mirror of
https://github.com/nexu-io/open-design.git
synced 2026-05-31 19:04:39 +07:00
4 commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
lefarcen
|
98a2c63973
|
feat(daemon): add Antigravity agent adapter (#3157)
* feat(daemon): add Antigravity agent adapter
Adds Google Antigravity (`agy` CLI) as a coding-agent runtime. Detection
picks up `agy` on PATH, the daemon spawns `agy -p "<prompt>"` for a
single non-interactive turn, and the assistant text reply streams back
on stdout. OAuth is shared with the Antigravity IDE through the system
keyring, so users who have signed into the desktop app are authenticated
on first run with no extra step.
`agy` v1.0.3 has no JSON / stream-json / ACP output mode (upstream issue
#119), no `--model` flag (issue #35), and no MCP forwarding hook yet —
the adapter ships with `streamFormat: 'plain'` and a single `default`
fallback model so the model picker doesn't mislead users into thinking
their choice is wired through. We will upgrade buildArgs + add a
dedicated event parser when upstream ships structured output.
Also gitignores `.antigravitycli/`, the project-local config directory
`agy` auto-creates on every run (upstream issue #175).
* fix(daemon): Antigravity adapter — stdin prompt, brand icon, form loop, empty-output guard
- Switch prompt delivery from argv to stdin (`agy -p -`) to avoid the
30KB maxPromptArgBytes limit that blocked real-world composed prompts
- Add official Antigravity brand SVG icon to agent picker
- Fix repeated question-form loop for plain agents by injecting an
OVERRIDE block when form answers are already present in the transcript
- Add empty-output guard for plain agents so expired auth or silent
failures surface a user-visible error instead of a blank "Done" turn
* feat(daemon): expand Antigravity adapter — model picker, form-loop fix, OAuth launcher, log-file classification
PR #3157 follow-up integrating four iterations from end-to-end manual
testing on Gemini 3.5 Flash + GPT-OSS 120B Medium through `agy` v1.0.3.
Each section is independently verifiable; combined they're what made
the first successful artifact generation work end-to-end.
## Model picker via settings.json (agy has no --model flag)
agy v1.0.3 ships no `--model` CLI flag (upstream issue #35), but the
TUI Switch-Model picker writes the chosen label to
`~/.gemini/antigravity-cli/settings.json`'s `"model"` field, and every
`-p` invocation re-reads that file on startup — verified by capturing
the `--log-file` line `Propagating selected model override to backend:
label="<model>"`. Antigravity's `fallbackModels` now lists the 8
labels its TUI exposes (Gemini 3.1 Pro / 3.5 Flash variants, Claude
Sonnet/Opus 4.6 Thinking, GPT-OSS 120B Medium) and `buildArgs`
persists the user's choice to settings.json right before spawn. The
synthetic `default` id is preserved — picking it leaves settings.json
untouched so a user who switches models from agy's own TUI keeps
their choice.
Introduces `RuntimeAgentDef.supportsCustomModel?: boolean`. AMR's
hardcoded blocklist in `SettingsDialog.tsx` migrates to the
declarative flag (it rejects free-form ids at the ACP layer), and
antigravity opts out because its label set is a server-side enum that
silently fails on unrecognised strings.
## Form-loop fix (transcript sanitizer + stronger OVERRIDE)
The discovery form loop on weak/medium plain-stream models (GPT-OSS
120B Medium, Gemini 3.5 Flash) had two reinforcing causes:
1. `buildDaemonTranscript` packed the prior assistant turn's
literal `<question-form>` markup into the user request on the
next turn, giving the model a template to echo. New
`sanitizePriorAssistantTurnForTranscript` strips
`<question-form>...</question-form>` blocks and ```json fences
that match form-schema shape, replacing them with a brief
placeholder. User content is preserved verbatim (a user who
legitimately mentions `<question-form>` in chat keeps their
message intact).
2. The OVERRIDE block on form-answered turns was 4 lines and only
banned the bare `<question-form>` tag — models still emitted the
fenced JSON, form-asking prose ("Got it — tell me the following"),
and fake system events ("subagents stopped"). The new
`FORM_ANSWERED_SYSTEM_OVERRIDE` enumerates each anti-pattern and
pins them via tests, so silently weakening any line reintroduces
the regression.
