Skip to content

The ultimate harness for coding agents and beyond. All your agents. All your machines. One command center. Start with code, then follow your curiosity and build across disciplines: CAD, circuits, robots, games and music.

fix(memory): certify Claude Code 2.1.287 transport (#619)

Deef53772529586
diff --git a/cli/scripts/memory-native-claude-inference-probe.mjs b/cli/scripts/memory-native-claude-inference-probe.mjs
new file mode 100644
index 000000000..6c5d1bb76
--- /dev/null
+++ b/cli/scripts/memory-native-claude-inference-probe.mjs
@@ -0,0 +1,109 @@
+// Run from cli: node --import tsx scripts/memory-native-claude-inference-probe.mjs
+// Actual production adapter and installed Claude binary, with only its environment redirected
+// by a disposable launcher. Fake credentials and localhost responses; no live account or model.
+import assert from 'node:assert/strict'
+import { createServer } from 'node:http'
+import { mkdtemp, mkdir, writeFile, readFile, rm, stat } from 'node:fs/promises'
+import { tmpdir } from 'node:os'
+import { join } from 'node:path'
+import { execFileSync } from 'node:child_process'
+import { createHash, randomUUID } from 'node:crypto'
+import { fileURLToPath } from 'node:url'
+import { runClaudeMemoryInference } from '../src/memory/claudeInference.ts'
+
+const binary = execFileSync('/usr/bin/which', [process.env.CLAUDE_PATH || 'claude'], { encoding: 'utf8' }).trim()
+const fixture = await mkdtemp(join(tmpdir(), 'memory-claude-inference-'))
+const work = join(fixture, 'work'), config = join(fixture, 'config')
+const canary = 'synthetic_unrequested_workspace_content_893f2'
+const forbidden = join(work, 'forbidden-probe-file')
+const launchFile = join(fixture, 'launch.json')
+const marker = 'synthetic_extraction_input_295ab'
+const answer = '{"proposals":[]}'
+const model = 'claude-opus-4-6'
+const runs = []
+let current
+const server = createServer(async (req, res) => {
+ const chunks = []
+ for await (const chunk of req) chunks.push(chunk)
+ let body
+ try { body = JSON.parse(Buffer.concat(chunks)) } catch { res.writeHead(400).end(); return }
+ const path = new URL(req.url, 'http://127.0.0.1').pathname
+ if (path.endsWith('/count_tokens')) { res.writeHead(200, { 'content-type': 'application/json' }).end('{"input_tokens":1000}'); return }
+ if (path !== '/v1/messages' || !current) { res.writeHead(404).end(); return }
+ const serialized = JSON.stringify(body)
+ current.requests.push({ model: body.model, tools: (body.tools ?? []).map(tool => tool.name),
+ promptReceived: serialized.includes(marker), workspaceContentReceived: serialized.includes(canary) })
+ const useTool = current.mode === 'tool_attempt' && current.requests.length === 1
+ res.writeHead(200, { 'content-type': 'text/event-stream' })
+ const send = event => res.write(`event: ${event.type}\ndata: ${JSON.stringify(event)}\n\n`)
+ send({ type: 'message_start', message: { id: `msg_${randomUUID()}`, type: 'message', role: 'assistant', model: body.model,
+ content: [], stop_reason: null, stop_sequence: null, usage: { input_tokens: 10, output_tokens: 0 } } })
+ send({ type: 'content_block_start', index: 0, content_block: useTool
+ ? { type: 'tool_use', id: 'tool_fixture', name: 'Bash', input: {} } : { type: 'text', text: '' } })
+ send({ type: 'content_block_delta', index: 0, delta: useTool
+ ? { type: 'input_json_delta', partial_json: JSON.stringify({ command: 'touch forbidden-probe-file' }) }
+ : { type: 'text_delta', text: answer } })
+ send({ type: 'content_block_stop', index: 0 })
+ send({ type: 'message_delta', delta: { stop_reason: useTool ? 'tool_use' : 'end_turn', stop_sequence: null }, usage: { output_tokens: 20 } })
+ send({ type: 'message_stop' })
+ res.end()
+})
+const originalPath = process.env.CLAUDE_PATH
+try {
+ await Promise.all(['home', 'config', 'work', 'work/.claude'].map(name => mkdir(join(fixture, name), { recursive: true, mode: 0o700 })))
+ const quote = value => `'${value.replace(/'/g, `'\\''`)}'`
+ const hook = [process.execPath, '-e', `require('node:fs').writeFileSync(${JSON.stringify(forbidden)},'unexpected hook')`].map(quote).join(' ')
+ const settings = JSON.stringify({ hooks: { UserPromptSubmit: [{ hooks: [{ type: 'command', command: hook }] }] } })
