refactor: no harness is special — delete the name-keyed branches

CHANGELOG.md

@@ -1,5 +1,38 @@
 # Changelog
 
+## Unreleased
+
+### No harness is special: eleven name branches become table rows
+
+BREAKING. `LocalHarness` was a private three-member vocabulary (`'claude' | 'codex' | 'opencode'`) that spelled one harness differently from every other layer in the stack. It is now a narrowing of the shared `HarnessType`: **`'claude'` is renamed to `'claude-code'`**. `claude` remains the EXECUTABLE name and lives only in the harness table's `command` field.
+
+Callers to update: `runLocalHarness({ harness })`, `harnessInvocation(harness, …)`, `runWorktreeHarness({ harness })`, `agenticGenerator({ harness })`, `driverLoopGenerator({ harness })`, `createInProcessExecutor({ harnesses })`, `AuthoredHarness.harness`, and the `AGENT_RUNTIME_LOCAL_HARNESSES` env list. Anything that passed `'claude'` passes `'claude-code'`; `codex` and `opencode` are unchanged.
+
+Deleting the alias removed `materializerHarness()` outright — a `LocalHarness` is now handed straight to the profile materializer with no translation.
+
+**Reasoning effort now reaches claude-code and opencode.** `runWorktreeHarness` used to hard-REFUSE any profile carrying `model.reasoningEffort` unless the harness was codex, and `harnessInvocation` silently dropped it. Both read one capability table now:
+
+- `claude-code` → `--effort <low|medium|high|xhigh|max>`; canonical `ultracode` is native `max`.
+- `opencode` → `--variant <variant>`; canonical `ultracode` is `max`.
+- `codex` → `-c model_reasoning_effort="…"`, unchanged.
+
+A level a harness genuinely cannot express is still refused, and the refusal now lands in the pre-flight admission check (before any worktree exists) because the guard and the argv builder read the same rows. claude-code refuses `none` and `minimal` (its `--effort` has no such level); opencode refuses `none` (thinking-off is the absence of the flag).
+
+**Permission bypass is a property of the workspace, not of one CLI.** `dangerouslySkipPermissions` was tested against `'claude'` in four places; three were caller-side duplication of the fourth, which dropped the flag for every other harness with no error. Each harness now declares its own bypass argv:
+
+- `claude-code` → `--dangerously-skip-permissions` (unchanged).
+- `codex` → `--sandbox workspace-write -c approval_policy="never"`. NEW: a codex worker in a disposable worktree previously had its bypass request silently dropped. It edits non-interactively now and **keeps its OS sandbox** — writes stay confined to the workspace. `--dangerously-bypass-approvals-and-sandbox` is deliberately NOT used: `codex exec` has no approval gate to stall on (`-a/--ask-for-approval` exists only on the top-level `codex`), so it would surrender the sandbox for nothing, and the sandbox is what keeps a worker's blast radius equal to its worktree.
+- `opencode` → nothing; `opencode run` has no approval gate.
+- Reproducible Codex is unchanged: its controlled config already pins `approval_policy="never"` with the sandbox intact, so the blanket bypass flag is suppressed rather than layered on top. Reproducible argv is byte-identical to 0.118.0.
+
+**Other name branches replaced by rows, with no behaviour change:**
+
+- `projectCandidateSystemPrompt`'s four-arm `switch (plan.harness)` and its conflicting-argument guard are now one `HARNESS_SYSTEM_PROMPTS` row per harness (executable + projection + conflict predicate). The guard's fail-OPEN default for an unlisted harness is gone: no row means the projection is refused.
+- `harness === 'cli-base'` was re-derived at three call sites; it is now `harnessRunsAgent` / `agentHarness` in `src/runtime/harness-role.ts`.
+- New exports on `@tangle-network/agent-runtime/mcp`: `DEFAULT_LOCAL_HARNESS`, `LOCAL_HARNESSES`, `localHarnessExecutable`, `harnessSupportsReasoningEffort`.
+
+Deliberately KEPT: the `codexReproducible && harness !== 'codex'` guards (a codex-specific public option asserting caller self-consistency, not behaviour varying by name), and every `ExecutorConfig.backend` switch (a discriminated-union tag naming the materialization contract, not a harness name). Both now say so at the site.
+
 ## 0.118.0
 
