fix(workspace): honor SUTANDO_WORKSPACE in 5 TS sites — round 3 audit, first TS slice (issue #839) by qingyun-wu · Pull Request #842 · sonichi/sutando

qingyun-wu · 2026-05-18T19:20:57Z

Summary

First TS slice of the workspace audit followup filed as #839. 5 bug sites across 2 files, same class as #836 / #833 / #838 (Python side, all merged today): code paths resolving runtime data against process.env.WORKSPACE_DIR (wrong env var name — should be SUTANDO_WORKSPACE) or process.cwd() instead of the canonical resolveWorkspace() helper introduced in #821.

Sites fixed

File	Site	Pre-fix behavior
`src/cartesia-tts.ts:29`	`getWorkspace`	`WORKSPACE_DIR` env (legacy, not set) → `process.cwd()` fallback. TTS audio files (`results/audio/<category>/<label>-<ts>.wav`) written to whatever cwd the launcher had.
`src/inline-tools.ts:471, 472`	`cancel_task`	`tasksDir`/`resultsDir` looked in `<cwd>/tasks/`. Voice-initiated cancels silently failed to find in-flight tasks on hosts with `SUTANDO_WORKSPACE` set — bridge writes to workspace, cancel_task was reading repo.
`src/inline-tools.ts:775`	`show_view`	`dynamic-content.json` written to `<cwd>/`. Sutando.app (post-#837) reads from `\$SUTANDO_WORKSPACE/`. Voice toggle fired but the app never saw it.
`src/inline-tools.ts:873`	`recent_context`	`state/voice-session-context.json` looked in `<cwd>/state/`. Core writer (CLAUDE.md "Voice session context") writes to `\$SUTANDO_WORKSPACE/state/`. Voice-agent's deictic-pronoun recovery returned "no context recorded yet" even when populated.
`src/inline-tools.ts:772`	`NOTES_DIR`	`sharedPersonalPath('notes', process.cwd())` passed cwd as the workspace arg to the helper, so the fallback inside the helper resolved to `<cwd>/notes/` instead of `\$SUTANDO_WORKSPACE/notes/`. read_note / search_notes looked in the wrong place.

Pattern (matches #821 + #836)

import { resolveWorkspace } from './workspace_default.js';
const WORKSPACE_DIR = resolveWorkspace();

Module-init resolution. Single source of truth per module. Mirrors the Python WORKSPACE_DIR = resolve_workspace() pattern from the Python side of the audit.

Not addressed in this PR (still in #839's backlog)

src/voice-agent.ts:97, 103 — WORKSPACE_DIR (the canonical voice-agent reference) + CALL_RESULTS_DIR. Highest-traffic sites, used across ~10 call sites in voice-agent.ts (PIDFILE, tasks/, results/, briefings, conversation.log). Worth a focused review — saved for the next PR.
src/util_paths.ts:26, 54 — the workspace ?? process.cwd() fallback inside the helper. Cleaner to fix every caller first (this PR is one step toward that) and then either delete the fallback or switch it to resolveWorkspace() so the helper has a sensible default.
src/inline-tools.ts:957, 1021 — join(process.cwd(), 'skills'). These are CODE-adjacent paths (skills ship with the code), so they need REPO_ROOT semantics, not WORKSPACE_DIR. Latent today (the launcher always cwd's to the repo) but worth fixing.
src/mentra-bridge.ts:88 — file is currently untracked (WIP per the commit history).

Test plan

npx tsc --noEmit — clean.
Diff is targeted: 2 files, 22+/-8.
CI on PR.

Followups

After this lands, 3 of the 4 followup buckets in #839 remain (voice-agent core, util_paths fallback, skills-path REPO_ROOT semantics). Once all are addressed, #839 can close.

Audit running total: 18 bug sites across 10 files (#815 + #836 + #833 + #838 + this PR), in the workspace-vs-repo class.

…ols sites (issue #839) Round-3 audit followup to #836/#833/#838 — first slice of the TS audit flagged in #839. Five bug sites across two files, all hitting the same class: code paths resolving workspace data against `process.env.WORKSPACE_DIR` (wrong env var name) or `process.cwd()` instead of `resolveWorkspace()`. **Sites fixed:** - `src/cartesia-tts.ts:29` (`getWorkspace`) — TTS audio output dir (`results/audio/<category>/<label>-<ts>.wav`). Pre-fix: `WORKSPACE_DIR` env (legacy name, not set) → `process.cwd()` fallback → audio written to whatever cwd the launcher had. Now resolves to `$SUTANDO_WORKSPACE`. - `src/inline-tools.ts:471, 472` (`cancel_task` tool) — `tasksDir` and `resultsDir` for the cancel-task lookup. The tool listed tasks from `<cwd>/tasks/` and wrote cancel results to `<cwd>/results/`. On hosts with `SUTANDO_WORKSPACE` set, the bridge writes tasks to the workspace but cancel_task was looking in the repo (or whatever cwd) — so voice- initiated cancels would never find an in-flight task and would silently write a stranded CANCEL_INSTRUCTION file. - `src/inline-tools.ts:775` (`show_view` tool) — `dynamic-content.json` for the menu-bar app's view toggle. Pre-fix the inline tool wrote to `<cwd>/dynamic-content.json` while Sutando.app (post-#837) reads from `$SUTANDO_WORKSPACE/dynamic-content.json`. Split-brain: voice toggle fired but the app never saw it on hosts with `SUTANDO_WORKSPACE`. - `src/inline-tools.ts:873` (`recent_context` tool) — voice-agent state context. Pre-fix the read path looked at `<cwd>/state/voice-session-context.json` but the writer (core via CLAUDE.md "Voice session context" section) writes to `$SUTANDO_WORKSPACE/state/`. Voice-agent's deictic-pronoun recovery ("what was the post?", "what's pending?") would return `{note: "no context recorded yet"}` even when the writer had populated the file. - `src/inline-tools.ts:772` (`NOTES_DIR`) — `sharedPersonalPath('notes', process.cwd())` passes cwd as the workspace argument to the helper. When `SUTANDO_PRIVATE_DIR` isn't set, the helper falls through to `join(workspace, 'notes')` which becomes `<cwd>/notes/` instead of `$SUTANDO_WORKSPACE/notes/`. read_note / search_notes silently look in the wrong place. **Pattern (matches #821):** ```ts import { resolveWorkspace } from './workspace_default.js'; const WORKSPACE_DIR = resolveWorkspace(); ``` Module-init resolution mirrors the Python `WORKSPACE_DIR = resolve_workspace()` canonical pattern. Single source of truth per module. **Not addressed in this PR (still in #839's backlog):** - `src/voice-agent.ts:97, 103` — `WORKSPACE_DIR`/`CALL_RESULTS_DIR`. The highest-traffic sites; defer to a focused review. - `src/util_paths.ts:26, 54` — the fallback inside the helper itself. Cleaner to fix every caller first, then either delete the fallback or switch it to `resolveWorkspace()` so the helper has a sensible default. - `src/inline-tools.ts:957, 1021` — `join(process.cwd(), 'skills')`. These are CODE-adjacent paths (skills ship with the code), so they need REPO_ROOT semantics, not WORKSPACE_DIR. Latent today because the launcher always cwd's to the repo, but worth a separate PR. - `src/mentra-bridge.ts:88` — file is currently untracked (WIP). **Verified:** - `npx tsc --noEmit` ✓ clean across the whole tree. - Diff is targeted: 2 files, 22+/-8.