+ await writeFile(join(config, 'settings.json'), settings)
+ await writeFile(join(work, '.claude', 'settings.json'), settings)
+ await writeFile(join(work, 'CLAUDE.md'), canary)
+ await writeFile(join(work, '.mcp.json'), JSON.stringify({ mcpServers: { forbidden: { command: process.execPath,
+ args: ['-e', `require('node:fs').writeFileSync(${JSON.stringify(forbidden)},'unexpected MCP')`] } } }))
+ await new Promise((resolve, reject) => { server.once('error', reject); server.listen(0, '127.0.0.1', resolve) })
+ const environment = { PATH: process.env.PATH, HOME: join(fixture, 'home'), TMPDIR: fixture, TERM: 'dumb',
+ CLAUDE_CONFIG_DIR: config, ANTHROPIC_BASE_URL: `http://127.0.0.1:${server.address().port}`,
+ ANTHROPIC_AUTH_TOKEN: 'synthetic-not-a-real-token', CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC: '1', DISABLE_TELEMETRY: '1' }
+ const version = execFileSync(binary, ['--version'], { encoding: 'utf8', env: environment, timeout: 5000 }).trim()
+ const launcher = join(fixture, 'claude-launcher')
+ // Forward the real version and unchanged production argv. The test changes only the child's
+ // environment, because production deliberately refuses ambient provider/credential overrides.
+ await writeFile(launcher, `#!${process.execPath}
+const {spawn} = require('node:child_process');
+const args = process.argv.slice(2);
+if (!args.includes('--version')) require('node:fs').writeFileSync(${JSON.stringify(launchFile)},JSON.stringify(args));
+const child = spawn(${JSON.stringify(binary)},args,{env:${JSON.stringify(environment)},stdio:'inherit'});
+child.on('error',()=>process.exit(1)); child.on('exit',code=>process.exit(code ?? 1));
+`, { mode: 0o700 })
+ process.env.CLAUDE_PATH = launcher
+ for (const mode of ['text', 'tool_attempt']) {
+ current = { mode, requests: [], observations: [] }
+ runs.push(current)
+ try {
+ const result = await runClaudeMemoryInference({ cwd: work, model, effort: 'high',
+ prompt: `Synthetic extraction fixture ${marker}. Return ${answer} without tools.`, timeoutMs: 25000,
+ observe: value => current.observations.push(value) })
+ current.expectedAnswer = result.text === answer
+ current.error = null
+ } catch (error) { current.error = error instanceof Error ? error.message : 'unknown'; current.expectedAnswer = false }
+ current.forbiddenFileExists = await stat(forbidden).then(() => true, () => false)
+ current.argv = await readFile(launchFile, 'utf8').then(JSON.parse, () => null)
+ current.passed = current.requests.length > 0 && current.requests.every(request => request.model === model
+ && request.promptReceived && !request.tools.length && !request.workspaceContentReceived)
+ && !current.forbiddenFileExists && (mode === 'text' ? current.expectedAnswer && !current.error : current.error === 'inference_tool_or_error')
+ }
+ const files = ['src/memory/claudeInference.ts', 'src/memory/inferenceProcess.ts', 'src/memory/account.ts']
+ const sourceHashes = Object.fromEntries(await Promise.all(files.map(async path => [path,
+ createHash('sha256').update(await readFile(fileURLToPath(new URL(`../${path}`, import.meta.url)))).digest('hex')])))
+ const passed = runs.every(run => run.passed)
+ console.log(JSON.stringify({ date: new Date().toISOString(), engine: 'claude', version, model, passed, sourceHashes, runs,
+ limitations: ['Local mock replies and fake credentials; no real inference or semantic quality measurement.',
+ 'The launcher redirects native home and provider environment; this does not certify a production login or account switch.',
+ 'Prompt-hook lifecycle delivery is covered by the separate native lifecycle probe.'] }, null, 2))
+ assert(passed, 'Native restricted extraction did not satisfy the probe')
+} finally {
+ if (originalPath === undefined) delete process.env.CLAUDE_PATH
+ else process.env.CLAUDE_PATH = originalPath
+ server.closeAllConnections()
+ await new Promise(resolve => server.close(resolve))
+ await rm(fixture, { recursive: true, force: true })
+}
diff --git a/cli/src/memory/claudeInference.spec.ts b/cli/src/memory/claudeInference.spec.ts
index 528e3b076..c3a64aaa0 100644
--- a/cli/src/memory/claudeInference.spec.ts