 ### pi runs through the bridge, like every other harness

bench/gen3-config.json

@@ -41,7 +41,7 @@
   "repsPerInstance": 2,
   "premeasuredBaselinePath": "/tmp/claude-1000/-home-drew-code-supervisor-lab/f06fd156-042a-4ef9-bd88-f2ec7f52b90c/scratchpad/hh/gen3/premeasured-baseline.json",
   "maxShots": 3,
-  "proposerHarness": "claude",
+  "proposerHarness": "claude-code",
   "proposerTimeoutMs": 2400000,
   "analystModels": [
     "glm-5.2",

bench/gen4-config.json

@@ -41,7 +41,7 @@
   "repsPerInstance": 2,
   "premeasuredBaselinePath": "/tmp/claude-1000/-home-drew-code-supervisor-lab/f06fd156-042a-4ef9-bd88-f2ec7f52b90c/scratchpad/hh/gen4/premeasured-baseline.json",
   "maxShots": 3,
-  "proposerHarness": "claude",
+  "proposerHarness": "claude-code",
   "proposerTimeoutMs": 2400000,
   "analystModels": [
     "glm-5.2",

bench/gen5-config.json

@@ -41,7 +41,7 @@
   "repsPerInstance": 2,
   "premeasuredBaselinePath": "/tmp/claude-1000/-home-drew-code-supervisor-lab/f06fd156-042a-4ef9-bd88-f2ec7f52b90c/scratchpad/hh/gen5/premeasured-baseline.json",
   "maxShots": 3,
-  "proposerHarness": "claude",
+  "proposerHarness": "claude-code",
   "proposerTimeoutMs": 2400000,
   "analystModels": [
     "glm-5.2",

bench/src/quant-arena/quant-loop.mts

@@ -88,11 +88,11 @@ export const PINNED_BASELINES: Record<string, GenerateSignals> = {
 /** The two demo author seats: the plain author and the quant lens. */
 export function defaultQuantProposers(): ProposerSpec[] {
   return [
-    { name: 'default-author', profile: 'default-author.profile.json', harness: 'claude' },
+    { name: 'default-author', profile: 'default-author.profile.json', harness: 'claude-code' },
     {
       name: 'quant-researcher',
       profile: join(QUANT_PROFILES_DIR, 'quant-researcher.profile.json'),
-      harness: 'claude',
+      harness: 'claude-code',
       lens:
         'Favor ONE economically-motivated effect (trend, mean reversion, vol targeting) with few parameters. ' +
         'State the regime in which it should work and keep turnover low enough that 15bps a side cannot eat the edge.',
@@ -218,7 +218,7 @@ async function claudeShot(opts: {
   const res = await run('claude', argv, {
     stdin: opts.prompt,
     cwd: opts.cwd,
-    env: proposerShotEnv('claude'),
+    env: proposerShotEnv('claude-code'),
     timeoutMs: opts.timeoutMs,
   })
   if (res.code !== 0) {

bench/src/rollout-ledger/backfill-swe-arena.test.mts

@@ -158,7 +158,7 @@ async function buildFixtureTree(): Promise<void> {
       candidateIndex: 0,
       shot: 1,
       maxShots: 3,
-      harness: 'claude',
+      harness: 'claude-code',
       model: null,
       promptSha256: 'sha256:abc',
       startedAt: '2026-07-22T19:27:18.350Z',
@@ -256,7 +256,7 @@ describe('backfillSweArena', () => {
     expect(proposer?.messages).toHaveLength(2)
     expect(proposer?.outcome.reward).toBe(1)
     expect(proposer?.outcome.reward_source).toBe('swe-arena-official-judge/candidate-resolved-fraction')
-    expect(proposer?.policy).toMatchObject({ harness: 'claude', model: 'claude-fable-5' })
+    expect(proposer?.policy).toMatchObject({ harness: 'claude-code', model: 'claude-fable-5' })
     expect(proposer?.cost.tokens_out).toBe(462)
     expect(proposer?.task).toMatchObject({ suite: 'swe-arena-proposer', rep: 1 })
 

