opencode-harness-kit 0.2.1 → 0.3.4

package/CHANGELOG.md

@@ -0,0 +1,303 @@
+# Changelog
+
+All notable changes to `opencode-harness-kit` are recorded here. The format
+follows [Keep a Changelog](https://keepachangelog.com/en/1.1.0/), and the
+project uses [semantic versioning](https://semver.org/spec/v2.0.0.html).
+
+This kit sits between a provider catalog and a config file: capability and
+privacy metadata in, a working harness out, with a command to re-check that
+the metadata has not rotted. Most of the entries below are therefore about
+**silent failures** — a config that looks correct while the thing it promises
+is not happening.
+
+---
+
+## [Unreleased]
+
+## [0.3.4] — 2026-10-20
+
+> 0.3.2 and 0.3.3 were never published. 0.3.2 carried the fix below but the
+> publish path it needed was not set up yet; 0.3.3 was **staged** by a
+> diagnostic that turned out to have side effects, and a staged version cannot
+> be published over directly. Both are folded into this release.
+
+### Added
+
+- **The publish job can use npm trusted publishing (OIDC), so the repository
+  needs no npm secret at all.** npm exchanges the runner's OIDC identity for a
+  short-lived publish token, scoped to the workflow that asked for it. There
+  is no long-lived token in the repository to leak, rotate or expire - and npm
+  is revoking the legacy automation tokens anyway.
+  - The job installs npm 11.5.1 or later, which is what OIDC needs and what
+    node 22 does not ship.
+  - The registry is configured by hand instead of through `setup-node`'s
+    `registry-url`, because that helper always writes an auth line, and an
+    empty auth line makes npm believe it is authenticated and skip the OIDC
+    exchange entirely.
+  - The credential check is conditional: `npm whoami` is the right check for
+    a token, and **the wrong** one for OIDC, because an OIDC identity is not
+    reflected by `npm whoami` at all. Asking it unconditionally failed a
+    publish that would have worked.
+  - A token still works: if `NPM_TOKEN` is set, the job uses it.
+- **`npm run release:trust` configures the trusted publisher from the command
+  line.** npm's web form has four fields and npm does not validate what you
+  type into it, so a typo is discovered at publish time and reported as
+  "OIDC token exchange error - package not found" - which reads like a missing
+  package. npm ships `npm trust` for exactly this, and it can read what is
+  stored, so the configuration is now compared against the identity the
+  workflow will actually present and replaced when it differs.
+- The workflow is named after its file, so the value that has to be typed on
+  npmjs.com is the value the Actions tab shows.
+
+### Fixed
+
+- **The publish preflight rejected the very token it wanted.** It called `npm
+  access list packages` and required the package to appear in the list. That
+  command lists the packages a *user* has access to, which needs
+  account-level read permission - and a granular token scoped to one package
+  is answered **403** for it. A correctly scoped publish token would have
+  been refused, and the user sent looking for a problem that was in the
+  check. Only `npm whoami` is checked now, which any valid token passes; the
+  write permission is proven by the publish itself, and its failure message
+  says what was missing.
+
+## [0.3.1] — 2026-10-20
+
+### Fixed
+
+- **`oc-harness verify` failed a good install over somebody else's skill
+  naming.** It checked every skill directory in the config and required the
+  `name:` in the frontmatter to match the directory name. OpenCode identifies
+  a skill by its **directory**, not by the frontmatter, so a skill whose two
+  differ loads perfectly well. An install carrying 220-odd community skills
+  reported five failures that were not failures at all, and the command is
+  the one people put in a commit hook.
+- The check now covers the kit's **own** skills, which do have to agree with
+  their directory — the loop refers to them by name, and one whose frontmatter
+  drifted is a step of the loop that will not load. Everyone else's skills are
+  counted and left alone.
+
+## [0.3.0] — 2026-10-20
+
+### The fleet
+
+The default profile is no longer a model. It is **ten jobs, ten free models,
+one session**.
+
+- `lanes.json` declares the jobs and says nothing about models. A profile
+  declares which lanes it enables and which model serves each one.
+- Each lane is an OpenCode agent with its own model, its own context window
+  and its own effort level. Delegating a job is how you use a second model;
+  switching models by hand never happens.
+- **A lane's model is a chain, not a pin.** Free routes get withdrawn in
+  days — four of this kit's five original profiles were dead within a month —
+  so every lane lists the models that could serve it, best first, and the
+  router walks the same chain when a route is spent.
+- **The chain is ordered inside the profile's own ceiling.** A profile
+  declares `maxRetention`; a lane takes the first rung at or below it. One
+  fleet file therefore covers ten jobs at two privacy levels: `fleet-safe`
+  (zero-retention only, the default) and `fleet-all` (every lane on the model
+  built for it).
+- `retention.yml` was rewritten against the OpenCode privacy page: seven new