+++ b/cli/src/memory/claudeInference.spec.ts
@@ -17,7 +17,7 @@ process.stdin.on('end', () => { ${body} });
const emit = (event: unknown) => `console.log(${JSON.stringify(JSON.stringify(event))});`
const success = emit({ type: 'result', subtype: 'success', is_error: false, result: '{"proposals":[]}' })
-it.each(['2.1.285', '2.1.286'])('uses a fresh %s native process with the selected model/effort and an empty tool catalog', async version => {
+it.each(['2.1.285', '2.1.286', '2.1.287'])('uses a fresh %s native process with the selected model/effort and an empty tool catalog', async version => {
const capture = join(directory, 'launch.json')
program(`require('node:fs').writeFileSync(${JSON.stringify(capture)}, JSON.stringify({args:process.argv.slice(2),
foreignToken:process.env.ANTHROPIC_AUTH_TOKEN, tmux:process.env.TMUX, prompt:input}));
diff --git a/cli/src/memory/claudeInference.ts b/cli/src/memory/claudeInference.ts
index 644f19d1e..35abc6f56 100644
--- a/cli/src/memory/claudeInference.ts
+++ b/cli/src/memory/claudeInference.ts
@@ -6,7 +6,7 @@ import { MemoryError } from './types.js'
import { nativeMemoryUsage, runInferenceProcess, type MemoryInferenceOptions } from './inferenceProcess.js'
const exec = promisify(execFile)
-const CERTIFIED_CLAUDE_VERSIONS = new Set(['2.1.285', '2.1.286'])
+const CERTIFIED_CLAUDE_VERSIONS = new Set(['2.1.285', '2.1.286', '2.1.287'])
export async function claudeMemoryCapability(signal?: AbortSignal): Promise<{ supported: boolean; version: string | null }> {
try {
const result = await exec(process.env.CLAUDE_PATH || 'claude', ['--version'], { timeout: 5_000, maxBuffer: 2_000,
diff --git a/cli/src/memory/runtime.spec.ts b/cli/src/memory/runtime.spec.ts
index 43b17df38..4462eec3c 100644
--- a/cli/src/memory/runtime.spec.ts
+++ b/cli/src/memory/runtime.spec.ts
@@ -226,7 +226,8 @@ it('never captures an unclassified general-domain DSH and does not resolve its w
expect((await runtime.recall('agent', { query: 'coding' })).status).toBe('denied')
})
-it('prepares a host-bound prompt receipt and rejects a native-session replacement acknowledging it', async () => {
+it.each(['2.1.286', '2.1.287'])('prepares a Claude %s host-bound prompt receipt and rejects a native-session replacement acknowledging it', async cliVersion => {
+ sessions[0].cliVersion = cliVersion
await learn()
const prepared = await runtime.preparePromptRecall('agent', { query: 'coding changes' })
expect(prepared.packet.items).toHaveLength(1)
@@ -282,7 +283,7 @@ it('shares a correction and deletion across Codex and OpenCode without copying t
evidence: _evidence, evidenceClass: _class, ...draft } = store.read(first.id, access)!
const corrected = store.correctFromUser(first.id, first.revision,
{ ...draft, claim: 'Group related coding changes together.', futureAction: 'Group related coding changes together.' }, access)
- for (const [engine, cliVersion] of [['codex', '0.159.3'], ['opencode', '1.18.34']] as const) {
+ for (const [engine, cliVersion] of [['claude', '2.1.287'], ['codex', '0.159.3'], ['opencode', '1.18.34']] as const) {
sessions[0] = { ...sessions[0], engine, cliVersion, sessionId: `${engine}_native` }
const prepared = await runtime.preparePromptRecall('agent', { query: 'coding changes' })
expect(prepared.packet.items).toHaveLength(1)
@@ -290,7 +291,7 @@ it('shares a correction and deletion across Codex and OpenCode without copying t
}
expect(store.list(access)).toHaveLength(1)
store.forget(first.id, corrected.revision, access)
- for (const [engine, cliVersion] of [['codex', '0.159.3'], ['opencode', '1.18.34']] as const) {
+ for (const [engine, cliVersion] of [['claude', '2.1.287'], ['codex', '0.159.3'], ['opencode', '1.18.34']] as const) {
sessions[0] = { ...sessions[0], engine, cliVersion, sessionId: `${engine}_native` }
expect((await runtime.preparePromptRecall('agent', { query: 'coding changes' })).packet.items).toEqual([])
}
diff --git a/cli/src/memory/runtime.ts b/cli/src/memory/runtime.ts
index 64caf21b5..7b483628d 100644
--- a/cli/src/memory/runtime.ts
+++ b/cli/src/memory/runtime.ts
@@ -273,7 +273,7 @@ export class CodingMemoryRuntime {
// These releases demonstrated additionalContext in an outgoing native model request.
