The floating assistant was broken end-to-end; Sami said "Do it that works".
Three user-visible bugs fixed + the promised LLM delegation implemented:
FRONTEND (status/js/ai-chat.js — rewritten):
- Response parsing: old code read data.data; the API returns the FLAT ok()
envelope — EVERY reply rendered "Sorry, something went wrong". Now reads
the real shape (contract asserted by route test).
- Suggestion chips: onclick="...send('"Deploy Plex"')" — nested quotes,
invalid JS, dead buttons (shipped like this). Now data-driven buttons with
addEventListener.
- Deploy button: called window.openAppSelector which NEVER existed anywhere
in the tree (grep-verified). Now opens the REAL deploy modal via new
window.openDeployConfigFor(templateId) exported by app-selector.js —
resolves against the live API template cache, reuses showDeployConfig
(same path a manual catalog click takes). Widgets route to the grid.
- Conversation history (last 6 turns) now rides along per request, engine
badge in header, ⚙ settings sub-panel (baseUrl/model/API key + Test +
Remove), all server text via textContent (no HTML injection), CSP-safe
(no inline handlers).
BACKEND (routes/ai-intent.js):
- TWO engines, one contract. LLM classifier (OpenAI-compatible chat
completions) with system prompt embedding the LIVE 78-app catalog + the
six wizard categories; regex patterns kept as the zero-config default and
the automatic failover (HTTP error, timeout via AbortController,
unparseable output -> pattern answers; response carries engine +
llmError so the UI can say which served — no silent degradation).
- Closed-vocabulary schema validation: intent MUST be in the known set else
'unknown'; appId MUST exist in APP_TEMPLATES else nulled (hallucinated
ids cannot leak into deploy plans); categories filtered to the wizard's
real keys; reply stripped of control chars + bounded to 600; history
role-filtered (only user/assistant pass — client-injected system turns
dropped) + capped; input capped to 1000 chars.
- enrich() rebuilds the response contract per intent; pattern-canned
response is dropped when the LLM rerouted the intent (no stale shape);
deployPlan body.port now fills from the template's defaultPort (was
hardcoded null); deployPlan shape asserted against wizard usage.
- Key storage: PUT /ai/settings stores {baseUrl,model,apiKey} in the
ENCRYPTED credential store (key 'ai:llm') via ctx.credentialManager —
NOT config.json (GET /config echoes the raw file to the browser; that
route's shape was verified before this decision). Blank apiKey on PUT
keeps the stored one. GET /ai/settings reports configured/source/model/
hasApiKey — the key value NEVER crosses the wire (asserted). Env
fallback AI_LLM_BASE_URL/AI_LLM_API_KEY/AI_LLM_MODEL + legacy
LLM_PROXY_URL for headless installs. POST /ai/settings/test = one-shot
real classify probe returning ok/sample or ok:false/reason (never
throws). DELETE clears.
- GET /ai/capabilities now reports engine + llm status (pattern was honest
marketing before; now it can be true).
MOUNT (src/app.js): aiIntentRoutes receives siteConfig, log, fetchImpl
(ctx.fetchT — test-injectable) and credentialManager.
Docs: OpenAPI +4 AI path entries (intent, capabilities, settings, settings/test); CHANGELOG entry; dist bundle rebuilt
(openDeployConfigFor present in features.js; SW cache c62511c40d).
Tests: 20 NEW in __tests__/routes/ai-intent-engine.test.js (engine
selection w/ and w/o store, off-catalog phrasing via mock LLM,
hallucinated-id nulling, illegal intent coercion, HTTP-500 failover,
AbortError timeout failover, markdown-fence repair, reply sanitization,
history cap+role filter, env config path, intent-disagreement rebuild w/
real jellyfin port 8096, flat envelope contract, 400s, 1000-char cap,
settings GET/PUT/DELETE/test incl. key-never-echoed assertions,
capabilities engine reporting) + the 18 original pattern tests unchanged +
composition-root e2e 8 pass. Full suite: 3643 passed; only the pre-existing
pdf-parse billing family red (6 failed in documented band [5,7], patch
touches no billing file).
Judge r2 (grade C) — all 3 blockers + both polish items:
1. REQUEST-vs-NOTIFICATION by PRESENCE, never truthiness: validateJsonRpc
uses hasOwnProperty('id'); id 0 and id '' are VALID request ids echoed
back BY VALUE (strict-equality tested), id:null treated as a request
under the documented error-id convention. mcp-core default branch fixed
the same way (`if (id)` → hasOwnProperty — id 0 / '' unknown-method now
answer -32601 with the id echoed instead of being swallowed as falsy).
Routes' malformed-notification silence path also switched to
presence-based detection (id:null + bad params → 400 envelope, not 204).
2. params:null REJECTED (-32600): contract is object-or-omitted ONLY
(strict JSON-RPC; the old code's m.params !== null guard contradicted
the documented contract). Boundary tests pin the policy.