sonichi

LGTM — round-3 TS audit, clean.

5 migrations checked:

src/cartesia-tts.ts getWorkspace() — drops the WORKSPACE_DIR env var alias + process.cwd() fallback. resolveWorkspace() consolidates. ✓
src/inline-tools.ts cancelTaskTool tasksDir/resultsDir — was process.cwd()/tasks, now WORKSPACE_DIR/tasks. ✓ Same drift class as PR #821 callsites.
src/inline-tools.ts NOTES_DIR — sharedPersonalPath('notes', process.cwd()) → sharedPersonalPath('notes', WORKSPACE_DIR). Note the docstring update too — "else cwd/notes" → "else /notes". Good attention to comment-rot.
src/inline-tools.ts dynamic-content.json — was repo-root drop. Now workspace. ✓
src/inline-tools.ts VOICE_SESSION_CONTEXT_PATH — same shape. ✓

Bug-citation comment is the right shape: explains why process.cwd() "happened to match the workspace" pre-fix only when voice-agent was launched from the repo with SUTANDO_WORKSPACE unset. That's the silent-divergence story.

Non-blocking nit: the docstring says "first TS slice (issue #839)" — if there are more TS sites worth migrating, worth listing them in #839 explicitly so the round-N work is enumerable. Up to you.

Both checks green (tsc+tests SUCCESS, license/cla SUCCESS — second is empty in rollup but PR view shows it). MERGEABLE. Ship.

…owup) Migrates voice-agent.ts's `WORKSPACE_DIR` + `CALL_RESULTS_DIR` constants to `resolveWorkspace()` (the canonical TS helper introduced in #821). Two const definitions cover ~15 call sites — all of which use the path for per-user runtime state (tasks/, results/, state/, conversation.log, briefings, audio, dynamic-content, PIDFILE, etc.). **Why this matters:** On hosts with `SUTANDO_WORKSPACE` set (e.g. this machine, where `SUTANDO_WORKSPACE=/Users/qingyunwu/Documents/github/sutando-workspace-001`), the pre-fix behavior was: 1. `WORKSPACE_DIR = process.env.WORKSPACE_DIR || <repo-path-via-import.meta.url>` - `WORKSPACE_DIR` env var was the *legacy* name (now `SUTANDO_WORKSPACE`). - Fallback resolved to the repo checkout, so voice-agent wrote tasks, results, conversation.log, briefings, the PIDFILE, audio dirs, etc. to the **repo**, while Sutando.app (post-#837), Python bridges (post-#762), and #842 inline-tools (just landed) all wrote/read from the **workspace**. 2. `CALL_RESULTS_DIR = <repo>/results/calls` — phone-call result records landed under the repo too. Net effect: split-brain on `SUTANDO_WORKSPACE`-enabled hosts. Phone-call `latest-result.json` lived under the repo; new tasks the voice-agent generated landed there too; bridges/Sutando.app couldn't see them. **Pattern (matches #821 / #836 / #842):** ```ts import { resolveWorkspace } from './workspace_default.js'; const WORKSPACE_DIR = resolveWorkspace(); ``` Module-init resolution, single source of truth. CALL_RESULTS_DIR becomes `join(WORKSPACE_DIR, 'results', 'calls')` — same workspace, just the subpath. **Doc-comment cleanup:** - Renamed env-doc at line 19 from `WORKSPACE_DIR` (legacy) to `SUTANDO_WORKSPACE` (canonical) with a pointer to workspace_default.ts. - Removed the stale "default to sutando/ so Claude Code subprocess picks up CLAUDE.md automatically" comment at line 96. Voice-agent doesn't spawn Claude Code anymore (`grep -n claude src/voice-agent.ts` is empty) — task-bridge handles that via the file pipeline. The dual-use rationale that conflated repo + workspace into one var is obsolete. **Verified:** - `npx tsc --noEmit` ✓ clean. - Diff is targeted: 1 file, 15+/-4. - All 15 existing call sites (lines 98, 142, 334, 366, 515, 547, 551, 553, 799, 985, 1110, 1111, 1113, 1299, 1302, 1387) keep their `WORKSPACE_DIR` reference — they automatically pick up the new resolution. No site-by-site edits required. **Operational note:** This change takes effect on the next voice-agent restart. Currently-running voice-agent processes keep their old paths until restart. No mid-flight disruption.

