@qadanm/opencode-smart-router 0.4.0 → 0.6.1

package/CHANGELOG.md

@@ -10,6 +10,174 @@ Notable changes to OpenCode Smart Router. Dates are the day the change landed on
 - **Launch kit** (`docs/LAUNCH-KIT.md`): paste-ready copy for Discord, Hacker
   News, X, and the opencode.cafe submission, plus the demo film assets.
 
+## 0.6.1 — 2026-11-06
+
+A defect in the router can no longer stop the editor.
+
+Version 0.6.0 could refuse every prompt, in every mode, until the plugin was
+removed by hand. Two defects compounded to cause it. This release fixes both,
+and adds the missing test layer that would have caught them. The full account is
+in [docs/INCIDENT-2026-11-06-prompt-lockout.md](docs/INCIDENT-2026-11-06-prompt-lockout.md).
+
+### Fixed
+
+- **An inherited shared core is no longer adopted.** The process-wide core in
+  `globalThis` is stamped with `SHARED_CORE_API`; a core from a different
+  version is replaced rather than reused. A new bundle could previously inherit
+  a core built by an older release, whose ledger had no `planPoolSpent()`, and
+  every prompt then threw. `init()` also repairs a ledger that is not a
+  `LedgerStore` instance.
+- **`routing.failClosed` defaults to `false`.** Blocking is now opt-in and means
+  one thing: what to do when this plugin cannot produce a decision. The old
+  default escalated an internal `TypeError` into a refusal to serve the prompt,
+  including in manual mode. An internal defect now degrades to "do not route",
+  which is what would have happened with the plugin uninstalled.
+- **Quota safety is a separate, unconditional gate.** A provider configured for
+  probing but with no reading is still refused, regardless of `failClosed`,
+  because that is a spending decision rather than a response to plugin failure.
+- **A failed setup installs no prompt hook.** The entry point used to leave a
+  "minimal fail-closed hook" behind when `initialize()` threw, turning a failed
+  setup into an editor that refused every message. A location whose setup failed
+  now has no router.
+- **Hooks cannot throw or hang at the host boundary.** OpenCode does not catch a
+  plugin hook that throws and does not time out one that never settles; both
+  were verified against a real OpenCode 2.2.5 process. Every hook is wrapped in
+  `containHook()`, and routing runs against a 10 second deadline
+  (`withDeadline()`), after which the prompt proceeds unrouted.
+- **Routing only considers models the host can serve.** The router offered the
+  whole seed catalog as candidates. On a fresh install it switched a session to
+  `opencode-go/claude-haiku-5-5`, which that machine did not have, and the next
+  step failed with "Model unavailable". Candidates are now filtered to the
+  host's own model list, and an empty result means no route rather than an
+  unfiltered fallback.
+- **Model discovery is awaited on the first routing decision.** Discovery starts
+  in the background, so the first prompt after startup could route against an
+  empty list, which is the window in which an unservable model was reachable.
+
+### Added
+
+- `scripts/isolated-e2e.sh`: runs a real OpenCode in a temporary HOME with its
+  own config, data, and cache, installs a deliberately broken plugin, and
+  asserts only that OpenCode still answers a prompt. It refuses to run if any
+  path escapes the sandbox. Seven scenarios pass; two are recorded as host
+  limitations because they demonstrate that OpenCode has no defense at that
+  boundary. It runs against a loopback model stub that it starts itself, so it
+  needs no credentials and cannot be blocked by provider quota. A CI job runs
+  it on every push.
+- Regression suites: `test/shared-core-skew.test.ts` and
+  `test/shared-core-adoption.test.ts` (a stale core must not be adopted or break
+  routing), `test/fail-open-contract.test.ts` (an internal defect must not block
+  a prompt), `test/containment.test.ts` (hooks cannot throw or hang),
+  `test/setup-failure-never-blocks.test.ts` (a failed setup leaves no router),
+  `test/routing-servable-models.test.ts` (only servable models are candidates).
+- Six new entries in the mutation harness, one per corrected property.
+
+### Changed
+
+- `config.example.json` ships `failClosed: false`. Copying the example used to
+  inherit the old blocking default.
+- `scripts/cleanroom-verify.sh` now overlays the working tree onto `HEAD` and
+  reports how many uncommitted paths it covered. Before this, a run with work in
+  progress verified the previous commit and reported a pass that said nothing
+  about the changes being made.
+
+### Fixed in CI
+
+- The `host-survival` job now installs OpenCode V2 from its release binaries
+  instead of Homebrew, which does not exist on Ubuntu runners, and verifies the
+  download against the published `SHA256SUMS`. It also asserts the major version
+  is 2, so a future change to the installer cannot quietly point the suite at
+  OpenCode 1, whose hook surface and CLI differ.
+- `test/history.test.ts` no longer fails on the declared Node 20.12 floor, where