+  entries, four dead models dropped, and a third value — `retained` — for
+  Beam, which does not train but keeps prompts for 30 days and is therefore
+  not zero-retention.
+- `oc-harness lanes` reports which lanes were actually used, so the fleet can
+  be cut to what earns its place rather than defended on principle.
+
+### Fixed — the measurement that decided the fleet was counting the wrong thing
+
+- **The lane counter counted events, not messages.** `opencode` emits
+  `message.updated` several times for a single assistant message — created,
+  streamed, completed — so one request was recorded three times, and
+  "build 7 calls" was really two messages. A lane record now carries the
+  message id and the reader merges on it.
+- **`chat.message` could not attribute a request to a lane.** On OpenCode
+  1.19 its `agent` field is `undefined`. It was the natural place to count a
+  call before the model answered, so that a call that never came back still
+  counted; it cannot, so it is no longer used for this.
+- **`oc-harness lanes` read the wrong file.** The plugin writes the telemetry
+  to the project root; the command read `process.cwd()`. Run from a
+  subdirectory, it reported "never called" for every lane — a fleet that
+  looked entirely unused.
+- **The unit is now named what it is.** A request is one assistant message,
+  which is one round trip to the provider, and a lane that makes three tool
+  calls makes three of them. The column says `reqs`.
+
+### Fixed — the fleet did not survive a spent route
+
+- **A spent session model collapsed the whole fleet onto one model.** The
+  router replaced the entire profile, which set every lane of the fallback
+  profile onto one model and left the other five lanes pointing at routes
+  that were also spent — the opposite of what the function's own docstring
+  promised. A fleet now keeps its lanes: the session model moves, each lane
+  moves down its own chain, and the router publishes the model it chose so
+  the effort hook pins the session to that one rather than back to the base
+  model that is spent.
+- **`effectiveRetention().worst` had two meanings.** It was a retention class
+  on one path and a model id on the other, and the header prints it as a
+  model name. A fleet that resolved to a `retained` lane said "but retained
+  makes the whole file retained". It is always a model id now.
+
+### Fixed — the numbers on screen did not mean what the label said
+
+- **The "input cap" was the base model's, labelled as the narrowest lane's.**
+  The config header, `oc-harness show` and `oc-harness list` each printed a
+  different number next to a prune threshold derived from another one, and
+  the README quoted a fourth. There is now one number: the input cap the
+  config can actually use.
+- **`profiles/index.json` carried a context of 200,000 for `fleet-all` and
+  `big-pickle`.** The cap that governs is 160,000, and the list recomputes it
+  from the profile rather than trusting the copy.
+- **The safe fleet inherited the unrestricted fleet's chain order**, so
+  bumping into the zero-retention ceiling dropped every lane onto whatever
+  rung happened to be third. Two lanes were genuinely mis-served: `locate`
+  asked for `low` effort and landed on a model that publishes no effort
+  levels at all, so the request was dropped on the floor; and `vision` landed
+  on the same model at 9–13 s when MiniMax reads the same image in 7.7 s.
+  The safe fleet now orders its own chains.
+- **The docs said four lanes lost their model under the ceiling. Six do.**
+  The number is now computed and asserted by a test, and the two fleets'
+  reserves (24,000 and 19,200) are both stated.
+- **Three measured latencies were wrong.** GLM 5.3 is 4.4–5.8 s on the work
+  a lane does, not 16–27 s; MiMo reads an image in 19.2 s, not 8 s; and
+  Nemotron Super VL does *not* hang on a tool read, only on a file attached
+  to the prompt — it answered a `grep` sweep in 4.2 s. The rule is "never
+  hand it a file as an attachment", not "it is broken".
+
+### Fixed — a shadowing agent file could hide in four places
+
+- **An agent markdown file replaces the config entry wholesale**, and a file
+  with no `model:` line makes the agent inherit the caller's model. Both were
+  measured on OpenCode 1.19 with the plugin out of the way (`--pure`): the
+  file's model wins. The model therefore lives in `agent.<lane>` in the
+  generated config, and the prompt bodies live in `prompts/`.
+- **`verify` checked one directory.** OpenCode reads `agent/` and `agents/`
+  alike, in the config dir and in `.opencode/`, and the plural one wins on a
+  name clash — so a shadow in the singular directory was invisible. All four
+  are scanned now.
+- `oc-harness install` removes the agent files an earlier version of this kit
+  left behind, from either directory.
+
+### Fixed — three lists of lane names, drifting apart
+
+- **`profiles/schema.json` named an `audit` lane that no longer existed and
+  was missing `longform`.** With `additionalProperties: false` the shipped
+  fleet profiles violated the shipped schema, and nothing said so.
+- **`lanes.json` pointed at `./lanes.schema.json`, which was not in the
+  repository.** It is now, and a test pins the lane table, the profile schema