… cwd (#839) Round-3 audit followup — `personalPath()` and `sharedPersonalPath()` both defaulted the optional `workspace` parameter to `process.cwd()` when the caller didn't pass one. That's the same cwd-as-workspace antipattern the rest of the audit family is fixing: it produces the right path only when the caller's cwd happens to be the workspace dir. This PR changes the default to `resolveWorkspace()` (the canonical helper introduced in #821). Callers that DO pass a workspace are unchanged. **Why this is the last util_paths.ts piece of the audit:** Most callers in src/ already pass `WORKSPACE_DIR` explicitly after the #842 / #843 migration: - src/voice-agent.ts:345 — passes WORKSPACE_DIR (resolveWorkspace post-#843) - src/inline-tools.ts:764 (post-#842) — passes WORKSPACE_DIR (resolveWorkspace) What this fix catches is the still-non-passing callers: - src/voice-agent.ts:556, 583 — `personalPath('stand-identity.json')` with no workspace arg. Falls through to the default. Plus any future caller that forgets to pass workspace — the default is now sensible. **Verified:** - `npx tsc --noEmit` ✓ clean. - Diff is targeted: 1 file, 8+/-3.

…owup) Migrates voice-agent.ts's `WORKSPACE_DIR` + `CALL_RESULTS_DIR` constants to `resolveWorkspace()` (the canonical TS helper introduced in #821). Two const definitions cover ~15 call sites — all of which use the path for per-user runtime state (tasks/, results/, state/, conversation.log, briefings, audio, dynamic-content, PIDFILE, etc.). **Why this matters:** On hosts with `SUTANDO_WORKSPACE` set (e.g. this machine, where `SUTANDO_WORKSPACE=/Users/qingyunwu/Documents/github/sutando-workspace-001`), the pre-fix behavior was: 1. `WORKSPACE_DIR = process.env.WORKSPACE_DIR || <repo-path-via-import.meta.url>` - `WORKSPACE_DIR` env var was the *legacy* name (now `SUTANDO_WORKSPACE`). - Fallback resolved to the repo checkout, so voice-agent wrote tasks, results, conversation.log, briefings, the PIDFILE, audio dirs, etc. to the **repo**, while Sutando.app (post-#837), Python bridges (post-#762), and #842 inline-tools (just landed) all wrote/read from the **workspace**. 2. `CALL_RESULTS_DIR = <repo>/results/calls` — phone-call result records landed under the repo too. Net effect: split-brain on `SUTANDO_WORKSPACE`-enabled hosts. Phone-call `latest-result.json` lived under the repo; new tasks the voice-agent generated landed there too; bridges/Sutando.app couldn't see them. **Pattern (matches #821 / #836 / #842):** ```ts import { resolveWorkspace } from './workspace_default.js'; const WORKSPACE_DIR = resolveWorkspace(); ``` Module-init resolution, single source of truth. CALL_RESULTS_DIR becomes `join(WORKSPACE_DIR, 'results', 'calls')` — same workspace, just the subpath. **Doc-comment cleanup:** - Renamed env-doc at line 19 from `WORKSPACE_DIR` (legacy) to `SUTANDO_WORKSPACE` (canonical) with a pointer to workspace_default.ts. - Removed the stale "default to sutando/ so Claude Code subprocess picks up CLAUDE.md automatically" comment at line 96. Voice-agent doesn't spawn Claude Code anymore (`grep -n claude src/voice-agent.ts` is empty) — task-bridge handles that via the file pipeline. The dual-use rationale that conflated repo + workspace into one var is obsolete. **Verified:** - `npx tsc --noEmit` ✓ clean. - Diff is targeted: 1 file, 15+/-4. - All 15 existing call sites (lines 98, 142, 334, 366, 515, 547, 551, 553, 799, 985, 1110, 1111, 1113, 1299, 1302, 1387) keep their `WORKSPACE_DIR` reference — they automatically pick up the new resolution. No site-by-site edits required. **Operational note:** This change takes effect on the next voice-agent restart. Currently-running voice-agent processes keep their old paths until restart. No mid-flight disruption.

… cwd (#839) Round-3 audit followup — `personalPath()` and `sharedPersonalPath()` both defaulted the optional `workspace` parameter to `process.cwd()` when the caller didn't pass one. That's the same cwd-as-workspace antipattern the rest of the audit family is fixing: it produces the right path only when the caller's cwd happens to be the workspace dir. This PR changes the default to `resolveWorkspace()` (the canonical helper introduced in #821). Callers that DO pass a workspace are unchanged. **Why this is the last util_paths.ts piece of the audit:** Most callers in src/ already pass `WORKSPACE_DIR` explicitly after the #842 / #843 migration: - src/voice-agent.ts:345 — passes WORKSPACE_DIR (resolveWorkspace post-#843) - src/inline-tools.ts:764 (post-#842) — passes WORKSPACE_DIR (resolveWorkspace) What this fix catches is the still-non-passing callers: - src/voice-agent.ts:556, 583 — `personalPath('stand-identity.json')` with no workspace arg. Falls through to the default. Plus any future caller that forgets to pass workspace — the default is now sensible. **Verified:** - `npx tsc --noEmit` ✓ clean. - Diff is targeted: 1 file, 8+/-3.