+  `node:sqlite` does not exist. The reader already degraded gracefully there; the
+  test now asserts that degradation and skips only the two cases that genuinely
+  require a real database. The suite passes on 20.12 with two honest skips rather
+  than two failures.
+- The eight calibration properties the mutation harness listed as unguarded now
+  have `test/calibration-policy.test.ts`. They were not missing guards, they were
+  pointed at a file that covers a different module, so the properties were never
+  exercised. All 69 mutations are now caught.
+
+## 0.6.0 — 2026-11-06
+
+Tier-3 work escalates by itself, paid from the plan-model pool.
+
+### Added
+
+- **Automatic flagship escalation** (`premium.mode: "auto"`, now the default):
+  a tier-3 task routes to the plan-model pool (Claude Sonnet 5.5, GPT-6.1 Sol)
+  without a per-task approval prompt while the pool has budget. When the pool
+  is spent, escalation degrades to an approval gate and the router serves the
+  task with the strongest ordinary model - a routing refusal never fails a
+  serviceable task. Tier-1 and tier-2 work always stays on the ordinary pool.
+- **Plan-model pool ledger**: monthly buckets tracking spend against the
+  OpenDesign plan-model allowance (Pro: $120/month), persisted across
+  restarts, separate from Design Plan credits and from other providers'
+  premium spend. The cap is read from the resolved plan catalog.
+- **Per-kind capability evidence**: `ModelRecord.capabilities` (coding,
+  debugging, agentic, architecture) from named benchmarks with provenance,
+  moving expected success by at most 4 points - enough to break a close
+  ranking, never enough to overrule a gate. Unscored models stay neutral:
+  absent evidence is unknown, not zero.
+- **Context probes** (`scripts/probe-plan-model-context.py`): end-to-end
+  800K-token verification with memorization sentinels and truncation checks.
+
+### Changed
+
+- GPT-6.1 Sol is scored (intelligence 55, Artificial Analysis v5.0.1) and
+  auto-approved for tier-3 escalation along with Sonnet 5.5, both by owner
+  instruction 2026-11-06. DeepSeek's observed 39.5 is untouched.
+- Catalog context windows raised from measured (328K/400K) to probed-verified
+  800K for both flagships after passed end-to-end probes on 2026-11-06; the
+  advertised 1M is deliberately NOT claimed, and Sol's long-context surcharge
+  above 272K was confirmed by the billed amounts.
+- MiniMax M3.1 Flash is classified as a Design Plan model. The gateway serves
+  it both inside and outside the Design Plan group and OpenDesign has not
+  published which applies, so the conservative choice is made: treating an
+  ordinary tier-2 workhorse as a plan model keeps routine work from draining
+  the escalation pool. Before this, a tier-3 task could land on it as a
+  premium model that a tier-2 flag itself made eligible.
+
+### Fixed
+
+- A partial settings file no longer produces a warning for every key it leaves
+  out. Absent and invalid are now distinct: a key you did not write takes its
+  default silently, while a key you did write with the wrong type still reports
+  the problem. Before this, a config that only set `providers.opendesign.plan`
+  emitted around fifty warnings and hid the ones that mattered.
+
+## 0.5.0 — 2026-11-06
+
+The router stops taking things that are not its to take.
+
+### Changed
+
+- **The router no longer renames sessions.** The `AUTO · OD · model — name`
+  title prefix is gone, for new chats and for every chat going forward. A chat
+  title is the owner's name for the work; status belongs in the panel, the
+  terminal footer and the decision record. The `ui.titleIndicator` setting is
+  removed and ignored if present.
+- `/router-title off` now strips prefixes this plugin wrote in earlier versions
+  and reports how many titles it cleaned. Titles that lost their separator are
+  left alone rather than rewritten.
+- Existing installs were repaired in place: every prefixed title was restored
+  from its separator, or from the session's own first user message when the
+  separator was already gone.
+
+### Added
+
+- Earlier review pass carried in this release: self-heal floor fallback, real
+  safety-margin enforcement, task-scoped rejection memory, strict failure joins,
+  request-clock threading, and metric-neutral decision notes.
+
+### Removed
+
+- **Sidebar project grouping (never released on npm).** The desktop app ships
+  this natively behind Settings -> Experimental -> Organize vertical sidebar ->
+  Projects (upstream commit `2bb4988`, after the installed v2.2.5). A plugin
+  copy would duplicate app code and fight the native version. The reasoning and
+  the prior-art notes are in `docs/SIDEBAR-GROUPING-TABLED.md`.
+
 ## 0.4.0 — 2026-11-05
 
 Self-healing routing, and npm distribution.

package/CONTRIBUTING.md

@@ -85,6 +85,13 @@ Explain why, not what. The diff already shows what. A commit that fixes a subtle
 bug should say how the bug showed up, because that part is not recoverable from
 the code.
 
