[vinhali]

Summary

Reconstructs O360 explicit-target routing directly on release/v3.8.51 commit 7bd7c75f1ea2a7a08b2b7719a1e584e91d1ee567. The diff is limited to combo routing, its existing quality baseline, and focused tests; current Azure STT, MAI WebM normalization, auth/audit work, and production fixes remain untouched.

• Adds highest-priority /route <target> handling, with task text accepted on the directive line or following lines.
• Removes only the directive from the current user turn and preserves the complete prior conversation.
• Adds one-turn-only routing with the existing recursive combo dispatch guard.
• Adds secondary natural-language routing from the last user message only, excluding fenced code, inline code, tool output, prior turns, and non-user messages.
• Adds built-in, case-insensitive aliases claude -> cc/claude-sonnet-5 and astra -> cx/gpt-5.6-sol-high; operator-configured synonyms retain precedence.
• Preserves exact, case-sensitive semantics for literal model IDs.
• Resolves combo names such as o360-coding and o360-cheap-worker through the existing combo resolver/dispatch path.
• Fails closed with no fallback for unknown slash targets and known unavailable targets. Weak unresolved natural-language phrases return to normal combo routing.
• Reuses the existing resolver, single-model/combo dispatch, durable-health gate, decision trace, and recursion guard. No database, migration, service, daemon, MCP, or classifier is introduced.
• Records explicit-routing source as slash_command or natural_language in the existing combo decision trace.

Acceptance coverage

• /route claude, /route astra, /route o360-coding, /route o360-cheap-worker, and /route cx/gpt-5.6-sol-high
• Same-line and next-line task content
• UNKNOWN_TARGET and TARGET_UNAVAILABLE with zero fallback
• Full history preservation and return to normal routing on the following turn
• Fenced/inline code exclusions
• use Claude natural-language routing
• Ordinary phrases including use a tool, use combo fallback, and manda a mensagem falling through normally
• Existing o360-coding ordering remains cx/gpt-5.6-sol-high then cc/claude-sonnet-5, with no GLM target

Validation

• Focused explicit-target suite: 110/110 passing
• Full combo suite: 296/296 passing
• Vitest: 465/465 passing
• npm run typecheck:core
• npm run check:file-size
• npm run check:complexity-ratchets
• Prettier check on all changed files
• ESLint on all changed TypeScript files

Live provider check

• Astra/Codex connection is active; a real cx/gpt-5.6-sol-high inference returned HTTP 200 with the requested sentinel.
• Claude connection is currently inactive/expired; direct cc/claude-sonnet-5 did not infer and did not substitute another model. With this branch, /route claude returns TARGET_UNAVAILABLE while that state persists.

[vinhali]

PR/release hygiene pass — commit classification, cleanup, separate issue filed

Per request, closing out three housekeeping items before this PR is considered mergeable (still NOT merging or deploying — that stays a separate, explicit decision).

1. Commit classification (verified via git show --stat on every commit, not commit messages)

A) Feature (explicit-target-routing):

• b9d6410f1 — initial implementation
• cb807a96a — gap 2/3 fixes ("com o" false-positive tightening, combo-vs-pin recursion test)

Both touch only open-sse/services/combo/* and tests/unit/combo/*.

B) CI-only workaround (used solely to build/run the isolated canary for the gap-1 E2E proof):

• 411dcee7f, 76121cd04, b24c1adf4, b6506fd33, 54d5c7380 — touched only .github/workflows/e2e-build-temp.yml and Dockerfile.e2e-temp, zero overlap with feature files.

2. PR scope restored

Removed both CI-only files (c8d4f2b59). git diff cb807a96a HEAD is now empty — PR #1's tree is byte-identical to the feature-only state. Current file list is exactly the 11 feature files (7 source + 3 test, plus the shared-module export touches).