3. Unknown-method boundary tests: id 0 → -32601 id:0, id '' → -32601 id:'',
id null → -32601 id:null — asserted at BOTH layer levels (route).
Polish:
- 16-case TABLE-DRIVEN boundary matrix covering every allowed/rejected
id/params combination incl. 202 vs 204 notification split.
- Verifier now probes writable /tmp up front (exit 73 + clear message if
unavailable; documented requirement) instead of dying at mktemp.
- Verifier target count 57 → 73 (DC-162 suites) with the boundary table
named in the assertion label.
Full DC-162 suite: 73 passed (route 50 + core 23). No transport/credential
semantics changed from rev2.
urn:ump:or2qqhcytp6lfzlyi7xru7hdqy3uar2hbjllsrrbxrbray3c5kpa
All three r3 blocking issues fixed + polish applied:
- Radicale provision route now FAILS CLOSED (409) unless the target container
is present, inspectable, and its image matches /radicale/i. Missing
containerId, failed inspect, or non-Radicale image → nothing is written to
the shared config and no source/credential is registered. 4 new tests cover
the three refusal paths + the verified-image happy path.
- ICS feed fetch now streams with a hard 5 MiB body cap (same bound as the
CalDAV path) via readBodyCapped(); oversized feeds reject mid-stream and
degrade to per-source error. Test covers a >5 MiB feed end-to-end.
- Verifier mode separation: `--verify` runs the same checks fully read-only
(all output under mktemp -d in /tmp, transcript printed to stdout, scratch
removed on exit — proven rc=0 with zero repo/tmp-side writes); default
transcript mode is explicitly declared a TRANSCRIPT GENERATOR in its header
and writes only under the script's own directory.
- Full-suite count assertions replaced with the documented invariant
(failed ∈ [5,7], failing set == known-flaky billing suite, extras pass
isolated rerun) — judge r3 polish.
- Stale credential-key comment in radicale-provisioner.js header corrected.
Calendar suites: 43 passed (38 + 5 new). Perf note: readBodyCapped streams;
no full-buffer read of unbounded bodies on the ICS path anymore.
All three r2 blocking issues fixed + polish applied:
- useExisting (adopted) Radicale containers no longer get a source registered
with credentials the operator never provisioned. Registration happens ONLY
when pre-start provisioning succeeded; adopted deployments get a
calendarHint in the deploy response pointing at the manual-wiring modal.
- Weekly BYDAY expansion: candidates earlier than DTSTART are skipped (they
do not consume COUNT); window inclusion now uses interval-overlap
semantics (occurrence span vs window span) for both recurring and
non-recurring events. Three new tests cover midweek DTSTART, a straddling
multi-day event, and a boundary-exactly-ending event.
- Verifier is now read-only on the graded tree: no stash/checkout/rewrite.
Base proof runs in a temporary git worktree of 448f535 under /tmp (removed
on exit); the flaky pdf-parse suite is proven pre-existing via "red at
base" + "patch touches no billing file" (empty diff), and a race-flaky
suite (DC-116, witnessed once under parallel load) is given one isolated
rerun which must pass.
Polish: PUT /sources normalizes the user field (and clamps name) like POST.
Calendar suites: 38 passed (35 + 3 new recurrence tests). The 3558/3564
full-suite numbers vary by one flaky pdf-parse member between runs; the
evidence now asserts the invariant (suite red at base, no billing file in
the patch, flaky suites green on isolated rerun) instead of an exact count.
All five r1 blocking issues fixed + polish items applied:
- Unified credential-key convention: registerAutoWiredSource now stores at
calendar:source:<source.id> — the same key the events fetcher resolves —
so auto-wired sources authenticate on first fetch (was calendar:radicale:
<subdomain>, fetched key never written = empty-password CalDAV auth).
- calendarStateManager is now passed INTO assembleContext(...) (the r1 patch
only extended the function signature, so ctx.calendarStateManager was
undefined and the deploy hook silently skipped auto-wiring).
- Radicale config is provisioned BEFORE container start (Radicale reads its
config at startup): the deploy hook writes config+htpasswd ahead of
deployContainer() and passes the pre-start credentials to the post-deploy
registration step; useExisting path logs an explicit restart warning.
- PUT /sources/:id now validates URL updates through the same shared
validateSourceUrl() used by POST (protocol + remote-plaintext rules).
- New end-to-end contract test (calendar-autowire.e2e.test.js): spins a real
HTTP CalDAV server, registers an auto-wired source, then fetches events
through the API — the server verifies Basic auth against the provisioned
password, so a credential-key or wiring regression fails the suite.
Polish: provisioning is idempotent (deterministic source id, update instead
of duplicate), unbiased rejection-sampling for password/salt generation,
writeRadicaleConfig docstring no longer claims idempotence.