Also adds RuntimeAgentDef.resumesSessionViaCli + RuntimeContext.
hasPriorAssistantTurn as forward-looking abstractions (skipTranscript
option on composeChatUserRequestForAgent). Antigravity does NOT opt
in — agy's `-c` resume activates an internal agentic loop with tool
retries and fallback-to-cached-response on tool errors that the OD
system prompt cannot steer; reverted after seeing byte-identical
form re-emissions caused by agy's own retry logic, not OD's transcript.
## One-click OAuth via system terminal
agy print mode can't complete Google Sign-In on its own (the OAuth
callback page asks the user to paste an auth code back into agy, but
`-p` has no input field). Before this commit the auth banner only
told the user to "open a terminal yourself."
Adds `POST /api/agents/antigravity/oauth-launch` and a cross-platform
launcher in `runtimes/terminal-launch.ts`:
- macOS: osascript → Terminal.app `do script "agy"` + activate
- Linux: tries x-terminal-emulator, gnome-terminal, konsole,
xfce4-terminal, xterm in order
- Windows: `cmd /c start "Open Design" cmd /k agy`
The endpoint hardcodes the `agy` command (no user input → no shell
injection surface) and is loopback-gated like the other daemon
endpoints. The chat's `AGENT_AUTH_REQUIRED` banner now renders a
"Sign in via terminal" button next to Retry; clicking it spawns the
terminal so the user can finish OAuth in one click.
## Silent-failure classification (auth vs quota via --log-file)
agy print mode is silent on stdout/stderr for both missing-OAuth AND
quota-exhausted failures — the upstream
`RESOURCE_EXHAUSTED (code 429): Individual quota reached` and the
`not logged into Antigravity` line only surface in agy's
`--log-file`. Without log inspection the daemon misread quota as
"auth required" and showed the wrong banner.
`RuntimeContext.agentLogFilePath` carries a daemon-owned per-run temp
path that antigravity's buildArgs translates to `--log-file <path>`.
The empty-output guard now reads that log on a `code === 0 &&
!childStdoutSeen` exit, feeds the tail to
`classifyAgentServiceFailure`, and routes:
- "not logged into Antigravity" → AGENT_AUTH_REQUIRED with
antigravityAuthGuidance
- "RESOURCE_EXHAUSTED" / "quota" / → RATE_LIMITED with
"Individual quota reached" antigravityQuotaGuidance
- none of the above (rare) → fall back to auth guidance
as the most likely cause
Both surface a terminal launcher in the auth banner: auth gets "Sign
in via terminal", quota gets "Switch model in terminal" — same
endpoint, contextual label. The handler is identical (open agy in a
terminal); the user either signs in or uses agy's Switch Model
picker to pick a model with available quota.
## Validation
- `pnpm guard` pass
- `pnpm --filter @open-design/daemon` runtime + telemetry suites:
192 passed, 1 skipped (the 1 pre-existing `task-type` failure on
origin/main is unrelated to this change)
- `pnpm --filter @open-design/web` typecheck pass; sse / amr-guidance
/ AgentIcon suites pass (51 web tests)
- Manual end-to-end on darwin + Gemini 3.5 Flash and GPT-OSS 120B
Medium: turn-1 question-form rendered correctly, turn-2 produced
`<artifact>` with full HTML (3.3KB Modern Minimal design) instead
of re-emitting the form. agy `--log-file` content correctly
classified as RATE_LIMITED when Gemini Pro quota was exhausted,
and as AGENT_AUTH_REQUIRED when keychain was cleared.
* fix(web/test): align amrAgent fixture with supportsCustomModel contract
The AMR agent definition in the daemon ships `supportsCustomModel: false`
so the Settings model picker hides the free-text "Custom…" option. The
PR changed `allowCustomModel` from `selected.id !== 'amr'` (hardcoded)
to `selected.supportsCustomModel !== false` (declarative), but the test
fixture was not updated to carry the same field — causing the
`__custom__` sentinel to appear in the picker under test.
Generated-By: looper 0.9.2 (runner=fixer, agent=claude-code)
* fix(daemon): align formAnswerTransition wording with main + scope build directive to discovery
CI surfaced two failures on the merge with main:
- chat-route.test marks submitted discovery form answers ... expected
the main-version wording 'Do not emit another <formId> form.'
- telemetry-message-finalization keeps non-discovery form answers
active ... expected task-type to fall through the else branch
('Treat these form answers as the active user turn'), not the
discovery RULE 2/RULE 3 build branch.