3. tls-client-node pipeline bug — separate issue, not hidden here

Filed as #2: the real, unmodified Dockerfile + runner-web target currently fails on any fresh build — bogdanfinn/tls-client's real v1.16.0 release assets are named tls-client-xgo-1.16.0-<platform>.<ext>, but the pinned tls-client-node@0.2.0 postinstall script still expects the old tls-client-<platform>-<version>.<ext> naming. Confirmed via a direct query against the real GitHub release (not a rate-limit, not assumed). This is a genuine upstream/packaging bug, independent of this patch, and needs its own fix decision from whoever owns the Dockerfile — not something bundled or silently worked around in the routing PR.

Final summary

ROUTING_V1_SOURCE_SHA=cb807a96a1... (full: see below)
CANARY_IMAGE_DIGEST=sha256:7fcb30ba4b810a1f5e9f50f95c7d6982eb7b763242f6b12d75490d5e9092e2df
REAL_AGENT_THREE_TURN_E2E=PASS
ROUTING_PATCH_CAUSED_PREVIOUS_CRASH=false
TEMP_BUILD_WORKAROUND_PRESENT_IN_PR=false
SEPARATE_PACKAGING_FIX_REQUIRED=true (tracked in #2)
PR1_READY_FOR_REVIEW=true

No merge, no production swap performed or scheduled.

🤖 Generated with Claude Code

https://claude.ai/code/session_01MPKHh3DGo2bJBmXGQwtvSb

[vinhali]

Reconciled onto the current release/v3.8.51 head (now includes PR #3 combo-audit, PR #4/#6 issue #5 fix, PR #7 Cloudflare Access identity) — clean merge (git merge-tree, zero conflicts), since this feature's diff is fully isolated to combo/explicit-target files none of those PRs touched. Cherry-picked only the 2 real feature commits (dropped the temp CI-workaround commits that were already reverted before the original merge). Fixed 2 real Fast Quality Gates findings surfaced against the new base: a complexity-ratchets regression in explicitTargetDetector.ts (extracted a helper) and a frozen file-size cap on combo.ts (rebaselined with justification, matching this file's own established precedent for irreducible dispatch-chokepoint wiring). All 46 explicit-target-routing unit tests pass. New head: 4dbdf6185. CI running.

[vinhali]

CI fixup: false-positive explicit-target detection + file-size baseline correction

Run 34261416326 surfaced two real issues after reconciliation against the moved release head (1fe18e5206969749f331df5374f8199a61e080d4), now fixed in 50ef8f20d:

1. Fast Quality Gates (file-size) — fixed. The earlier rebaseline of open-sse/services/combo.ts (4111) was computed before lint-staged's Prettier reflow of several pre-existing >100-col lines (unrelated to this PR, same class as the _rebaseline_2026_09_07_ci_baseline_repair_prettier_reflow precedent) had run against the final reconciled diff. Verified via git diff -w that the reflow is whitespace-only. Corrected the frozen cap to the true value (4148) with an updated justification in config/quality/file-size-baseline.json.

2. Unit Tests fast-path (1/4, 2/4, 3/4) — fixed, real regression. explicitTargetDetector.ts's VERB_FIRST_PATTERN had no structural safeguard against matching ordinary sentences: "Use combo fallback" and "Use a tool to look something up" were detected as explicit-target commands (alias combo/tool), failed to resolve, and hard-failed the request closed with 400 UNKNOWN_TARGET — breaking 3 pre-existing tests in tests/unit/chat-route-coverage.test.ts and tests/unit/combo-roundrobin-compat-fallback-6238.test.ts.

Root cause: every required positive test uses a capitalized alias ("Astra", "Claude"); every false positive uses a lowercase common noun. Fix requires the captured alias to start with an uppercase letter (\p{Lu}), applied to both VERB_FIRST_PATTERN and COM_O_PATTERN for consistency — verified against all 29 pre-existing detector tests (still green) plus 3 new regression tests added for this exact class of false positive. A lowercase-named real target stays reachable via the admin-configured synonym tier.