Calendar suites: 35 passed (32 prior + 3 new). Lint: 0 errors; complexity
warnings in deploy.js pre-exist at base (61 vs 68 now — the hook did not
worsen the trend line materially; calendar files themselves lint clean).
- DC-160: /catalog/:appId read app.image/ports/env/volumes which live under
app.docker.* — every detail response was empty. Now reads docker.* and
surfaces defaultPort/subdomain/healthCheck/mediaMount/claimToken/difficulty/
setupInstructions for the deploy form. List/search use the templates own
category field (old hand-rolled CATEGORY_MAP mis-categorized apps).
- Shield (routes/shield.js + 5-line app.js mount) committed to source after
weeks as prod-only drift; regression test asserts the mount exists.
- New composition-root e2e suite: boots REAL createApp() on an ephemeral
port, probes the historically-broken routes through the REAL mount table
(wizard wiring, shield, catalog detail, deploys bridge) — the class of bug
all 3 P7 dogfood findings belonged to. 5 tests.
- Full suite 3499/3499 (155 suites); table-driven DC-160 test covers all 73
image templates.
Dogfood finding #3: upsertPerson resolved existing records by id only.
The wizard upserts a bootstrap admin AFTER a provider sync has already
created a person for the same Tailscale user (no id passed), so the same
providerUserId was duplicated into two records. Resolve existing by
providerUserId when no id is given; bootstrap Admin group + protectedAt are
preserved on the merge. New regression test; full suite 3414/3414.
Dogfood finding #2: createTailscaleProvider() built with coordClient=null;
nothing in prod ever called attachCoordClient() (the wiring the provider
docblock promised never landed), so wizard credential validation always
failed not_configured even with a valid stored token. Pass a delegating
proxy that fetches a fresh coord client per call via the sanctioned
ctx.tailscaleCoord.getClient pattern (token re-read from credentialManager
each call). Full suite 3413/3413.
The /network/wizard mount in app.js omitted providers, so the wizard built
with providers={} and every mode selection returned 503 provider_unavailable
in production. Unit tests never caught it because they inject their own
provider map. Pass built tailscale + wireguard instances.
Full suite: 3413/3413 (152 suites).
- routes/ca.js: declare format before pfx/pem/crt dispatch (was ReferenceError
on every request that passed validation); add CA_CERT_FORMATS single source
of truth + hardened format extraction (string-coerce, whitelist)
- routes/caddycode.js: fix upstream-validation regexes (control-char classes)
- routes/logs.js: SSE/validation lint fixes
- routes/openclaw.js: remove useless regex escape in ALLOWED_PATH_RE
- fleet-validation.js + http-caddy-admin-origin test: eslint-disable for
intentional control-regex security sentinels
- ca-dc076.routes.test.js: regression test for declared format + behavioral
coverage of format validation (now pre-PKI)
Judge: Qwen lane (qwen3.8-max) grade B, 0 blocking, verdict
/tmp/judge-batch1-verdict.json. API suite 2859/2859 green.
The host-side uptime watchdog and on-host cron jobs curl Caddy with a stock
curl/8.5.0 UA from loopback and the host tailscale IP — ~300 GET /api/health
401 warn-events/day burying real perimeter signal (census: 339+20 of 5000
access.log lines). Dropping on UA alone would blind the store to external
curl scanners, so generic tool UAs (curl/) are now dropped ONLY when the
source remote_ip is one of this host's own addresses (DASHCADDY_SELF_IPS,
default loopback; start.sh derives 127.0.0.1 + tailscale ip -4, empty-safe).
remote_ip (TCP peer) is used, never the spoofable client_ip. DashCaddy-*
probe UAs stay unconditionally dropped. 9 regression pins cover the full
conjunction matrix incl. external-IP+curl KEPT and spoofed-XFF KEPT.
2858/2858 green (128 suites).
Judge: glm-4.6@zai-coding-paas cold-read round 1 = A clean (0 blocking),
both polish notes folded. URN urn:ump:s2sgitfepze65crtp57dpdw4gk4w7upsoi4tqcvfwahcsicyepba
- query().total: full-scan true count (was capped at offset+limit by an
early break — dashboard 24h stat read ≤1000 vs real ~46k)
- trim trigger/curer mismatch: byte trigger + line curer never converged
when avg line > ~524B (live avg 355B); now keeps maxDisk lines OR
≤80% byte budget, whichever retains fewer (TRIM_TARGET_FACTOR);
budget env-overridable via SECURITY_EVENT_TRIM_BYTES / opts.trimSizeLimit
- trim/append race: trim renamed over the file mid-append losing events
to the unlinked inode; now single-flight, queue-empty gated, holds the
write lock, error paths always release + lazily re-kick (no hot loop)
Judge: glm-4.6@zai-coding-paas adversarial cold-read, grade B clean
(0 blocking, 4 polish — 2 folded, 1 already-satisfied, 1 deferred as
judge-endorsed safer). 127 suites / 2849 tests green.
Hermes