+  and `LANE_NAMES` to each other.
+- `MANAGED_ROLES` was a hand-written list of nine lanes that had drifted, and
+  was no longer read by anything. It is derived from the lane table now.
+
+### Removed
+
+- `agents/` — the prompt bodies moved to `prompts/`, where a file cannot
+  shadow a config entry.
+- The `space-bunny-free`, `x-preview-f-free`, `deepseek-v4-flash-free`,
+  `muse-spark-1.3-contributor-free` and `muse-split-free` profiles. Their
+  models were withdrawn from the provider. `oc-harness list` now names them
+  under **no longer on the tier**, with what replaced them.
+
+### Migration
+
+0.2.x → 0.3.0 changes the profile ids, the config layout and the meaning of
+the telemetry. See [MIGRATION.md](MIGRATION.md).
+
+---
+
+## [0.2.1] — 2026-10-19
+
+Shipped: the `space-finish` skill's duplicated steps removed, one demo video
+instead of three, and the package metadata pointed at the right repository.
+
+### Fixed
+
+- **`oc-harness use` deleted hand-tuned config.** The JSONC parser returned
+  `{}` for a config it could not read, so the merge was onto an empty base
+  and the user's own settings were written back out deleted, with no message.
+  The parser walks the text once, tracking string state, so comments and
+  trailing commas are both handled correctly.
+- **Five further defects that lose data, break CI, or send users nowhere** —
+  including `--dryrun` being accepted and then ignored, so a dry run silently
+  overwrote the real config.
+- **`oc-harness agents` reported "no project agents found" forever** while
+  none of them were installed at all, because it scanned only the project and
+  the kit's own agents live in the global config directory.
+- The demo video was replaced with a real recorded take: the slide renderer
+  was describing a test count it was not measuring.
+
+---
+
+## [0.2.0] — 2026-10-15
+
+Shipped: the evaluation harness, integration tests that load the built plugin
+the way OpenCode does, npm packaging, `oc-harness link`, `oc-harness agents`,
+the tool-output caps, and the `scout` → `locate` agent rename.
+
+### Fixed
+
+- **The guard matched text a shell would not run.** `cd repo && rm -rf /`,
+  `sudo rm -rf /` and `bash -c "rm -rf /"` are now caught, while
+  `git log --grep push` and `echo push` are not. A heredoc body is data, not
+  commands.
+- **`tool_output` was pinned at 200 lines / 8 KB**, about 6× tighter than
+  OpenCode's default. It truncated 415 of 2,966 `read` calls and 418 calls
+  were issued with an `offset` to fetch the rest: the truncation *causes* the
+  re-read, one for one, at a measured 17.3 s median per request. The cap is
+  now the default, and pruning rather than truncation bounds the context.
+- **`tail_turns: 15` was inert config.** Compaction is keyed off the full
+  context ceiling that a normal session never reaches; it had fired zero times
+  in 153 measured sessions. The key was dropped rather than tuned.
+- **The permission-pressure figure was a transcription of the allowlist**, so
+  the measured number was whatever the copy looked like when it was last
+  edited. `baseline` imports the real list now.
+- **The `46.6%` that effort tiering saves was a projection, not a
+  measurement** — cross-model medians, because the tiering had never run on
+  the zero-retention model. `eval` now prints it as `PROJECTION` with
+  `measured: no`.
+
+### Known at the time
+
+- `use --refresh` and the eval harness could not reach a user's real
+  install from CI, so the live half of the gate ran only locally.
+- The `scout` agent shadowed OpenCode's own built-in `scout` subagent, which
+  exists to read a dependency's source.
+
+---
+
+## [0.1.0] — 2026-10-12
+
+First release: profiles, the effort resolver, the config compiler, the
+`oc-harness` CLI, the four plugins (guard, prune, router, telemetry),
+`GLOBAL.md`, the agent files, the skills, the test suite and
+`baseline.mjs`.
+
+### What it promised
+
+A model-agnostic harness kit. Everything model-specific lives in
+`profiles/<id>.json`; no plugin in `src/` names a model. Switching models is
+a data edit, not a code change.
+
+### Fixed in later releases
+
+The defects in 0.1.0 and 0.2.0 are listed under those releases, because the
+fix and the version that carried it are the same fact. Two of them were
+silent in the worst way:
+
+- **`oc-harness use` could write back a config it had failed to read**, with
+  the user's own settings deleted and no message. Fixed in 0.2.1.
+- **The default profile pointed at a model the provider had withdrawn.** The
+  config kept validating and kept claiming zero-retention. The tier turned
+  over again within a month, which is why 0.3.0 gives every lane a chain
+  instead of a pin.
+
+---
+
+[Unreleased]: https://github.com/pham-canh/space-harness/compare/v0.3.4...HEAD
+[0.3.4]: https://github.com/pham-canh/space-harness/compare/v0.3.1...v0.3.4
+[0.3.1]: https://github.com/pham-canh/space-harness/compare/v0.3.0...v0.3.1
+[0.3.0]: https://github.com/pham-canh/space-harness/compare/v0.2.1...v0.3.0
+[0.2.1]: https://github.com/pham-canh/space-harness/compare/v0.2.0...v0.2.1
+[0.2.0]: https://github.com/pham-canh/space-harness/compare/v0.1.0...v0.2.0
+[0.1.0]: https://github.com/pham-canh/space-harness/releases/tag/v0.1.0