Also caught and fixed a real regex-engine trap along the way: the i flag case-folds \p{Lu} in V8/JS, silently defeating an uppercase-only check — confirmed empirically, documented in the code comment.

Verified locally: detector's own 32/32 tests pass; direct module invocation confirms detectExplicitTargetInBody now returns null for all 3 previously-misdetected message strings (letting normal combo dispatch proceed as before); check-file-size.mjs --base-ref 1fe18e520696 passes clean.

Awaiting fresh CI on 50ef8f20d.

[vinhali]

Semantic correction: hybrid /route + conservative natural language (replaces capitalization heuristic)

Commit 1da6e4b97 replaces 50ef8f20d's capitalization-based precision fix with the intended hybrid design: a deterministic /route <target> control command plus a natural-language convenience path that never fails a request closed on its own.

1. /route <target> (new, routeSlashCommand.ts) — the deterministic control API. Recognized only as the first non-blank line of the last user message (a real command is never itself wrapped in a code fence/inline-code span — the fence/span markers occupy that leading position instead, so this alone satisfies "must not route inside code" with no extra logic). Case never matters for resolution (/route claude / Claude / CLAUDE all resolve identically); literal model ids with / and . (cx/gpt-5.6-sol-high) and lowercase-hyphenated combo names (o360-cheap-worker) both route correctly. An unresolved /route target always fails closed (UNKNOWN_TARGET) — the syntax is unambiguous. Strips the control line from the forwarded task text (full prior history untouched) and fails closed with MISSING_TASK if the command was the turn's only content.

2. Natural language ("use <alias>", simplified explicitTargetDetector.ts) — capitalization is no longer a routing signal at all. An unresolved candidate now always falls through to ordinary chat instead of failing closed, since ordinary prose ("use combo fallback", "manda a mensagem") takes the identical shape with zero routing intent, and real targets are routinely lowercase anyway. Removed the previous commit's model-keyword/trailing-purpose-clause "strength" tiering entirely — /route now owns 100% of the deterministic fail-closed-on-unknown guarantee, so that complexity is gone.

tryExplicitTargetDispatch.ts checks /route first (highest priority); only if absent does it fall back to natural language. Same resolver, health gate, and recursion guard as before — each audit-trace decision now also records its source (slash_command vs natural_language).

Test coverage added/updated

• combo-route-slash-command.test.ts (new, 23 tests): extraction, case-insensitivity, literal model ids, fenced/inline-code exclusion, first-line-only requirement, last-user-message-only scoping.
• combo-explicit-target-detector.test.ts (41 tests): simplified back to pure extraction (no strength field), lowercase aliases, lowercase-hyphenated combo names, literal model ids.
• combo-explicit-target-dispatch.test.ts (29 tests, real handleComboChat end-to-end): case-insensitive /route alias/combo/model-id resolution, /route unknown → fail closed, natural-language unknown → fall through (including the exact previously-regressed phrases), TARGET_UNAVAILABLE/context-cache-pin-override/audit-trace/three-turn scenarios for both routing sources, directive-stripping and earlier-turn-immutability for /route.
• combo-explicit-target-resolver.test.ts: unchanged, still 8/8 passing (resolver needed zero changes — it already supported case-insensitive combo/synonym/label lookups and literal model ids with //.; the gap was purely in what the detector could extract).

Local verification (this sandbox lacks a full CI-equivalent build; real CI will confirm)

• All 4 targeted test files pass individually (101 tests total across detector/slash-command/dispatch/resolver).
• Previously-regressed tests (chat-route-coverage.test.ts's "Use combo fallback" case, combo-roundrobin-compat-fallback-6238.test.ts) verified passing again under the new mechanism.
• check-file-size.mjs --base-ref 50ef8f20d242 → OK, no rebaseline needed.
• check-complexity-ratchets.mjs --base-ref 50ef8f20d242 → OK, 0 new violations.
• check-ts7-diagnostics-ratchet.mjs --base-ref 50ef8f20d242 → OK, 0 new diagnostics.
• eslint on all changed/new files → clean.
• semgrep not runnable in this sandbox (not installed) — deferred to CI.