+## Releasing
+
+See [README → Releases](README.md#releases). Publishing happens in CI through npm
+trusted publishing, so a maintainer never handles a long-lived token or relays a
+one-time password. Never publish from a working copy: the workflow checks the tag
+against `package.json` and re-runs the full suite first.
+
 ## Out of scope
 
 Likely to be declined:

package/README.md

@@ -7,6 +7,8 @@
 <p>Every other router asks what a token costs. This one asks what your plan allows, what is about to expire, and what you already paid for. MIT licensed.</p>
 
 <p>
+  <a href="https://www.npmjs.com/package/@qadanm/opencode-smart-router/v/0.4.0"><img src="https://img.shields.io/npm/v/%40qadanm%2Fopencode-smart-router?label=npm&color=cb3837&logo=npm&logoColor=white" alt="npm version"></a>
+  <a href="https://www.npmjs.com/package/@qadanm/opencode-smart-router"><img src="https://img.shields.io/npm/dm/%40qadanm%2Fopencode-smart-router?label=installs&color=cb3837&logo=npm&logoColor=white" alt="npm monthly installs"></a>
   <a href="LICENSE"><img src="https://img.shields.io/badge/License-MIT-blue.svg" alt="License: MIT"></a>
   <img src="https://img.shields.io/badge/OpenCode-2.x-000000.svg" alt="OpenCode 2.x">
   <img src="https://img.shields.io/badge/tests-759%20passing-2ea44f.svg" alt="759 tests">
@@ -23,6 +25,10 @@
 
 <br>
 
+<a href="assets/smart-router-demo.gif"><img alt="One session, 22 seconds: a prompt arrives, the router shows the decision card with its reasons and blocked candidates, the budget windows fill in, and a rejected turn moves itself to the next best model and continues" src="assets/smart-router-demo.gif" width="880"></a>
+
+<br>
+
 <picture>
   <source media="(prefers-color-scheme: dark)" srcset="assets/architecture-dark.svg">
   <img alt="Your prompt goes to the smart router, which classifies and prices it against your real quota before OpenCode sends the step to one provider" src="assets/architecture-light.svg" width="880">
@@ -188,15 +194,44 @@ panel's last decision records what was rejected and what took over.
 
 - Never buys credits, changes billing, or enables top-ups.
 - Never routes to a provider you did not approve.
-- Never spends premium quota without an explicit one-shot approval or a budget you set.
+- Never spends the plan-model pool without the budget you set; in the default
+  auto mode tier-3 work escalates to it automatically, and when it runs out the
+  router falls back to the strongest ordinary model instead of failing.
 - Never switches a session to a model OpenCode cannot currently resolve.
 - Never fails open on missing quota data when other providers already reported.
 - Never blocks your other work when one provider is at its limit.
+- **Never stops you sending a message.** If the router breaks, it is skipped.
+
+### When the plugin is wrong, the editor still works
+
+Routing is an optimization. OpenCode works without it, and it keeps working when
+the router is wrong. That is a contract, not a best effort:
+
+- an internal defect degrades to "do not route", never to "do not work". The
+  prompt is served with the model already selected and the reason is recorded in
+  warnings;
+- every hook is wrapped so it cannot throw into OpenCode, and routing runs
+  against a deadline so it cannot hold a prompt indefinitely;
+- if setup fails for a location, that location simply has no router;
+- a shared core left behind by a different plugin version is replaced rather
+  than reused.
+
+Blocking a prompt is available, but it is opt-in through
+`routing.failClosed: true`, and it covers plugin faults only. Quota safety is
+separate and always on: a provider configured for probing with no reading is
+refused, because that is a spending decision rather than a response to failure.
+
+This was made a contract the hard way. Version 0.6.0 turned an internal
+`TypeError` into a refusal to send any prompt, in every mode, and the plugin had
+to be removed by hand. The account is in
+[docs/INCIDENT-2026-11-06-prompt-lockout.md](docs/INCIDENT-2026-11-06-prompt-lockout.md),
+and the guarantee is tested against a real OpenCode process by
+`scripts/isolated-e2e.sh`, not only by unit tests.
 