package/GLOBAL.md

@@ -2,7 +2,7 @@
 
 You are working inside a harness that has already been tuned for the model in
 use. Everything the config and the hooks can enforce is enforced already: the
-tools, the permissions, the command allowlist, the guard, the per-role effort
+tools, the permissions, the command allowlist, the guard, the per-lane effort
 and step caps. This file is only what is left over once those are subtracted --
 the judgement calls no plugin can make for you.
 
@@ -35,12 +35,47 @@ Everything else is reversible and yours. Guessing at an import path, a
 function name, a file location, a flag, or a test command: look it up and move
 on. If you looked and could not find it, say that.
 
+A `git push` is not one of those decisions. Committing the branch you were
+asked to finish and pushing it is the last step of the job, not something to
+escalate. Push, then say what you pushed.
+
 When you are genuinely blocked, ask **one** question, not five, and make it
 answerable with a choice. "Should I use `X` or `Y`" is a question. "What are
 your thoughts on the architecture" is a delay.
 
 Never ask what a `grep` would answer.
 
+## Send the work to the lane that has it
+
+Each lane is a separate agent on its own model, with its own context window
+and its own effort level. You do not switch models: you hand the job over.
+
+| The job | The lane | What it owns |
+|---|---|---|
+| Turn a vague request into a plan | `plan` | the requirement, the tasks, the verify commands |
+| Write and verify the change | `build` | the edit, the test run, the report |
+| Read a change in two passes | `reviewer` | blockers first, then quality |
+| Find where something lives | `locate` | `file:line` evidence, and nothing else |
+| Run the project's checks | `verify` | the real output of a real command |
+| Read an image, a screenshot, a diagram | `vision` | what is in the picture |
+| Read a long document or a spec | `doc` | what it says, and where the code disagrees |
+| Produce JSON or a JSON schema | `schema` | output a program can parse |
+| Sweep a lot of code mechanically | `sweep` | matches, with no judgement |
+| Read a megabyte, or write alternatives | `longform` | very long context, or several answers |
+
+Rules for the hand-off:
+
+- **Delegate the job, not the whole session.** A lane returns an answer; you
+  keep the thread. Do not hand over the conversation.
+- **One lane per job.** `sweep` finds the matches, `reviewer` judges them.
+  Asking one lane to do both gets you the weaker answer.
+- **Do the lane's job yourself when it is small.** Handing a one-line lookup to
+  a separate model costs a round trip and buys nothing. The lanes exist for
+  the jobs that need their own context window, their own model, or their own
+  effort level.
+- **A lane that reports a blocker is the loop working.** Read the file:line it
+  gives you and fix that.
+
 ## Locate before you read
 
 On the measured install, `read` ran 2,966 times against 135 `grep` and 114
@@ -54,7 +89,9 @@ to avoid this.
 