The colleague's earlier
|
||
kami
|
59a9867cf3
|
fix(daemon): surface discovery form answers to agents (#2071) | ||
|
lefarcen
|
4d8d233ce0
|
Fix Langfuse report finalization hook (#1402) | ||
|
lefarcen
|
afb331a288
|
feat: add opt-in Langfuse telemetry (#800)
* docs(specs): add langfuse telemetry change spec
Captures the design for forwarding completed agent runs to Langfuse,
including data-model mapping, field-budget caps, privacy gates,
build-secret injection, GDPR right-to-deletion approach, and the
resolved decisions on default consent, identifier shape, region, and
ownership.
* feat(daemon): add langfuse-trace module and telemetry prefs
Adds the dependency-free building blocks for forwarding completed
agent runs to Langfuse. Two layers:
- AppConfigPrefs gains installationId and a TelemetryPrefs object with
metrics / content / artifactManifest gates. The daemon validator
treats telemetry like agentModels — replace-on-write, drop-when-empty,
reject non-boolean inner values.
- New langfuse-trace.ts builds a {trace-create, generation-create}
pair from a ReportContext, capping prompt at 8 KB, output at 16 KB,
artifacts at 50 entries, and dropping any batch larger than 1 MB
before send. reportRunCompleted is no-op when LANGFUSE_PUBLIC_KEY /
LANGFUSE_SECRET_KEY are unset (so dev runs and forks never emit) and
short-circuits on prefs.metrics === false.
Server-side wiring into the run-close path lands in a follow-up.
* fix(langfuse): default to US Langfuse region
End-to-end smoke against the project's actual dev key on 2026-05-07
returned 401 from cloud.langfuse.com (EU) and 207 from
us.cloud.langfuse.com (US), confirming the org lives in US. Update the
default base URL, the matching test, and the spec's Q3 decision row to
match. Self-hosted or EU-region operators can still override via the
LANGFUSE_BASE_URL env var.
* feat(daemon): wire langfuse trace forwarding into run-close
Adds the daemon-side glue to forward completed agent runs:
- runs.ts gains an optional onTerminate hook fired once per run after it
reaches a terminal state. Errors thrown from the hook are caught and
logged, never propagated, so telemetry can never break the run path.
- New langfuse-bridge.ts assembles a ReportContext from the in-memory
run record, the conversation's persisted assistant message, and the
user's app-config preferences. It tolerates a missing message (e.g.
when web has not yet PUT the final delta) and a missing app-config.
- server.ts stashes the original user prompt on the run object inside
startChatRun so the bridge can include it without crossing the
createChatRunService boundary, and registers the hook callback when
building the run service.
Behavior remains a no-op unless LANGFUSE_PUBLIC_KEY / LANGFUSE_SECRET_KEY
are set in the daemon env AND telemetry.metrics is true in app-config.
A live smoke against us.cloud.langfuse.com on 2026-05-07 confirmed the
matching trace + generation schema is accepted (HTTP 207, both events
201 created).
* fix(langfuse): address PR #800 review feedback
P1 — Move trace forwarding off the daemon-internal run-close hook and
onto the message-persistence path. The original onTerminate hook ran
inside finish() the moment the SSE 'end' event was emitted, which is
*before* the web client's onDone handler refreshes project files and
PUTs producedFiles + final assistant content back to SQLite. Reading
SQLite at that moment routinely missed both. The fix: drop the runs.ts
hook entirely and trigger from PUT /api/projects/:id/conversations/:cid/
messages/:mid when the saved row carries a terminal runStatus. A
reportedRuns Set guards against the multiple PUT calls web makes per
turn (each retry / state update). Set entries auto-evict after the same
30 min TTL the runs map uses. Web persists a terminal-status message in
all three completion paths — onDone (succeeded), onError (failed), and
cancel (canceled) — so this catches every run shape.
P2 — postLangfuseBatch now parses the 207 Multi-Status response body.
Langfuse legacy ingestion always returns 207, and response.ok is true
for 207, so per-event validation errors used to slip through silently.
We now warn when body.errors is non-empty. Two new unit tests.
P2 — truncate() and the HARD_BATCH cap now compare UTF-8 byte length,
not String.length (which counts UTF-16 code units). A 4096-character
CJK prompt occupies 12 KB, well over the 8 KB input cap. truncate also
walks backwards to a UTF-8 leading byte so the cut never lands inside a
multi-byte codepoint. New unit test covers '设'.repeat(4096).