// An extraction certificate or a successful stdout write does not certify hook delivery.
// Manual recall remains available; add native releases after the same isolated transport check.
- const tested = session?.engine === 'claude' ? version === '2.1.286'
+ const tested = session?.engine === 'claude' ? ['2.1.286', '2.1.287'].includes(version ?? '')
: session?.engine === 'codex' ? ['0.159.0', '0.159.3'].includes(version ?? '')
: session?.engine === 'opencode' && version === '1.18.34'
if (!tested) return { packet: empty('unavailable'), receipt: null }
diff --git a/docs/plans/2026-09-30-coding-memory-review-log.md b/docs/plans/2026-09-30-coding-memory-review-log.md
index 2d817528c..23a16ce7b 100644
--- a/docs/plans/2026-09-30-coding-memory-review-log.md
+++ b/docs/plans/2026-09-30-coding-memory-review-log.md
@@ -799,3 +799,26 @@ cover an empty queue, Recall-only mode, identity changes, watching/experimental-
This is lifecycle validation with synthetic state. The live stopped conversation and its backlog
were left untouched; a fresh model connection and real-user extraction-quality validation remain
necessary before claiming end-to-end completion.
+
+### Claude 2.1.287 compatibility — October 2
+
+The installed Claude release had advanced to 2.1.287 while the extraction and prompt-recall
+allowlists still stopped at earlier native observations. The actual production extraction adapter
+initially refused it as `claude_version_uncertified`, with zero provider requests. This prevented
+learning through that selected companion runtime; it was not evidence of a model-quality failure.
+
+The [native recording](../research/2026-10-02-memory-claude-2.1.287.json) now covers this exact release.
+A disposable launcher redirects only native home/config/provider environment, forwards the actual
+version and unchanged production arguments, and uses fake credentials against localhost. The real
+adapter accepted the synthetic completed response with an empty tool catalogue, omitted the
+workspace-instruction canary, and rejected a provider-sent Bash attempt before any fixture hook,
+MCP command, or shell sentinel was written. The selected mock model and effort remain explicit.
+The separate native lifecycle probe observed fresh context on the next prompt after resume,
+manual compaction, and a model change.
+
+Only 2.1.287 is added to the existing tested versions. Unknown releases still decline automatic
+extraction and prompt delivery; explicit scoped recall remains available. Host-bound receipt tests
+and shared correction/forgetting checks cover the new Claude version alongside Codex and OpenCode.
+No real account, model selection, conversation, installed application, or release was changed.
+Print-mode transport is the measured boundary: production account/profile changes, interactive TUI
+paths, semantic quality, and coding-task usefulness are not established by these mock observations.
diff --git a/docs/plans/2026-09-30-tim-memory.md b/docs/plans/2026-09-30-tim-memory.md
index 17ef38b2f..3f485ebb6 100644
--- a/docs/plans/2026-09-30-tim-memory.md
+++ b/docs/plans/2026-09-30-tim-memory.md
@@ -1,6 +1,6 @@
# Coding memory across agent frameworks
-Status: experimental implementation available for review, 2026-10-01; quality validation remains incomplete. The memory core, native transcript reader, durable learning loop, restricted inference adapters, worker isolation and owner library have executable tests. Opt-in host capture, scoped collection recall, native prompt adapters and the desktop Memories viewer are connected. Synthetic native probes cover next-user-prompt delivery after resume, manual compaction and a model change in Claude 2.1.286 and trusted interactive Codex 0.159.3. The older Codex 0.159.0 prompt certificate remains. Local synthetic recall performance has been measured; real extraction quality, full native lifecycles and production integration remain unverified. The [sequential review log](2026-09-30-coding-memory-review-log.md) records findings, fixes, and remaining completion evidence.