…owup) (#843) Migrates voice-agent.ts's `WORKSPACE_DIR` + `CALL_RESULTS_DIR` constants to `resolveWorkspace()` (the canonical TS helper introduced in #821). Two const definitions cover ~15 call sites — all of which use the path for per-user runtime state (tasks/, results/, state/, conversation.log, briefings, audio, dynamic-content, PIDFILE, etc.). **Why this matters:** On hosts with `SUTANDO_WORKSPACE` set (e.g. this machine, where `SUTANDO_WORKSPACE=/Users/qingyunwu/Documents/github/sutando-workspace-001`), the pre-fix behavior was: 1. `WORKSPACE_DIR = process.env.WORKSPACE_DIR || <repo-path-via-import.meta.url>` - `WORKSPACE_DIR` env var was the *legacy* name (now `SUTANDO_WORKSPACE`). - Fallback resolved to the repo checkout, so voice-agent wrote tasks, results, conversation.log, briefings, the PIDFILE, audio dirs, etc. to the **repo**, while Sutando.app (post-#837), Python bridges (post-#762), and #842 inline-tools (just landed) all wrote/read from the **workspace**. 2. `CALL_RESULTS_DIR = <repo>/results/calls` — phone-call result records landed under the repo too. Net effect: split-brain on `SUTANDO_WORKSPACE`-enabled hosts. Phone-call `latest-result.json` lived under the repo; new tasks the voice-agent generated landed there too; bridges/Sutando.app couldn't see them. **Pattern (matches #821 / #836 / #842):** ```ts import { resolveWorkspace } from './workspace_default.js'; const WORKSPACE_DIR = resolveWorkspace(); ``` Module-init resolution, single source of truth. CALL_RESULTS_DIR becomes `join(WORKSPACE_DIR, 'results', 'calls')` — same workspace, just the subpath. **Doc-comment cleanup:** - Renamed env-doc at line 19 from `WORKSPACE_DIR` (legacy) to `SUTANDO_WORKSPACE` (canonical) with a pointer to workspace_default.ts. - Removed the stale "default to sutando/ so Claude Code subprocess picks up CLAUDE.md automatically" comment at line 96. Voice-agent doesn't spawn Claude Code anymore (`grep -n claude src/voice-agent.ts` is empty) — task-bridge handles that via the file pipeline. The dual-use rationale that conflated repo + workspace into one var is obsolete. **Verified:** - `npx tsc --noEmit` ✓ clean. - Diff is targeted: 1 file, 15+/-4. - All 15 existing call sites (lines 98, 142, 334, 366, 515, 547, 551, 553, 799, 985, 1110, 1111, 1113, 1299, 1302, 1387) keep their `WORKSPACE_DIR` reference — they automatically pick up the new resolution. No site-by-site edits required. **Operational note:** This change takes effect on the next voice-agent restart. Currently-running voice-agent processes keep their old paths until restart. No mid-flight disruption.

… cwd (#839) (#844) Round-3 audit followup — `personalPath()` and `sharedPersonalPath()` both defaulted the optional `workspace` parameter to `process.cwd()` when the caller didn't pass one. That's the same cwd-as-workspace antipattern the rest of the audit family is fixing: it produces the right path only when the caller's cwd happens to be the workspace dir. This PR changes the default to `resolveWorkspace()` (the canonical helper introduced in #821). Callers that DO pass a workspace are unchanged. **Why this is the last util_paths.ts piece of the audit:** Most callers in src/ already pass `WORKSPACE_DIR` explicitly after the #842 / #843 migration: - src/voice-agent.ts:345 — passes WORKSPACE_DIR (resolveWorkspace post-#843) - src/inline-tools.ts:764 (post-#842) — passes WORKSPACE_DIR (resolveWorkspace) What this fix catches is the still-non-passing callers: - src/voice-agent.ts:556, 583 — `personalPath('stand-identity.json')` with no workspace arg. Falls through to the default. Plus any future caller that forgets to pass workspace — the default is now sensible. **Verified:** - `npx tsc --noEmit` ✓ clean. - Diff is targeted: 1 file, 8+/-3.