P2 — Spec R7 now lists the actual Langfuse trace deletion endpoint
(DELETE /api/public/traces/{traceId} for single, DELETE /api/public/traces
with body for batch). Verified by curl on us.cloud.langfuse.com:
DELETE /api/public/traces/X → 200; the path the original spec named
(POST /api/public/trace/X) returns 404. Reference link points at
langfuse.com/docs/administration/data-deletion.
P3 — Q4 (legacy ingestion vs OTel) moved from Open Questions to
Resolved Decisions. The implementation already commits to legacy and
the trade-off was discussed during design; the open-question status was
stale.
* feat(web): privacy consent surface + Settings → Privacy tab
Adds the user-facing half of the telemetry feature so the daemon-side
hook from PR #800 has something to talk to.
- AppConfig gains optional `installationId` (anonymous v4 uuid generated
on first opt-in; null after explicit decline; undefined when the user
has never seen the consent surface) and `telemetry: TelemetryConfig`
({metrics, content, artifactManifest}). syncConfigToDaemon round-trips
both fields so the bridge module sees the same prefs.
- SettingsDialog grows a Privacy section with two states. When the user
has never made a consent decision (typical first-run path), the
section renders the GDPR-aligned consent card: a kicker, the disclosure
body listing both metrics and conversation content as separate bullets,
and two equally-prominent buttons ("Share usage data" / "Don't share").
The Don't-share path keeps the app fully usable (core app must work
with all tracking declined). After a decision the same panel switches
to three independent toggles + the anonymous ID + a "Delete my data"
button that rotates the ID and turns everything off.
- App.tsx points the welcome modal at the new Privacy section so the
consent decision is the first thing a fresh installation sees.
- 17 i18n keys land in en + zh-CN + zh-TW with hand-translated copy,
and as English placeholders in the remaining 14 locales — enough for
the parity check to pass while leaving room for proper localisation
in a follow-up. Dict type updated.
- Minimal index.css for the consent card + toggle rows so the panel is
legible without depending on follow-up design polish.
Telemetry remains a no-op end-to-end until the user clicks Share usage
data: the daemon gate (prefs.metrics === true) keeps every code path
short-circuited otherwise.
* refactor(web): rebuild Privacy panel using project-native settings primitives
The first cut used custom .settings-privacy-* classes + raw HTML
checkboxes that didn't match any other Settings tab. Replace with the
shell other sections already use:
- settings-subsection containers with section-head + h4 + .hint
- seg-control / seg-btn pill toggles ("active" / "offline") for each of
the three telemetry preferences, mirroring NotificationsSection
- a 2-cell seg-control for the consent card so Share usage data and
Don't share carry identical visual weight (the GDPR equal-prominence
requirement that the previous accent / outline split missed)
- ghost button + readonly text input for the installation id row,
mirroring the API-key field pattern elsewhere
Drop the bespoke CSS block in favor of inheriting the existing
settings-section / seg-control / ghost styling. The only privacy-
specific style left is a tight definition list inside the consent card
for the metrics + content disclosure rows.
* refactor(web): use .toggle-row iOS switch for Privacy preferences
Active/offline pills (the seg-control single-cell pattern that
NotificationsSection uses) read awkwardly for a flat preference list.
Switch the three telemetry toggles to .toggle-row — the same control
NewProjectPanel uses for "speaker notes" / "animations": label + hint
on the left, iOS-style sliding switch on the right, full-row click
target. The consent card's two-button seg-control stays as-is — there
the equal-weight pill pair is exactly what GDPR equal-prominence wants.
* feat(web): standalone first-run privacy consent banner
Replaces the Settings-dialog-as-onboarding hack with a dedicated
bottom-right banner card that mounts whenever the user has never made
a privacy decision (cfg.installationId === undefined). The banner is
prominent (anchored to the corner with a soft shadow) but
non-blocking, mirrors cookie-consent UX, and shares the project's
panel styling — same .modal-elevated background, --radius-lg corners,
--shadow-lg lift.
Wiring:
- App.tsx imports PrivacyConsentModal and renders it at the root,
gated on installationId === undefined && !settingsOpen so it doesn't
double up with the Privacy tab's own consent card when Settings is
already showing.
- Share / Don't share both go through handleConfigPersist, so the
resulting installationId + telemetry prefs land in localStorage and
the daemon at the same time, reusing the existing autosave plumbing.
- The previous attempt that pinned the welcome SettingsDialog to the
Privacy section is reverted; onboarding now stays focused on agent
configuration, and the consent decision lives in its own surface.