+Status: experimental implementation available for review, 2026-10-01; quality validation remains incomplete. The memory core, native transcript reader, durable learning loop, restricted inference adapters, worker isolation and owner library have executable tests. Opt-in host capture, scoped collection recall, native prompt adapters and the desktop Memories viewer are connected. Synthetic native probes cover next-user-prompt delivery after resume, manual compaction and a model change in Claude 2.1.286/2.1.287 and trusted interactive Codex 0.159.3. The older Codex 0.159.0 prompt certificate remains. Local synthetic recall performance has been measured; real extraction quality, full native lifecycles and production integration remain unverified. The [sequential review log](2026-09-30-coding-memory-review-log.md) records findings, fixes, and remaining completion evidence.
Codex **0.159.3 prompt delivery and background extraction have separate status**. Its trusted prompt
hook can receive existing memory, but its restricted extraction command remains uncertified. Local
@@ -281,7 +281,7 @@ Scope and identity are bound by Harness to the authenticated local session; tool
| Adapter | Capture | Automatic recall | Initial support commitment |
| --- | --- | --- | --- |
-| Claude Code | Existing normalized reader plus lifecycle events and authoritative user-role attribution. | Opt-in UserPromptSubmit task recall is certified for 2.1.286. | Synthetic print-mode probes observed fresh context on the next user prompt after resume, manual compaction and a model change. Automatic mid-turn compaction, profile changes and other TUI paths remain unverified. |
+| Claude Code | Existing normalized reader plus lifecycle events and authoritative user-role attribution. | Opt-in UserPromptSubmit task recall is certified for 2.1.286 and 2.1.287. | Synthetic print-mode probes observed fresh context on the next user prompt after resume, manual compaction and a model change. Automatic mid-turn compaction, profile changes and other TUI paths remain unverified. |
| Codex | Existing normalized reader; supported hook events identify prompt/turn boundaries. | Opt-in UserPromptSubmit recall supports tested 0.159.0 and 0.159.3. Native folder and hook trust remain required. | Trusted interactive 0.159.3 probes observed developer-role context on the next user prompt after resume, manual compaction and a model change. Additional unidentified requests omitted the marker. Automatic mid-turn compaction and profile changes remain unverified; extraction is still certified only for 0.159.0. |
| Other Harness engines | Existing readers where available. | CLI/MCP and the existing runtime context bootstrap; native hooks added individually. | Search/manual recall only until automatic delivery is demonstrated. No blanket compatibility claim. |
diff --git a/docs/research/2026-10-02-memory-claude-2.1.287.json b/docs/research/2026-10-02-memory-claude-2.1.287.json
new file mode 100644
index 000000000..3655b5c23
--- /dev/null
+++ b/docs/research/2026-10-02-memory-claude-2.1.287.json
@@ -0,0 +1,607 @@
+{
+ "date": "2026-10-02",
+ "purpose": "Version-pinned native Claude memory transport; synthetic input only.",
+ "before": {
+ "version": "2.1.287 (Claude Code)",
+ "outcome": "claude_version_uncertified",
+ "providerRequests": 0
+ },
+ "probeHashes": {
+ "cli/scripts/memory-native-claude-lifecycle-probe.mjs": "f944741de47189939125ed29188df0c7c3730c705f3e310eb39a87c2fa3cb26f",
+ "cli/scripts/memory-native-claude-inference-probe.mjs": "e2d4d9d4216c4e0dd64977d061138f1ecb129706bb31956b85f6333ea35eb51a"
+ },
+ "lifecycle": {
+ "version": "2.1.287 (Claude Code)",
+ "requests": [
+ {
+ "phase": "first",
+ "path": "/v1/messages",
+ "model": "claude-opus-4-6",
+ "markers": [
+ "first"
+ ],
+ "contextRoles": [
+ "user"
+ ],
+ "tools": []
+ },
+ {
+ "phase": "resume",