 ## Never
 
-- Ask before `git push`, publish, or opening a pull request. Publishing is a human decision.
+- Force-push, rewrite published history, delete a branch or tag, or publish to
+  a registry (`npm publish`, `gh release`). Push without being asked; ask before
+  any of these, because they are the ones with no undo.
 - Commit a secret. If you find one, report it; do not move it.
 - Read or write `.env`, `.npmrc`, `.netrc`, `.git-credentials`. The guard
   blocks these and will keep blocking them.
@@ -77,12 +114,12 @@ next session more than a small code change.
 
 ## Reasoning budget
 
-Per-role effort and step caps are set in `opencode.json` and enforced by the
-harness, not by anything written here. `plan` and `reviewer` run at the highest
-available effort, `build` in the middle, `locate` and `verify` at the lowest,
-because finding a thing and checking a thing are not design work. If you are
-playing one of the low-effort roles and feel yourself designing an
-architecture, stop and report the evidence instead.
+Per-lane effort and step caps are set in the generated config and enforced by
+the harness, not by anything written here. `plan` and `reviewer` run at the
+highest available effort, `build` runs at the top of the range too, and
+`locate` and `verify` at the lowest, because finding a thing and checking a
+thing are not design work. If you are playing one of the low-effort lanes and
+feel yourself designing an architecture, stop and report the evidence instead.
 
 ## When context grows
 
@@ -96,10 +133,10 @@ you need from it and get only that.
 
 ## Model etiquette
 
-If a request fails with a rate limit or a usage cap, **stop**. Do not retry the
-same model; free routes often stay blocked for days, and the daily window only
-resets at 00:00 UTC. Report it and ask for a new session. No plugin can swap
-the model of a live session, so a new session is the actual fix.
+If a request fails with a rate limit or a usage cap, **stop**. Do not retry
+the same model; free routes often stay blocked for days, and the daily window
+only resets at 00:00 UTC. Report it and ask for a new session. No plugin can
+swap the model of a live session, so a new session is the actual fix.
 
 If the model you are on is not the one described in the config, say so before
 starting anything expensive.
@@ -107,4 +144,4 @@ starting anything expensive.
 ## When you are stuck
 
 Say what you tried, what you expected, and what happened instead. A precise
-report is more useful than a third attempt at the same approach.
+report is more useful than a third attempt at the same approach.