…loses #840 properly) Two stacked issues fixed in one shot. **Production bug:** Python writers migrated core-status.json to $SUTANDO_WORKSPACE in #836 (per CLAUDE.md workspace contract), but TS readers in `src/web-client.ts:readCoreStatus` and `src/inline-tools.ts: getCoreStatusTool` still read it from REPO_ROOT via `import.meta.url`- relative paths. Same silent-divergence shape as #821/#842/#843 fixed for other paths — voice-agent reading "core idle" while Python actually wrote "core working" elsewhere. Fix: both TS readers now `join(WORKSPACE_DIR, 'core-status.json')`. **Test flake (#840):** `tests/agent-state-endpoint.test.ts` and `tests/ get-core-status-tool.test.ts` both wrote to `<REPO_ROOT>/core-status.json` in their fixtures. node:test runs files in parallel by default; the get-core-status-tool test writes `status:'running'` as one of its cases right when agent-state-endpoint is asserting `state:'listening'`. Race manifested as `expected listening, got working` on whichever subtest happened to call effectiveAgentState() during the foreign 'running' write. Real fix is test isolation: each test process creates its own `mkdtempSync` workspace dir and exports `SUTANDO_WORKSPACE=$TEMP_DIR`. The spawned web-client inherits the env; the in-process tool import reads the same temp dir. No more shared state, no more race. PR #841's `beforeEach` reset stayed (still useful for tool-track sequencing within the single describe block), but #840 wasn't fully closed by that — this PR closes it for real. Verified: - agent-state-endpoint.test.ts: 9/9 pass - get-core-status-tool.test.ts: 6/6 pass - Running both together (the failure mode): 15/15 pass - Full TS suite: 173/173 pass Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…loses #840 properly) (#849) Two stacked issues fixed in one shot. **Production bug:** Python writers migrated core-status.json to $SUTANDO_WORKSPACE in #836 (per CLAUDE.md workspace contract), but TS readers in `src/web-client.ts:readCoreStatus` and `src/inline-tools.ts: getCoreStatusTool` still read it from REPO_ROOT via `import.meta.url`- relative paths. Same silent-divergence shape as #821/#842/#843 fixed for other paths — voice-agent reading "core idle" while Python actually wrote "core working" elsewhere. Fix: both TS readers now `join(WORKSPACE_DIR, 'core-status.json')`. **Test flake (#840):** `tests/agent-state-endpoint.test.ts` and `tests/ get-core-status-tool.test.ts` both wrote to `<REPO_ROOT>/core-status.json` in their fixtures. node:test runs files in parallel by default; the get-core-status-tool test writes `status:'running'` as one of its cases right when agent-state-endpoint is asserting `state:'listening'`. Race manifested as `expected listening, got working` on whichever subtest happened to call effectiveAgentState() during the foreign 'running' write. Real fix is test isolation: each test process creates its own `mkdtempSync` workspace dir and exports `SUTANDO_WORKSPACE=$TEMP_DIR`. The spawned web-client inherits the env; the in-process tool import reads the same temp dir. No more shared state, no more race. PR #841's `beforeEach` reset stayed (still useful for tool-track sequencing within the single describe block), but #840 wasn't fully closed by that — this PR closes it for real. Verified: - agent-state-endpoint.test.ts: 9/9 pass - get-core-status-tool.test.ts: 6/6 pass - Running both together (the failure mode): 15/15 pass - Full TS suite: 173/173 pass Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Followup to #849 — same workspace-contract fix for `voice-state.json` that #849 just landed for `core-status.json`. Issue #853 filed explicitly to track this. **Sites migrated:** - `src/voice-agent.ts:853` (writer) — was `writeFileSync('voice-state.json', ...)` (cwd-relative; effectively REPO_ROOT when launched from there). Now `writeFileSync(join(WORKSPACE_DIR, 'voice-state.json'), ...)`. `WORKSPACE_DIR` is already in scope post-#843. - `src/web-client.ts:3137` (reader) — was `new URL('../voice-state.json', import.meta.url)` (REPO_ROOT-relative). Now `join(WORKSPACE_DIR, 'voice-state.json')`. `WORKSPACE_DIR` is already in scope post-#821. **Test isolation update (mirrors #849's pattern):** - `tests/agent-state-endpoint.test.ts` — `VOICE_STATE_PATH` was joined against REPO_ROOT (`dirname(fileURLToPath(import.meta.url))`). Now joined against the per-process `TEMP_WORKSPACE` introduced in #849. The 3 subtests at lines 192/211/230 that write/unlink the file pick up the new path automatically. - Dropped the `savedVoiceState` stash/restore — TEMP_WORKSPACE is fresh per test run + `rmSync`'d on teardown, so no prod-file leak is possible. - Removed now-unused imports: `readFileSync`, `existsSync`, `dirname`, `fileURLToPath`. Net: 32+/-23 in the test file. **Why this bug bites:** Pre-fix, voice-agent's writer used `writeFileSync('voice-state.json', ...)` — a cwd-relative path. On hosts where the bridge / voice-agent was launched from the repo, cwd ≈ REPO_ROOT and both sides aligned. But: - On `$SUTANDO_WORKSPACE`-set hosts, the writer kept writing to repo while the reader (post-#849, post-#842) read from $SUTANDO_WORKSPACE. - On any host where the launcher's cwd drifted from REPO_ROOT (IDE spawn, `npx tsx` from elsewhere), the writer landed in an unexpected location. `dm-result.py`'s voiceConnected query goes through web-client's `/sse-status` → `readVoiceState()` → now uniformly resolves through WORKSPACE_DIR. Same path for the writer. Split-brain closed. **Verified:** - `npx tsc --noEmit` ✓ clean. - Diff is targeted: 3 files, 29 insertions / 23 deletions. **Audit family — TS side**: this is the last known core-state file left at REPO_ROOT. After this lands, all workspace runtime data flows through `$SUTANDO_WORKSPACE` consistently across Python + TS + Swift writers and readers.

…) (#855) Followup to #849 — same workspace-contract fix for `voice-state.json` that #849 just landed for `core-status.json`. Issue #853 filed explicitly to track this. **Sites migrated:** - `src/voice-agent.ts:853` (writer) — was `writeFileSync('voice-state.json', ...)` (cwd-relative; effectively REPO_ROOT when launched from there). Now `writeFileSync(join(WORKSPACE_DIR, 'voice-state.json'), ...)`. `WORKSPACE_DIR` is already in scope post-#843. - `src/web-client.ts:3137` (reader) — was `new URL('../voice-state.json', import.meta.url)` (REPO_ROOT-relative). Now `join(WORKSPACE_DIR, 'voice-state.json')`. `WORKSPACE_DIR` is already in scope post-#821. **Test isolation update (mirrors #849's pattern):** - `tests/agent-state-endpoint.test.ts` — `VOICE_STATE_PATH` was joined against REPO_ROOT (`dirname(fileURLToPath(import.meta.url))`). Now joined against the per-process `TEMP_WORKSPACE` introduced in #849. The 3 subtests at lines 192/211/230 that write/unlink the file pick up the new path automatically. - Dropped the `savedVoiceState` stash/restore — TEMP_WORKSPACE is fresh per test run + `rmSync`'d on teardown, so no prod-file leak is possible. - Removed now-unused imports: `readFileSync`, `existsSync`, `dirname`, `fileURLToPath`. Net: 32+/-23 in the test file. **Why this bug bites:** Pre-fix, voice-agent's writer used `writeFileSync('voice-state.json', ...)` — a cwd-relative path. On hosts where the bridge / voice-agent was launched from the repo, cwd ≈ REPO_ROOT and both sides aligned. But: - On `$SUTANDO_WORKSPACE`-set hosts, the writer kept writing to repo while the reader (post-#849, post-#842) read from $SUTANDO_WORKSPACE. - On any host where the launcher's cwd drifted from REPO_ROOT (IDE spawn, `npx tsx` from elsewhere), the writer landed in an unexpected location. `dm-result.py`'s voiceConnected query goes through web-client's `/sse-status` → `readVoiceState()` → now uniformly resolves through WORKSPACE_DIR. Same path for the writer. Split-brain closed. **Verified:** - `npx tsc --noEmit` ✓ clean. - Diff is targeted: 3 files, 29 insertions / 23 deletions. **Audit family — TS side**: this is the last known core-state file left at REPO_ROOT. After this lands, all workspace runtime data flows through `$SUTANDO_WORKSPACE` consistently across Python + TS + Swift writers and readers. Co-authored-by: Chi Wang <4250911+sonichi@users.noreply.github.com>