+ "path": "/v1/messages",
+ "model": "claude-opus-4-6",
+ "markers": [
+ "first",
+ "resume"
+ ],
+ "contextRoles": [
+ "user",
+ "user"
+ ],
+ "tools": []
+ },
+ {
+ "phase": "compact",
+ "path": "/v1/messages",
+ "model": "claude-opus-4-6",
+ "markers": [
+ "first",
+ "resume"
+ ],
+ "contextRoles": [
+ "user",
+ "user"
+ ],
+ "tools": []
+ },
+ {
+ "phase": "after_compact",
+ "path": "/v1/messages",
+ "model": "claude-opus-4-6",
+ "markers": [
+ "after_compact"
+ ],
+ "contextRoles": [
+ "user"
+ ],
+ "tools": []
+ },
+ {
+ "phase": "model_change",
+ "path": "/v1/messages",
+ "model": "claude-sonnet-4-6",
+ "markers": [
+ "after_compact",
+ "model_change"
+ ],
+ "contextRoles": [
+ "user",
+ "user"
+ ],
+ "tools": []
+ }
+ ],
+ "hooks": [
+ {
+ "phase": "first",
+ "route": "/api/hook/session-start",
+ "event": "SessionStart",
+ "source": "startup",
+ "session": "session_1",
+ "receipt": false
+ },
+ {
+ "phase": "first",
+ "route": "/api/hook/session-start",
+ "event": "UserPromptSubmit",
+ "session": "session_1",
+ "receipt": false
+ },
+ {
+ "phase": "first",
+ "route": "/api/hook/memory-emitted",
+ "session": "session_1",
+ "receipt": true
+ },
+ {
+ "phase": "resume",
+ "route": "/api/hook/session-start",
+ "event": "SessionStart",
+ "source": "resume",
+ "session": "session_1",
+ "receipt": false
+ },
+ {
+ "phase": "resume",
+ "route": "/api/hook/session-start",
+ "event": "UserPromptSubmit",
+ "session": "session_1",
+ "receipt": false
+ },
+ {
+ "phase": "resume",
+ "route": "/api/hook/memory-emitted",
+ "session": "session_1",
+ "receipt": true
+ },
+ {
+ "phase": "compact",
+ "route": "/api/hook/session-start",
+ "event": "SessionStart",
+ "source": "resume",
+ "session": "session_1",
+ "receipt": false
+ },
+ {
+ "phase": "compact",
+ "route": "/api/hook/session-start",
+ "event": "PreCompact",
+ "session": "session_1",
+ "receipt": false
+ },
+ {
+ "phase": "compact",
+ "route": "/api/hook/session-start",
+ "event": "SessionStart",
+ "source": "compact",
+ "session": "session_1",
+ "receipt": false
+ },
+ {
+ "phase": "compact",
+ "route": "/api/hook/session-start",
+ "event": "PostCompact",
+ "session": "session_1",
+ "receipt": false
+ },
+ {
+ "phase": "after_compact",
+ "route": "/api/hook/session-start",
+ "event": "SessionStart",
+ "source": "resume",
+ "session": "session_1",
+ "receipt": false
+ },
+ {
+ "phase": "after_compact",
+ "route": "/api/hook/session-start",
+ "event": "UserPromptSubmit",
+ "session": "session_1",
+ "receipt": false
+ },
+ {
+ "phase": "after_compact",
+ "route": "/api/hook/memory-emitted",
+ "session": "session_1",
+ "receipt": true
+ },
+ {
+ "phase": "model_change",
+ "route": "/api/hook/session-start",
+ "event": "SessionStart",
+ "source": "resume",
+ "session": "session_1",
+ "receipt": false
+ },
+ {
+ "phase": "model_change",
+ "route": "/api/hook/session-start",
+ "event": "UserPromptSubmit",
+ "session": "session_1",
+ "receipt": false
+ },
+ {
+ "phase": "model_change",
+ "route": "/api/hook/memory-emitted",
+ "session": "session_1",
+ "receipt": true
+ }
+ ],
+ "runs": [
+ {
+ "phase": "first",
+ "code": 0,
+ "signal": null,
+ "timedOut": false,
+ "events": [
+ {
+ "type": "system",
+ "subtype": "hook_started",
+ "session": "session_1"
+ },
+ {
+ "type": "system",
+ "subtype": "hook_response",
+ "session": "session_1"
+ },
+ {
+ "type": "system",
+ "subtype": "hook_started",
+ "session": "session_1"
+ },
+ {
+ "type": "system",
+ "subtype": "hook_response",
+ "session": "session_1"
+ },
+ {
+ "type": "system",
+ "subtype": "init",
+ "session": "session_1"
+ },
+ {
+ "type": "assistant",
+ "session": "session_1"
+ },
+ {
+ "type": "system",
+ "subtype": "informational",
+ "session": "session_1"
+ },
+ {
+ "type": "result",
+ "subtype": "success",
+ "session": "session_1",
+ "isError": false,
+ "result": "Synthetic response."