 ## Status
 
-Version 0.4.0. Routing runs on the maintainer's machine in automatic mode;
-premium escalation and purchased-credit spending are disabled. Current behavior
+Version 0.6.1. Routing runs on the maintainer's machine in automatic mode;
+tier-3 tasks escalate to the plan-model pool automatically within its budget. Current behavior
 and open questions are in
 [docs/PRO-MIGRATION-2026-10-29.md](docs/PRO-MIGRATION-2026-10-29.md).
 
@@ -226,6 +261,7 @@ command panel closed.
 | `/router-premium-budget <usd>` | Set the premium budget |
 | `/router-reserve provider monthly weekly fiveHour` | Adjust provider reserves |
 | `/router-reset-metrics --confirm` | Reset metrics and routing state, not conversations |
+| `/router-title off` | Strip prefixes this plugin wrote into titles before 0.5.0 |
 
 Selecting a model in the desktop picker is detected and respected. The session
 reports `manual` until you run `/router-auto`.
@@ -245,7 +281,10 @@ It cannot loosen global restrictions.
 Worth knowing:
 
 - Mode: `balanced`
-- Premium: `ask`, $0 budget, one-shot approval available
+- Premium: `auto`, $120/month plan-model pool, one-shot approval available.
+  Auto mode escalates tier-3 tasks to the plan-model pool (Sonnet 5.5,
+  GPT-6.1 Sol) without asking, and degrades to an approval gate when the pool
+  is spent. Tier-1 and tier-2 work always stays on the ordinary pool.
 - Training-enabled and free promotional models: blocked
 - Purchased extra credits: blocked
 - Prompt text storage: disabled
@@ -287,7 +326,7 @@ npm run verify    # clean room: empty HOME, fresh install, typecheck, tests, bui
 npm run mutate    # prove each safety test fails when its property breaks
 ```
 
-759 tests across 63 files cover classification, eligibility, quota parsing,
+822 tests across 75 files cover classification, eligibility, quota parsing,
 credential discovery and its security properties, ledger accounting, pool
 reconciliation, documented peak pricing, cache-aware routing, premium protection,
 purchased-credit blocking, manual overrides, failure classification, configuration
@@ -330,6 +369,51 @@ A change that alters routing behavior needs a test that fails without it. See
 [CONTRIBUTING.md](CONTRIBUTING.md) for conventions, and [SECURITY.md](SECURITY.md)
 for reporting a vulnerability.
 
+## Releases
+
+Releases are cut from `main` and published by CI, not from a laptop. The package
+is published with [npm trusted publishing](https://docs.npmjs.com/trusted-publishers),
+which exchanges a GitHub Actions identity for a short-lived publish credential, so
+there is no `NPM_TOKEN` in this repository and no one-time password to relay.
+
+1. Bump `version` in `package.json` and add the matching `CHANGELOG.md` section.
+2. Run `npm run check`, commit, and push to `main`.
+3. Tag and push the tag:
+
+   ```bash
+   git tag -a v0.6.1 -m "0.6.1: <one line>"
+   git push origin main --tags
+   ```
+
+4. `.github/workflows/release.yml` checks the tag against `package.json`,
+   re-runs the full suite, and publishes.
+5. Create the GitHub release from the changelog section:
+
+   ```bash
+   gh release create v0.6.1 --title "0.6.1: <one line>" --notes-file notes.md
+   ```
+
+If the tag and `package.json` disagree, the workflow refuses to publish: a
+version that reaches the registry can never be replaced.
+
+### One-time npm setup
+
+Trusted publishing is authorized on the npm side, not in this repository. On the
+package settings page for `@qadanm/opencode-smart-router`, add a trusted publisher:
+
+| Field | Value |
+|---|---|
+| Provider | GitHub Actions |
+| Organization or user | `qadanm` |
+| Repository | `opencode-smart-router` |
+| Workflow filename | `release.yml` |
+| Environment | (leave empty) |
+
+Until that entry exists, the release workflow fails at the publish step with a
+`404 Not Found`, which is npm's way of saying the OIDC identity is not on the
+package's trusted list. The check, build, and tag verification still run and
+still gate the release.
+
 ## License
 
 MIT. See [LICENSE](LICENSE).

package/config.example.json

@@ -59,7 +59,7 @@
     "largeContextHysteresisPct": 35,
     "maxEscalationsPerTask": 2,
     "preferSameProvider": true,
-    "failClosed": true,
+    "failClosed": false,
     "protectEscalationPool": true,
     "longContextFloorTokens": 1000000
   },
@@ -90,7 +90,6 @@
     "slowRouteMultiple": 2.5
   },
   "ui": {
-    "titleIndicator": false,
     "notifications": true,
     "compactMode": false
   },