save_note resolved NOTES_DIR against the workspace (correct since #842), but never ensured the notes/ subdir exists. On a fresh install — or any workspace without a prior note — the first save_note failed with ENOENT and surfaced to the user as a generic "issue saving the file". Add mkdirSync(NOTES_DIR, { recursive: true }) before writeFileSync so the first save always succeeds. Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…ols sites (issue sonichi#839) (sonichi#842) Round-3 audit followup to sonichi#836/sonichi#833/sonichi#838 — first slice of the TS audit flagged in sonichi#839. Five bug sites across two files, all hitting the same class: code paths resolving workspace data against `process.env.WORKSPACE_DIR` (wrong env var name) or `process.cwd()` instead of `resolveWorkspace()`. **Sites fixed:** - `src/cartesia-tts.ts:29` (`getWorkspace`) — TTS audio output dir (`results/audio/<category>/<label>-<ts>.wav`). Pre-fix: `WORKSPACE_DIR` env (legacy name, not set) → `process.cwd()` fallback → audio written to whatever cwd the launcher had. Now resolves to `$SUTANDO_WORKSPACE`. - `src/inline-tools.ts:471, 472` (`cancel_task` tool) — `tasksDir` and `resultsDir` for the cancel-task lookup. The tool listed tasks from `<cwd>/tasks/` and wrote cancel results to `<cwd>/results/`. On hosts with `SUTANDO_WORKSPACE` set, the bridge writes tasks to the workspace but cancel_task was looking in the repo (or whatever cwd) — so voice- initiated cancels would never find an in-flight task and would silently write a stranded CANCEL_INSTRUCTION file. - `src/inline-tools.ts:775` (`show_view` tool) — `dynamic-content.json` for the menu-bar app's view toggle. Pre-fix the inline tool wrote to `<cwd>/dynamic-content.json` while Sutando.app (post-sonichi#837) reads from `$SUTANDO_WORKSPACE/dynamic-content.json`. Split-brain: voice toggle fired but the app never saw it on hosts with `SUTANDO_WORKSPACE`. - `src/inline-tools.ts:873` (`recent_context` tool) — voice-agent state context. Pre-fix the read path looked at `<cwd>/state/voice-session-context.json` but the writer (core via CLAUDE.md "Voice session context" section) writes to `$SUTANDO_WORKSPACE/state/`. Voice-agent's deictic-pronoun recovery ("what was the post?", "what's pending?") would return `{note: "no context recorded yet"}` even when the writer had populated the file. - `src/inline-tools.ts:772` (`NOTES_DIR`) — `sharedPersonalPath('notes', process.cwd())` passes cwd as the workspace argument to the helper. When `SUTANDO_PRIVATE_DIR` isn't set, the helper falls through to `join(workspace, 'notes')` which becomes `<cwd>/notes/` instead of `$SUTANDO_WORKSPACE/notes/`. read_note / search_notes silently look in the wrong place. **Pattern (matches sonichi#821):** ```ts import { resolveWorkspace } from './workspace_default.js'; const WORKSPACE_DIR = resolveWorkspace(); ``` Module-init resolution mirrors the Python `WORKSPACE_DIR = resolve_workspace()` canonical pattern. Single source of truth per module. **Not addressed in this PR (still in sonichi#839's backlog):** - `src/voice-agent.ts:97, 103` — `WORKSPACE_DIR`/`CALL_RESULTS_DIR`. The highest-traffic sites; defer to a focused review. - `src/util_paths.ts:26, 54` — the fallback inside the helper itself. Cleaner to fix every caller first, then either delete the fallback or switch it to `resolveWorkspace()` so the helper has a sensible default. - `src/inline-tools.ts:957, 1021` — `join(process.cwd(), 'skills')`. These are CODE-adjacent paths (skills ship with the code), so they need REPO_ROOT semantics, not WORKSPACE_DIR. Latent today because the launcher always cwd's to the repo, but worth a separate PR. - `src/mentra-bridge.ts:88` — file is currently untracked (WIP). **Verified:** - `npx tsc --noEmit` ✓ clean across the whole tree. - Diff is targeted: 2 files, 22+/-8. Co-authored-by: Chi Wang <4250911+sonichi@users.noreply.github.com>