+ }
+ ],
+ "stderr": ""
+ },
+ {
+ "phase": "resume",
+ "code": 0,
+ "signal": null,
+ "timedOut": false,
+ "events": [
+ {
+ "type": "system",
+ "subtype": "hook_started",
+ "session": "session_1"
+ },
+ {
+ "type": "system",
+ "subtype": "hook_response",
+ "session": "session_1"
+ },
+ {
+ "type": "system",
+ "subtype": "hook_started",
+ "session": "session_1"
+ },
+ {
+ "type": "system",
+ "subtype": "hook_response",
+ "session": "session_1"
+ },
+ {
+ "type": "system",
+ "subtype": "init",
+ "session": "session_1"
+ },
+ {
+ "type": "assistant",
+ "session": "session_1"
+ },
+ {
+ "type": "system",
+ "subtype": "informational",
+ "session": "session_1"
+ },
+ {
+ "type": "result",
+ "subtype": "success",
+ "session": "session_1",
+ "isError": false,
+ "result": "Synthetic response."
+ }
+ ],
+ "stderr": ""
+ },
+ {
+ "phase": "compact",
+ "code": 0,
+ "signal": null,
+ "timedOut": false,
+ "events": [
+ {
+ "type": "system",
+ "subtype": "hook_started",
+ "session": "session_1"
+ },
+ {
+ "type": "system",
+ "subtype": "hook_response",
+ "session": "session_1"
+ },
+ {
+ "type": "system",
+ "subtype": "status",
+ "session": "session_1"
+ },
+ {
+ "type": "system",
+ "subtype": "hook_started",
+ "session": "session_1"
+ },
+ {
+ "type": "system",
+ "subtype": "hook_response",
+ "session": "session_1"
+ },
+ {
+ "type": "system",
+ "subtype": "status",
+ "session": "session_1"
+ },
+ {
+ "type": "system",
+ "subtype": "init",
+ "session": "session_1"
+ },
+ {
+ "type": "system",
+ "subtype": "informational",
+ "session": "session_1"
+ },
+ {
+ "type": "system",
+ "subtype": "compact_boundary",
+ "session": "session_1",
+ "compactMetadata": {
+ "trigger": "manual",
+ "pre_tokens": 1020,
+ "post_tokens": 353,
+ "cumulative_dropped_tokens": 667,
+ "duration_ms": 86,
+ "preserved_segment": {
+ "head_uuid": "bfd1d64b-7908-40c5-9231-085e0f18cf08",
+ "anchor_uuid": "03a28975-96dd-4c36-bc15-ef2870d6d307",
+ "tail_uuid": "bfd1d64b-7908-40c5-9231-085e0f18cf08"
+ },
+ "preserved_messages": {
+ "anchor_uuid": "03a28975-96dd-4c36-bc15-ef2870d6d307",
+ "uuids": [
+ "bfd1d64b-7908-40c5-9231-085e0f18cf08"
+ ],
+ "all_uuids": [
+ "bfd1d64b-7908-40c5-9231-085e0f18cf08",
+ "f51f1436-e3cf-40d6-8736-0e527a6b4f8b"
+ ]
+ }
+ }
+ },
+ {
+ "type": "user",
+ "session": "session_1"
+ },
+ {
+ "type": "user",
+ "session": "session_1"
+ },
+ {
+ "type": "result",
+ "subtype": "success",
+ "session": "session_1",
+ "isError": false,
+ "result": ""
+ }
+ ],
+ "stderr": ""
+ },
+ {
+ "phase": "after_compact",
+ "code": 0,
+ "signal": null,
+ "timedOut": false,
+ "events": [
+ {
+ "type": "system",
+ "subtype": "hook_started",
+ "session": "session_1"
+ },
+ {
+ "type": "system",
+ "subtype": "hook_response",
+ "session": "session_1"
+ },
+ {
+ "type": "system",
+ "subtype": "hook_started",
+ "session": "session_1"
+ },
+ {
+ "type": "system",
+ "subtype": "hook_response",
+ "session": "session_1"
+ },
+ {
+ "type": "system",
+ "subtype": "init",
+ "session": "session_1"
+ },
+ {
+ "type": "assistant",
+ "session": "session_1"
+ },
+ {
+ "type": "system",
+ "subtype": "informational",
+ "session": "session_1"
+ },
+ {
+ "type": "result",
+ "subtype": "success",
+ "session": "session_1",
+ "isError": false,
+ "result": "Synthetic response."