package/MIGRATION.md

@@ -0,0 +1,140 @@
+# Migrating from 0.2.x to 0.3.0
+
+0.3.0 replaces the single-model profile with a **fleet**: ten jobs, ten free
+models, one session. Three things change for a user — the profile ids, the
+config layout, and what the telemetry file means.
+
+Nothing in 0.2.x is deleted by the upgrade in a way that loses data. The
+profiles whose models were withdrawn from the provider are gone because the
+models are gone; `oc-harness list` names them under **no longer on the tier**
+and says what replaced them.
+
+---
+
+## 1. The profile ids changed
+
+The old ids named a model. The new ones name a fleet, because that is what
+they are.
+
+| 0.2.x | 0.3.0 | note |
+|---|---|---|
+| `space-bunny-free` (default) | `fleet-safe` (default) | the model was withdrawn; this is the safe fleet |
+| `muse-split-free` | `fleet-all` | the unrestricted fleet: every lane on the model built for it |
+| `deepseek-v4-flash-free` | `deepseek-v4.1-flash-free` | the old one was withdrawn |
+| `x-preview-f-free` | — | withdrawn |
+| `muse-spark-1.3-contributor-free` | — | withdrawn |
+| `longcat-2.5-preview-free` | `longcat-2.5-preview-free` | unchanged |
+| — | `beam-free`, `big-pickle`, `glm-5.3-flash-free`, `mimo-v2.6-flash-free`, `minimax-m3.1-flash-free`, `nemotron-3-ultra-free`, `nemotron-3.5-lightning-free`, `nemotron-3.5-super-vl-free` | new single-model profiles for the rest of the tier |
+
+```bash
+oc-harness use fleet-safe
+oc-harness install
+oc-harness verify
+```
+
+If you had `OPENCODE_HARNESS_PROFILE` set, it names the old id and will not
+resolve:
+
+```bash
+export OPENCODE_HARNESS_PROFILE=fleet-safe
+```
+
+A single-model profile still exists for every model, so a user who wants one
+route and wants to know exactly which one still has it:
+
+```bash
+oc-harness use glm-5.3-flash-free
+```
+
+## 2. The config layout changed
+
+The prompt bodies moved out of the agent files and into `prompts/`, and the
+model moved into the config.
+
+| | 0.2.x | 0.3.0 |
+|---|---|---|
+| where the model lives | `agents/<lane>.md` frontmatter | `agent.<lane>.model` in the generated config |
+| where the prompt lives | the same file | `prompts/<lane>.md`, referenced with `{file:...}` |
+| lanes | five: plan, build, reviewer, locate, verify | ten: the five above plus vision, doc, schema, sweep, longform |
+| `compaction.reserved` | sized by the base model | sized by the **narrowest** lane model |
+
+**The move was not cosmetic.** An agent markdown file of the same name
+replaces the config entry wholesale, and a file with no `model:` line makes
+the agent inherit the caller's model. Both were measured on OpenCode 1.19.
+With the model in the config, a lane cannot be quietly re-pointed at a
+different provider by a file that happens to share its name.
+
+`oc-harness install` removes the old agent files for you, and `oc-harness
+verify` fails if one of them is still shadowing a lane.
+
+The lanes are OpenCode agents, so a delegation reaches a different model
+inside one session. That was measured, not assumed:
+
+```
+agent=build    modelID=minimax-m3.1-flash-free
+agent=reviewer modelID=nemotron-3-ultra-free
+agent=build    modelID=minimax-m3.1-flash-free
+```
+
+**Your own agent files are not touched** unless they are named after a lane
+this kit manages. A file called `scout.md` or `deploy.md` is left alone.
+
+## 3. The telemetry file means something different
+
+`.space/telemetry.jsonl` is gitignored and local. In 0.2.x the kit wrote
+`session`, `guard` and `provider-event` records to it; in 0.3.0 it also
+writes one record per assistant message, carrying the lane, the model and the
+duration.
+
+If you were reading that file, the units changed:
+
+- **0.2.x and earlier: one record per event.** `opencode` emits
+  `message.updated` several times per message, so a request was recorded
+  three times.
+- **0.3.0: one request = one assistant message**, deduplicated by message
+  id. A lane that makes three tool calls makes three requests.
+
+Records written by an older version have no message id and are **not
+counted** — the number they produce is inflated threefold. Delete the file
+to start clean:
+
+```bash
+rm -f .space/telemetry.jsonl
+```
+
+`oc-harness lanes` says which file it read, and how many records it could not
+count.
+
+## 4. The privacy gate now reads the whole fleet
+
+`oc-harness use` refuses a profile whose models have no stated retention
+terms, and demands `--yes` when any model in the fleet may train. In 0.2.x it
+read the **base** model's terms only, so a profile whose base was
+zero-retention while its roles were may-train installed with no warning at
+all.
+
+For `fleet-safe` nothing changes: every model is zero-retention, so there is
+nothing to consent to. For `fleet-all` the consent is now asked for, and the
+warning names the model responsible.
+
+## 5. What did not change
+
+- The plugins keep their hook names and their behaviour: `guard`,
+  `prune`, `router`, `telemetry`.
+- `retention.yml` keeps its shape, with one new value: `retained`, for a
+  provider that does not train but keeps the data for a stated window.
+- `GLOBAL.md` and the skills are installed the same way.
+- `oc-harness baseline`, `oc-harness doctor`, `oc-harness agents`,
+  `oc-harness show` and `oc-harness refresh` keep their arguments.
+
+## 6. If you want the old behaviour
+
+Install a single-model profile instead of a fleet. One route, one model, and
+you know exactly which one:
+
+```bash
+oc-harness use glm-5.3-flash-free
+```
+
+The fleet is a default, not a requirement. `docs/ROUTING.md` explains which
+lane each model owns and what was measured.