… followup) (sonichi#843) Migrates voice-agent.ts's `WORKSPACE_DIR` + `CALL_RESULTS_DIR` constants to `resolveWorkspace()` (the canonical TS helper introduced in sonichi#821). Two const definitions cover ~15 call sites — all of which use the path for per-user runtime state (tasks/, results/, state/, conversation.log, briefings, audio, dynamic-content, PIDFILE, etc.). **Why this matters:** On hosts with `SUTANDO_WORKSPACE` set (e.g. this machine, where `SUTANDO_WORKSPACE=/Users/qingyunwu/Documents/github/sutando-workspace-001`), the pre-fix behavior was: 1. `WORKSPACE_DIR = process.env.WORKSPACE_DIR || <repo-path-via-import.meta.url>` - `WORKSPACE_DIR` env var was the *legacy* name (now `SUTANDO_WORKSPACE`). - Fallback resolved to the repo checkout, so voice-agent wrote tasks, results, conversation.log, briefings, the PIDFILE, audio dirs, etc. to the **repo**, while Sutando.app (post-sonichi#837), Python bridges (post-sonichi#762), and sonichi#842 inline-tools (just landed) all wrote/read from the **workspace**. 2. `CALL_RESULTS_DIR = <repo>/results/calls` — phone-call result records landed under the repo too. Net effect: split-brain on `SUTANDO_WORKSPACE`-enabled hosts. Phone-call `latest-result.json` lived under the repo; new tasks the voice-agent generated landed there too; bridges/Sutando.app couldn't see them. **Pattern (matches sonichi#821 / sonichi#836 / sonichi#842):** ```ts import { resolveWorkspace } from './workspace_default.js'; const WORKSPACE_DIR = resolveWorkspace(); ``` Module-init resolution, single source of truth. CALL_RESULTS_DIR becomes `join(WORKSPACE_DIR, 'results', 'calls')` — same workspace, just the subpath. **Doc-comment cleanup:** - Renamed env-doc at line 19 from `WORKSPACE_DIR` (legacy) to `SUTANDO_WORKSPACE` (canonical) with a pointer to workspace_default.ts. - Removed the stale "default to sutando/ so Claude Code subprocess picks up CLAUDE.md automatically" comment at line 96. Voice-agent doesn't spawn Claude Code anymore (`grep -n claude src/voice-agent.ts` is empty) — task-bridge handles that via the file pipeline. The dual-use rationale that conflated repo + workspace into one var is obsolete. **Verified:** - `npx tsc --noEmit` ✓ clean. - Diff is targeted: 1 file, 15+/-4. - All 15 existing call sites (lines 98, 142, 334, 366, 515, 547, 551, 553, 799, 985, 1110, 1111, 1113, 1299, 1302, 1387) keep their `WORKSPACE_DIR` reference — they automatically pick up the new resolution. No site-by-site edits required. **Operational note:** This change takes effect on the next voice-agent restart. Currently-running voice-agent processes keep their old paths until restart. No mid-flight disruption.

… cwd (sonichi#839) (sonichi#844) Round-3 audit followup — `personalPath()` and `sharedPersonalPath()` both defaulted the optional `workspace` parameter to `process.cwd()` when the caller didn't pass one. That's the same cwd-as-workspace antipattern the rest of the audit family is fixing: it produces the right path only when the caller's cwd happens to be the workspace dir. This PR changes the default to `resolveWorkspace()` (the canonical helper introduced in sonichi#821). Callers that DO pass a workspace are unchanged. **Why this is the last util_paths.ts piece of the audit:** Most callers in src/ already pass `WORKSPACE_DIR` explicitly after the sonichi#842 / sonichi#843 migration: - src/voice-agent.ts:345 — passes WORKSPACE_DIR (resolveWorkspace post-sonichi#843) - src/inline-tools.ts:764 (post-sonichi#842) — passes WORKSPACE_DIR (resolveWorkspace) What this fix catches is the still-non-passing callers: - src/voice-agent.ts:556, 583 — `personalPath('stand-identity.json')` with no workspace arg. Falls through to the default. Plus any future caller that forgets to pass workspace — the default is now sensible. **Verified:** - `npx tsc --noEmit` ✓ clean. - Diff is targeted: 1 file, 8+/-3.