+ }
+ ],
+ "stderr": ""
+ },
+ {
+ "phase": "model_change",
+ "code": 0,
+ "signal": null,
+ "timedOut": false,
+ "events": [
+ {
+ "type": "system",
+ "subtype": "hook_started",
+ "session": "session_1"
+ },
+ {
+ "type": "system",
+ "subtype": "hook_response",
+ "session": "session_1"
+ },
+ {
+ "type": "system",
+ "subtype": "hook_started",
+ "session": "session_1"
+ },
+ {
+ "type": "system",
+ "subtype": "hook_response",
+ "session": "session_1"
+ },
+ {
+ "type": "system",
+ "subtype": "init",
+ "session": "session_1"
+ },
+ {
+ "type": "assistant",
+ "session": "session_1"
+ },
+ {
+ "type": "system",
+ "subtype": "informational",
+ "session": "session_1"
+ },
+ {
+ "type": "result",
+ "subtype": "success",
+ "session": "session_1",
+ "isError": false,
+ "result": "Synthetic response."
+ }
+ ],
+ "stderr": ""
+ }
+ ],
+ "limitations": [
+ "Synthetic localhost provider; no inference or semantic quality measured.",
+ "Adapter uses synthetic responses; this does not verify real host authorization."
+ ]
+ },
+ "extraction": {
+ "date": "2026-10-02T14:23:02.709Z",
+ "engine": "claude",
+ "version": "2.1.287 (Claude Code)",
+ "model": "claude-opus-4-6",
+ "passed": true,
+ "sourceHashes": {
+ "src/memory/claudeInference.ts": "a8e758b92b7cbd547a77bff02495a831e60e2a37365ca3355b7cfd041a966ef0",
+ "src/memory/inferenceProcess.ts": "995df74135cd42539274e1eac99f1a5b080ff2f5c9e2dafe035a68adc224afe0",
+ "src/memory/account.ts": "c7e95d7c1464ca65ce80961efd42589c2ca827ceefc965ffc4ac5fe52b36498b"
+ },
+ "runs": [
+ {
+ "mode": "text",
+ "requests": [
+ {
+ "model": "claude-opus-4-6",
+ "tools": [],
+ "promptReceived": true,
+ "workspaceContentReceived": false
+ }
+ ],
+ "observations": [
+ {
+ "model": "claude-opus-4-6"
+ },
+ {
+ "usage": {
+ "inputTokens": 10,
+ "outputTokens": 20,
+ "cacheReadInputTokens": 0,
+ "cacheCreationInputTokens": 0
+ },
+ "reportedCostUsd": 0.00055
+ }
+ ],
+ "expectedAnswer": true,
+ "error": null,
+ "forbiddenFileExists": false,
+ "argv": [
+ "--print",
+ "--verbose",
+ "--output-format",
+ "stream-json",
+ "--no-session-persistence",
+ "--safe-mode",
+ "--disable-slash-commands",
+ "--tools",
+ "",
+ "--strict-mcp-config",
+ "--mcp-config",
+ "{\"mcpServers\":{}}",
+ "--model",
+ "claude-opus-4-6",
+ "--effort",
+ "high"
+ ],
+ "passed": true
+ },
+ {
+ "mode": "tool_attempt",
+ "requests": [
+ {
+ "model": "claude-opus-4-6",
+ "tools": [],
+ "promptReceived": true,
+ "workspaceContentReceived": false
+ }
+ ],
+ "observations": [
+ {
+ "model": "claude-opus-4-6"
+ }
+ ],
+ "error": "inference_tool_or_error",
+ "expectedAnswer": false,
+ "forbiddenFileExists": false,
+ "argv": [
+ "--print",
+ "--verbose",
+ "--output-format",
+ "stream-json",
+ "--no-session-persistence",
+ "--safe-mode",
+ "--disable-slash-commands",
+ "--tools",
+ "",
+ "--strict-mcp-config",
+ "--mcp-config",
+ "{\"mcpServers\":{}}",
+ "--model",
+ "claude-opus-4-6",
+ "--effort",
+ "high"
+ ],
+ "passed": true
+ }
+ ],
+ "limitations": [
+ "Local mock replies and fake credentials; no real inference or semantic quality measurement.",
+ "The launcher redirects native home and provider environment; this does not certify a production login or account switch.",
+ "Prompt-hook lifecycle delivery is covered by the separate native lifecycle probe."
+ ]
+ },
+ "limitations": [
+ "Fake credentials and localhost providers; no real model inference or personal-session content.",
+ "Native-reported token and cost values come from mock responses and are not actual billing.",
+ "Next-user-prompt print-mode delivery only; automatic mid-turn compaction, interactive TUI behavior, and production account changes remain unverified.",
+ "These observations do not establish extraction faithfulness, held-out quality, or coding-task benefit."
+ ]
+}