* fix(web): keep privacy banner visible while Settings welcome modal is open
The banner gated itself on `!settingsOpen` to avoid double-rendering
with the Privacy tab's consent card. But the first-run path opens the
Settings welcome modal automatically when `onboardingCompleted=false`,
which fired immediately after bootstrap — so the banner flashed for a
moment and then vanished behind the modal backdrop.
Drop the `!settingsOpen` clause so the banner stays mounted whenever
the user has not yet made a privacy decision, and bump its z-index
above the modal backdrop (200 vs 100) so first-run users can actually
reach the consent buttons. The minor visual overlap with the Privacy
tab's own card is fine: clicking either copy resolves both surfaces.
* copy(privacy): soften consent button labels
Banner action buttons now read "Help improve Open Design" / "Not now"
(en, with hand translations in zh-CN / zh-TW and English placeholders
in the other 13 locales) instead of "Share usage data" / "Don't share".
The new wording aligns the affirmative action with the kicker copy
("Help us improve Open Design") and reads less alarming, while the
disclosure list above still names both data categories explicitly so
the consent stays informed under GDPR. The decline button stays as a
soft "Not now" rather than an aggressive "Don't share" so the reject
path doesn't read as hostile to the user.
No structural change — the two-cell seg-control still gives the buttons
identical visual weight, and the underlying side-effects are unchanged
(installationId is generated on Help / nulled on Not now, and the
telemetry prefs flip the same way).
* feat(telemetry): expand trace fields for evals & dataset construction
Each Langfuse trace now ships the full per-turn + per-install fact
sheet that the eval/dataset workflow needs, instead of only the bare
turn id + token count from before. Everything below is gated by
`prefs.metrics === true`; nothing here is content (those gates remain
separate).
Per-turn:
- model — first-class generation.model field, drives Langfuse cost
lookup and model-grouping in the UI; also mirrored in trace.metadata
and trace.tags so list-view filters work.
- reasoning — generation.modelParameters.{ reasoning } so the Model
Parameters card lights up; mirrored in metadata.
- skillId / designSystemId — metadata + tags, so dataset slices can
group by which skill/DS produced which output.
Per-process / build (constant within one daemon run, cached at start):
- appVersion / appChannel / packaged from app-version.ts
- nodeVersion (process.version), os (platform()), osRelease,
arch (os.arch())
- clientType — desktop vs web, derived from a new X-OD-Client header
the web layer sets in providers/daemon.ts (with a User-Agent sniff
fallback for third-party callers).
Plumbing:
- startChatRun stashes model / reasoning / skillId / designSystemId
on the run object alongside the existing userPrompt stash.
- POST /api/runs reads X-OD-Client and stores run.clientType.
- langfuse-bridge collects RuntimeInfo once per process and merges
per-run client carrier; ReportContext gains optional `turn` +
`runtime` blocks; existing fields stay backward compatible.
Spec gains a "Telemetry Fields Catalog" section enumerating every
field, its source, and the gate it lives under, so the eval team has a
single place to look up what's available without reading the trace
schema by example.
Tests:
- new langfuse-trace tests cover turn tags, runtime tags, generation
model/modelParameters promotion, modelParameters omission when
reasoning is unset, and metadata mirroring.
- langfuse-bridge gains an end-to-end "turn-level config" test that
threads model/reasoning/skill/DS/clientType + appVersion through
the bridge and asserts the Langfuse payload shape.
- existing tests adjusted to tolerate host-dependent os tag.
* copy(privacy): trim Share button to verb phrase only
"Help improve Open Design" overflowed the equal-width 2-cell
seg-control on the consent banner — the product name is already in
the kicker + headline above the buttons, so the button itself only
needs the verb phrase. Drop the product name from all locales:
- en: Help improve Open Design → Help improve
- zh-CN: 帮助改进 Open Design → 帮助改进
- zh-TW: 協助改進 Open Design → 協助改進
The decline button ("Not now" / "暂不" / "暫不") was already short, so
the two buttons now have comparable length and the equal-prominence
seg-control fits cleanly. Standalone Settings → Privacy panel uses
the same labels for consistency.
* fix(web): defer Settings welcome modal until privacy decision is made
Previously bootstrap raced two surfaces against each other on first
launch: the privacy consent banner (gated on installationId ===
undefined) and the Settings welcome modal (gated on
onboardingCompleted === false). The banner's higher z-index kept it
above the backdrop visually, but having two foreground surfaces at
once is still confusing UX.