Awaiting fresh CI on 1da6e4b97. Still not merged, per standing instruction.

[vinhali]

Real CI failure found and fixed: NL must never reach tier-4 literal-model-id resolution

CI on `1da6e4b97` (the hybrid /route + NL redesign) caught a genuine regression: "Use combo fallback" and "use FooModel para revisar" — both previously-regressed phrases — failed with a raw `400 "Unable to determine provider for model 'combo'"` instead of falling through to normal combo dispatch.

Root cause (confirmed via a direct production stack trace, not assumed): `resolveExplicitTarget`'s tier 4 (literal model id, gated on `isModelAvailable`) was reachable for natural-language candidates too. The real `isModelAvailable` implementation (`checkModelAvailable`, `src/sse/handlers/chat.ts`) is written for a model string the caller already believes might be real — it calls `getModelInfo()`, which doesn't throw for a nonsense string like "combo" (it returns a result object with `errorType: "model_not_found"`), and that doesn't reliably propagate as "not available" back up the chain. Fed an arbitrary word extracted from ordinary prose, it can answer "yes" — letting the resolver treat "combo" as a genuine literal model id and dispatch straight to it. This was not reproducible with an isolated unit-test stub (which only returns what you tell it to) — it only surfaced going through the real production check end-to-end via `handleChat`, which is why I chased it down with an actual stack trace before concluding anything.

Fix (`038ce3397`): natural language now only ever resolves through tiers 1-3 (settings synonym, exact combo name, exact target label) — all local, deterministic lookups with no fail-open risk. Tier 4 stays available for `/route `, where the user explicitly typed the string as a target and `/route cx/gpt-5.6-sol-high` needs it. Added a regression test that simulates `isModelAvailable` always answering "yes" (not "no") to prove natural language ignores tier 4 regardless of what that check says, rather than passing coincidentally.

All 30 dispatch tests + 41 detector tests + 23 slash-command tests + 8 resolver tests pass locally, including a direct reproduction of the original failing scenario via a standalone script hitting the real `handleChat` path (confirmed 200, not 400, after the fix).

Awaiting fresh CI on `038ce3397`. Still not merged.

[vinhali]

Fast Quality Gates fixup: max-lines-per-function

CI on `038ce3397` (the tier-4 fix) got all 4 Unit Test shards green (confirming the tier-4 fix itself) but failed Fast Quality Gates on `complexity-ratchets`. Root cause, confirmed via a scoped eslint run: cyclomatic complexity was never the issue (every function in the file measures well under the repo's `complexity>=15` threshold) — the `allowLiteralModelIdLookup` branching pushed `dispatchResolvedTarget` to 97 lines, over `max-lines-per-function>80`.

Fixed in `85b383dde` by extracting the model-kind and combo-kind response branches into their own `dispatchToModel`/`dispatchToCombo` helpers — pure extraction, no behavior change (all 30 dispatch tests pass unchanged). `dispatchResolvedTarget` now just resolves, checks unknown/ambiguous, and delegates.

Awaiting fresh CI on `85b383dde`. Still not merged.

[vinhali]

CI green — all checks passing on commit 85b383dde

All 13 checks pass (Fast Quality Gates, all 4 Unit Test shards, semgrep, typecheck, docs, eslint). The one failure seen after the max-lines-per-function fix was a 27-minute hang in an unrelated "RTK strip-code-comments" test on shard 4/4, followed by a runner-level cancellation — confirmed transient by re-running just that job, which passed cleanly on retry with no code changes.

READY_FOR_HUMAN_APPROVAL=true. Still not merged, awaiting your review.