…loses sonichi#840 properly) (sonichi#849) Two stacked issues fixed in one shot. **Production bug:** Python writers migrated core-status.json to $SUTANDO_WORKSPACE in sonichi#836 (per CLAUDE.md workspace contract), but TS readers in `src/web-client.ts:readCoreStatus` and `src/inline-tools.ts: getCoreStatusTool` still read it from REPO_ROOT via `import.meta.url`- relative paths. Same silent-divergence shape as sonichi#821/sonichi#842/sonichi#843 fixed for other paths — voice-agent reading "core idle" while Python actually wrote "core working" elsewhere. Fix: both TS readers now `join(WORKSPACE_DIR, 'core-status.json')`. **Test flake (sonichi#840):** `tests/agent-state-endpoint.test.ts` and `tests/ get-core-status-tool.test.ts` both wrote to `<REPO_ROOT>/core-status.json` in their fixtures. node:test runs files in parallel by default; the get-core-status-tool test writes `status:'running'` as one of its cases right when agent-state-endpoint is asserting `state:'listening'`. Race manifested as `expected listening, got working` on whichever subtest happened to call effectiveAgentState() during the foreign 'running' write. Real fix is test isolation: each test process creates its own `mkdtempSync` workspace dir and exports `SUTANDO_WORKSPACE=$TEMP_DIR`. The spawned web-client inherits the env; the in-process tool import reads the same temp dir. No more shared state, no more race. PR sonichi#841's `beforeEach` reset stayed (still useful for tool-track sequencing within the single describe block), but sonichi#840 wasn't fully closed by that — this PR closes it for real. Verified: - agent-state-endpoint.test.ts: 9/9 pass - get-core-status-tool.test.ts: 6/6 pass - Running both together (the failure mode): 15/15 pass - Full TS suite: 173/173 pass Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…nichi#853) (sonichi#855) Followup to sonichi#849 — same workspace-contract fix for `voice-state.json` that sonichi#849 just landed for `core-status.json`. Issue sonichi#853 filed explicitly to track this. **Sites migrated:** - `src/voice-agent.ts:853` (writer) — was `writeFileSync('voice-state.json', ...)` (cwd-relative; effectively REPO_ROOT when launched from there). Now `writeFileSync(join(WORKSPACE_DIR, 'voice-state.json'), ...)`. `WORKSPACE_DIR` is already in scope post-sonichi#843. - `src/web-client.ts:3137` (reader) — was `new URL('../voice-state.json', import.meta.url)` (REPO_ROOT-relative). Now `join(WORKSPACE_DIR, 'voice-state.json')`. `WORKSPACE_DIR` is already in scope post-sonichi#821. **Test isolation update (mirrors sonichi#849's pattern):** - `tests/agent-state-endpoint.test.ts` — `VOICE_STATE_PATH` was joined against REPO_ROOT (`dirname(fileURLToPath(import.meta.url))`). Now joined against the per-process `TEMP_WORKSPACE` introduced in sonichi#849. The 3 subtests at lines 192/211/230 that write/unlink the file pick up the new path automatically. - Dropped the `savedVoiceState` stash/restore — TEMP_WORKSPACE is fresh per test run + `rmSync`'d on teardown, so no prod-file leak is possible. - Removed now-unused imports: `readFileSync`, `existsSync`, `dirname`, `fileURLToPath`. Net: 32+/-23 in the test file. **Why this bug bites:** Pre-fix, voice-agent's writer used `writeFileSync('voice-state.json', ...)` — a cwd-relative path. On hosts where the bridge / voice-agent was launched from the repo, cwd ≈ REPO_ROOT and both sides aligned. But: - On `$SUTANDO_WORKSPACE`-set hosts, the writer kept writing to repo while the reader (post-sonichi#849, post-sonichi#842) read from $SUTANDO_WORKSPACE. - On any host where the launcher's cwd drifted from REPO_ROOT (IDE spawn, `npx tsx` from elsewhere), the writer landed in an unexpected location. `dm-result.py`'s voiceConnected query goes through web-client's `/sse-status` → `readVoiceState()` → now uniformly resolves through WORKSPACE_DIR. Same path for the writer. Split-brain closed. **Verified:** - `npx tsc --noEmit` ✓ clean. - Diff is targeted: 3 files, 29 insertions / 23 deletions. **Audit family — TS side**: this is the last known core-state file left at REPO_ROOT. After this lands, all workspace runtime data flows through `$SUTANDO_WORKSPACE` consistently across Python + TS + Swift writers and readers. Co-authored-by: Chi Wang <4250911+sonichi@users.noreply.github.com>

save_note resolved NOTES_DIR against the workspace (correct since sonichi#842), but never ensured the notes/ subdir exists. On a fresh install — or any workspace without a prior note — the first save_note failed with ENOENT and surfaced to the user as a generic "issue saving the file". Add mkdirSync(NOTES_DIR, { recursive: true }) before writeFileSync so the first save always succeeds. Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

qingyun-wu requested a review from sonichi May 18, 2026 19:21

sonichi reviewed May 18, 2026

View reviewed changes

This was referenced May 18, 2026

fix(voice-agent): honor SUTANDO_WORKSPACE for runtime data (issue #839) #843

Merged

fix(util_paths): default workspace fallback to resolveWorkspace() not cwd (issue #839) #844

Merged

Merge branch 'main' into fix/ts-workspace-audit

7351b4e

sonichi merged commit 12ac8fd into main May 18, 2026
2 checks passed

sonichi deleted the fix/ts-workspace-audit branch May 18, 2026 19:33

This was referenced May 18, 2026

docs(claude-md): split per-session context from on-demand reference (-153 lines) #845

Merged

fix(inline-tools): skill-loader scans REPO_ROOT/skills, not cwd/skills — closes #839 TS audit #848

Merged

sonichi mentioned this pull request May 18, 2026

fix(core-status): TS readers honor SUTANDO_WORKSPACE + isolate tests (closes #840 properly) #849

Merged

4 tasks done

liususan091219 mentioned this pull request May 18, 2026

feat(screen-companion): implement vision_query, take_note, look_up_reference #819

Merged

5 tasks

This was referenced May 18, 2026

voice-state.json: next workspace-contract leak after #849 lands #853

Closed

fix(voice-state): TS reader+writer honor SUTANDO_WORKSPACE (closes #853) #855

Merged

qingyun-wu mentioned this pull request May 18, 2026

lint: ban process.cwd() / Path.cwd() writes outside canonical resolvers #863

Closed

liususan091219 mentioned this pull request May 20, 2026

refactor(workspace): relocate root status files into state/, conversation.log into logs/ #940

Merged

jsun-m mentioned this pull request May 21, 2026

fix(notes): mkdir -p the notes dir before save_note write #979

Merged

sonichi mentioned this pull request Oct 16, 2026

feat(skills): tutorial — voice walkthrough + AG2 Space Getting started page #5597

Merged

Navigation Menu

Repository navigation

fix(workspace): honor SUTANDO_WORKSPACE in 5 TS sites — round 3 audit, first TS slice (issue #839) - #842

fix(workspace): honor SUTANDO_WORKSPACE in 5 TS sites — round 3 audit, first TS slice (issue #839)#842
sonichi merged 2 commits into
mainfrom
fix/ts-workspace-audit

qingyun-wu commented May 18, 2026

Uh oh!

sonichi left a comment

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

qingyun-wu commented May 18, 2026

Summary

Sites fixed

Pattern (matches #821 + #836)

Not addressed in this PR (still in #839's backlog)

Test plan

Followups

Uh oh!

sonichi left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants