(checkers.map(async (checker) => {

package/dist/triage/types.d.ts

@@ -28,14 +28,7 @@ export interface RulesValidatorResult {
     passed: boolean;
     rulesChecked: number;
 }
-export interface IntentDriftResult {
-    checker: 'intent-drift';
-    issues: TriageIssue[];
-    summary: string;
-    passed: boolean;
-    sessionsChecked: number;
-}
-export type CheckerResult = CodeReviewResult | RulesValidatorResult | IntentDriftResult;
+export type CheckerResult = CodeReviewResult | RulesValidatorResult;
 export interface TriageResult {
     passed: boolean;
     results: CheckerResult[];