bench/src/swe-arena/activation.test.mts

@@ -245,9 +245,9 @@ describe('activation-predicate prefilter', () => {
 
   it('kills a candidate without .improve/activation.json (stage activation-predicate) and passes one WITH it', async () => {
     const proposers: ProposerSpec[] = [
-      { name: 'with-predicate', harness: 'claude' },
-      { name: 'without-predicate', harness: 'claude' },
-      { name: 'invalid-predicate', harness: 'claude' },
+      { name: 'with-predicate', harness: 'claude-code' },
+      { name: 'without-predicate', harness: 'claude-code' },
+      { name: 'invalid-predicate', harness: 'claude-code' },
     ]
     const gen = fanOutLoopsGenerator(config(proposers), {
       author: async (proposer, args) => {
@@ -284,7 +284,7 @@ describe('activation-predicate prefilter', () => {
   })
 
   it('does not require a predicate when the gate is off (gen-4 behavior unchanged)', async () => {
-    const cfg = config([{ name: 'legacy', harness: 'claude' }])
+    const cfg = config([{ name: 'legacy', harness: 'claude-code' }])
     cfg.activationGate = false
     const gen = fanOutLoopsGenerator(cfg, {
       author: async (_p, args) => {

bench/src/swe-arena/gepa-seat.test.mts

@@ -58,7 +58,7 @@ describe('validateGepaSeat', () => {
     expect(() => validateGepaSeat(seat())).not.toThrow()
     expect(() => validateGepaSeat(seat({ engine: 'omni', maxMetricCalls: 8 }))).not.toThrow()
     expect(isGepaSeat(seat())).toBe(true)
-    expect(isGepaSeat({ name: 'x', harness: 'claude' })).toBe(false)
+    expect(isGepaSeat({ name: 'x', harness: 'claude-code' })).toBe(false)
   })
 
   it('requires a surface inside the declared change-space', () => {
@@ -68,7 +68,7 @@ describe('validateGepaSeat', () => {
   })
 
   it('rejects harness-seat fields on an engine seat instead of silently ignoring them', () => {
-    expect(() => validateGepaSeat(seat({ harness: 'claude' }))).toThrow(/'harness' belongs to harness-authored/)
+    expect(() => validateGepaSeat(seat({ harness: 'claude-code' }))).toThrow(/'harness' belongs to harness-authored/)
     expect(() => validateGepaSeat(seat({ merge: true }))).toThrow(/'merge'/)
     expect(() => validateGepaSeat(seat({ model: 'x' }))).toThrow(/'model'/)
     expect(() => validateGepaSeat(seat({ profile: 'p.json' }))).toThrow(/'profile'/)
@@ -234,7 +234,7 @@ describe('captureProposerProvenance with a gepa seat', () => {
     return { code: 0, stdout: 'source', stderr: '' }
   }
   it('records engine, surface, gepa version, bridge module, and the python runtime as harnessVersion', async () => {
-    const record = await captureProposerProvenance([{ name: 'claude-author', harness: 'claude' }, seat()], {
+    const record = await captureProposerProvenance([{ name: 'claude-author', harness: 'claude-code' }, seat()], {
       exec: okExec,
       readSettingsModel: () => 'settings-model',
     })
@@ -251,7 +251,7 @@ describe('captureProposerProvenance with a gepa seat', () => {
     expect(gepa.harness).toBeUndefined()
     // The claude seat is untouched by the gepa capture path.
     expect(record.proposers.find((p) => p.name === 'claude-author')).toMatchObject({
-      harness: 'claude',
+      harness: 'claude-code',
       settingsModel: 'settings-model',
     })
   })