Sequence them instead: bootstrap only opens the welcome modal when
the user has *already* resolved consent (installationId !== undefined).
Until then the banner owns the foreground alone. Once the user clicks
Help improve / Not now, the corresponding handler hands off to the
welcome modal if onboarding is still pending. End state matches what
it was before — just without the simultaneous-render flash.
* debug(privacy): log banner gate state to track sudden disappearance
Two console.log points to find which setCfg call (or stale bundle) is
flipping cfg.installationId from undefined to a value while the banner
is visible. To remove once the regression is reproduced.
* fix(privacy): keep installationId + telemetry out of localStorage
Daemon is now the single source of truth for the privacy decision.
Why this matters: the consent banner gates on
\`config.installationId === undefined\`, but loadConfig() merges
localStorage on top of the daemon's reply, so a stale uuid in
\`open-design:config\` (left over from a previous opt-in) was
re-hydrating the React state and immediately syncing back to the
daemon — defeating "Delete my data" and re-suppressing the banner
within milliseconds of every page load.
The deeper reason to fix it here, not just patch the gate: a privacy
identifier persisted in browser storage that the user can't see or
clear without DevTools is a compliance liability. Anything users can
revoke needs one canonical place to store it. Daemon \`app-config.json\`
already serves that role for everything else gated through
syncConfigToDaemon, so installationId + telemetry now ride that path
exclusively:
- saveConfig() strips both keys before writing localStorage.
- loadConfig() strips both keys when reading older stale payloads,
so existing installs migrate transparently on next launch.
- syncConfigToDaemon() / mergeDaemonConfig still round-trip them, so
the React state stays in sync with the daemon as before.
Net effect: clearing app-config.json (or hitting "Delete my data") now
fully resets the install identity, with no residual cohort key in
browser storage.
* feat(privacy): scrub secrets + PII from prompt/output before send
When prefs.content is on, daemon now runs the prompt and assistant
text through a regex scrubber (apps/daemon/src/redact.ts) before
posting to Langfuse. The scrubber is the simplest thing that gives
the user-facing copy a truthful claim — pure regex, zero new
dependencies, fully auditable in this Apache-2.0 repo (vs. pulling a
single-maintainer 5-month-old npm package into a core process).
Categories covered (each replaced with [REDACTED:<kind>]):
- Anthropic / OpenAI sk- keys (incl. proj/live/test/ant variants)
- Langfuse pk-lf- / sk-lf- (specific rule wins over generic sk-)
- GitHub gh[opsur]_ tokens
- AWS access key ids (AKIA + 16 uppercase)
- Google API keys (AIza + 35)
- Slack xox[abprs]- tokens
- Stripe live/test keys
- JWT header.payload.signature triples
- Bearer-header values (scheme word stays readable)
- Emails, IPv4, US-style phone numbers
- Credit cards — 13–19 digit runs that pass a Luhn check, so order ids
and unix-nanos timestamps that fail Luhn pass through unchanged
Not covered, stated openly in spec + i18n: names, postal addresses,
business-secret semantics, raw 40-hex tokens (too high a false-positive
cost for artifact slugs). Those would require an ML layer.
Wired in:
- apps/daemon/src/redact.ts — exports redactSecrets() +
redactSecretsWithCounts() helper for future audit-summary metadata.
- apps/daemon/src/langfuse-bridge.ts — runs both prompt and output
through redactSecrets() before they reach the trace builder.
- 18 unit tests cover every pattern plus negative cases (Luhn-failing
digit runs, out-of-range IPv4 octets, idempotence on re-redacted
text, ordinary prose passthrough).
- i18n privacyContentHint on en + zh-CN + zh-TW (plus 14 locale
placeholders) enumerates the categories so the consent disclosure
matches the implementation — the GDPR informed-consent requirement.
- spec gains a Pre-send Redaction subsection with the regex shape
table + intentional non-coverage list.
Drive-by: dropped the [privacy] debug logs that traced the now-fixed
bootstrap regression.
* fix(telemetry): make Langfuse reporting resilient
* feat(telemetry): nest Langfuse turn observations
* feat(telemetry): emit Langfuse tool spans
* fix(telemetry): report after finalized message writes
* fix(telemetry): honor persisted terminal status
* fix(web): let consent banner yield page clicks
* fix(telemetry): report current turn prompt only
|