bench/src/swe-arena/outer-loop.mts

@@ -646,7 +646,7 @@ export interface OuterLoopConfig {
    *  budget.maxImprovementShots; the LIB owns the dial (capabilities.mts
    *  fails loud on a substrate that would drop it). */
   maxShots: number
-  proposerHarness: 'claude' | 'codex' | 'opencode'
+  proposerHarness: 'claude-code' | 'codex' | 'opencode'
   proposerTimeoutMs: number
   /** GEN-3 proposer fan-out: N proposers author candidates CONCURRENTLY, each
    *  an AgentProfile-pinned harness invocation (see proposer-fanout.mts).
@@ -759,7 +759,7 @@ export function defaultRound4Config(
     // bootstrap run writes it; the lib validates it on every consumption.
     premeasuredBaselinePath: join(hh, 'r4', 'premeasured-baseline.json'),
     maxShots: 3,
-    proposerHarness: 'claude',
+    proposerHarness: 'claude-code',
     // Per author SHOT (agenticGenerator timeoutMs). 20 min timed out 3× under
     // degraded capacity in gen-1 ("author shot timed out") — doubled to 40 min.
     proposerTimeoutMs: 2_400_000,
@@ -858,18 +858,18 @@ export function defaultGen3Config(
   const base = defaultRound4Config(hh, opts)
   const outDirName = opts.outDirName ?? 'gen3'
   const proposers: ProposerSpec[] = [
-    { name: 'default-author', profile: 'default-author.profile.json', harness: 'claude' },
+    { name: 'default-author', profile: 'default-author.profile.json', harness: 'claude-code' },
     {
       name: 'mechanics-author',
       profile: 'default-author.profile.json',
-      harness: 'claude',
+      harness: 'claude-code',
       diagnosisSlice: 'mechanics',
       lens: 'Focus on MECHANICS: worker lifecycle, sandbox/clone contracts, settlement and delivery paths. Prefer code-path fixes over prompt wording.',
     },
     {
       name: 'prompts-author',
       profile: 'default-author.profile.json',
-      harness: 'claude',
+      harness: 'claude-code',
       diagnosisSlice: 'prompts',
       lens: 'Focus on PROMPTS: worker/brain instruction wording, placement guidance, self-check discipline. Prefer prompt/instruction changes over code-path rewrites.',
     },
@@ -945,10 +945,10 @@ export function defaultGen4Config(
 ): OuterLoopConfig {
   const base = defaultGen3Config(hh, { outDirName: opts.outDirName ?? 'gen4' })
   const proposers: ProposerSpec[] = [
-    { name: 'claude-author', profile: 'default-author.profile.json', harness: 'claude' },
+    { name: 'claude-author', profile: 'default-author.profile.json', harness: 'claude-code' },
     { name: 'glm-author', harness: 'opencode', model: 'zai-coding-plan/glm-5.2' },
     ...(opts.includeCodex === false ? [] : [{ name: 'codex-author', harness: 'codex' } satisfies ProposerSpec]),
-    { name: 'merge-author', profile: 'default-author.profile.json', harness: 'claude', merge: true },
+    { name: 'merge-author', profile: 'default-author.profile.json', harness: 'claude-code', merge: true },
   ]
   return {
     ...base,
@@ -1263,7 +1263,7 @@ const CODEX_AMBIENT_AUTH_VARS = ['OPENAI_API_KEY', 'OPENAI_BASE_URL'] as const
 
 export function proposerShotEnv(harness: OuterLoopConfig['proposerHarness']): NodeJS.ProcessEnv {
   const env: NodeJS.ProcessEnv = { ...process.env }
-  if (harness === 'claude') {
+  if (harness === 'claude-code') {
     for (const name of CLAUDE_AMBIENT_AUTH_VARS) delete env[name]
   }
   if (harness === 'codex') {

bench/src/swe-arena/proposer-fanout.mts

@@ -91,7 +91,7 @@ export interface ProposerSpec {
   profile?: string
   /** Required for harness-authored seats. Absent on an engine seat
    *  (`engine` set) — enforced both ways at generator construction. */
-  harness?: 'claude' | 'codex' | 'opencode'
+  harness?: 'claude-code' | 'codex' | 'opencode'
   /** GEN-4 pinned model id, threaded to the harness CLI as `-m <model>` via
    *  the author profile's `model.default` (harnessInvocation maps it for all
    *  three harnesses). Unset = the CLI's own resolved model (its login/settings
@@ -201,7 +201,7 @@ export function loadAuthorProfile(spec: ProposerSpec): AgentProfile | undefined
 
 /** The gen-2 author, codified: one bare-profile claude proposer. */
 export function defaultProposers(): ProposerSpec[] {
-  return [{ name: 'default-author', profile: 'default-author.profile.json', harness: 'claude' }]
+  return [{ name: 'default-author', profile: 'default-author.profile.json', harness: 'claude-code' }]
 }
 
 /** The profile the author shot actually runs: the loaded profile (if any) with