feat(intelligence): Observe + Mode-0 Intelligence SDK wrapper + effort/billing boundary by drewstone · Pull Request #277 · tangle-network/agent-runtime

drewstone · 2026-06-13T20:56:06Z

What

The 2-week go-live slice of docs/intelligence-sdk.md — the drop-in wrapper a product wraps its agent with, plus the intelligence-OFF billing boundary Drew asked for. Thin best-effort layer over the shipped createOtelExporter; no improvement-loop/Gated-PR modes (later phases).

src/intelligence/effort.ts — EffortPolicy as pure data. resolveEffort(tier, overrides?) compiles off | eco | standard | thorough | max → {analysts, corpus, fanout, loops, intelligenceBudgetUsd}. off ⇒ every intelligence axis off + budget 0; standard default; every field overridable.
src/intelligence/index.ts — createIntelligenceClient, withTangleIntelligence(agent, config), traceRun(meta, fn), doctor() (network-free mode readiness), flush().
src/intelligence/redact.ts — value-shaped PII/secret scrubber.
examples/intelligence-drop-in/ (~93 LOC) — wrap an agent, trace lands, survives a dead endpoint, OFF passthrough, doctor().
./intelligence subpath wired into exports + tsup + the packed-tarball import smoke.

The OFF billing boundary (Drew's ask)

effort: 'off' ⇒ the agent runs as pure passthrough (zero intelligence spawns) with best-effort telemetry still on (OFF and Observe are orthogonal — the upsell funnel stays lit). Usage is classified{inferenceUsd, intelligenceUsd} per trace (not a budget-pool split), and at OFF intelligenceUsd is clamped to 0 on export — so even a mis-reporting caller can't fabricate intelligence spend. Inference is still billed; OFF is inference-only, not free. The architecture: the billing line falls on the spawn line.

Honesty firewall

Aligned to the post-POWER-16 thesis (docs/go-live-plan.md): the SDK sells verifiably smarter and cheaper — 'smarter' only with its gate+CI+n attached. This slice makes no efficacy claim (Observe + passthrough only).

Verify

pnpm run build clean (dist/intelligence.{js,d.ts}), tsc --noEmit clean, pnpm run lint clean, full suite 837/837 + 21 new intelligence tests (independently re-run: 21/21). Example verified end-to-end against a real local OTLP collector.

…t/billing boundary The 2-week go-live slice of docs/intelligence-sdk.md: a thin best-effort layer over createOtelExporter that wraps a generic agent, ships one trace per call, and proves inference-only billing at the OFF tier. - src/intelligence/effort.ts — EffortPolicy as pure data. resolveEffort(tier, overrides) compiles 'off'|'eco'|'standard'|'thorough'|'max' into a flat {analysts,corpus,fanout,loops,intelligenceBudgetUsd} settings object; 'standard' is the default; every field individually overridable; 'off' ⇒ every intelligence knob off + intelligenceBudgetUsd:0. - src/intelligence/index.ts — createIntelligenceClient, withTangleIntelligence (passthrough wrapper, swallows export failures, propagates agent errors), traceRun(meta, fn) with recordOutput/recordOutcome, and doctor() (network-free mode readiness). Usage is classified per-trace as {inferenceUsd, intelligenceUsd}; at OFF intelligenceUsd is clamped to 0 — the mechanism that proves an OFF customer paid inference-only (a trace classification, not a budget-pool split, so budget.ts is untouched). - src/intelligence/redact.ts — value-shaped PII/secret scrubber (default + caller override + explicit opt-out), cycle-safe and total. - examples/intelligence-drop-in/ — <150 LOC drop-in over a local OTLP collector: trace lands, survives a dead endpoint, OFF passthrough. - ./intelligence subpath wired into package.json exports, tsup, and the package-export verifier. Behavior-changing modes (improvement loop, Gated PRs) are out of scope for this slice; the wrapper only Observes and passes through, so there is no abort path.

tangletools

✅ Auto-approved PR — `d690add1`

Blanket team auto-approval is enabled for this reviewer service.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.

_{tangletools · auto-approval · reason: blanket_auto_approve · 2026-06-13T20:56:13Z}

drewstone · 2026-06-13T20:56:06Z

What

The 2-week go-live slice of docs/intelligence-sdk.md — the drop-in wrapper a product wraps its agent with, plus the intelligence-OFF billing boundary Drew asked for. Thin best-effort layer over the shipped createOtelExporter; no improvement-loop/Gated-PR modes (later phases).

src/intelligence/effort.ts — EffortPolicy as pure data. resolveEffort(tier, overrides?) compiles off | eco | standard | thorough | max → {analysts, corpus, fanout, loops, intelligenceBudgetUsd}. off ⇒ every intelligence axis off + budget 0; standard default; every field overridable.
src/intelligence/index.ts — createIntelligenceClient, withTangleIntelligence(agent, config), traceRun(meta, fn), doctor() (network-free mode readiness), flush().
src/intelligence/redact.ts — value-shaped PII/secret scrubber.
examples/intelligence-drop-in/ (~93 LOC) — wrap an agent, trace lands, survives a dead endpoint, OFF passthrough, doctor().
./intelligence subpath wired into exports + tsup + the packed-tarball import smoke.

The OFF billing boundary (Drew's ask)

effort: 'off' ⇒ the agent runs as pure passthrough (zero intelligence spawns) with best-effort telemetry still on (OFF and Observe are orthogonal — the upsell funnel stays lit). Usage is classified{inferenceUsd, intelligenceUsd} per trace (not a budget-pool split), and at OFF intelligenceUsd is clamped to 0 on export — so even a mis-reporting caller can't fabricate intelligence spend. Inference is still billed; OFF is inference-only, not free. The architecture: the billing line falls on the spawn line.

Honesty firewall

Aligned to the post-POWER-16 thesis (docs/go-live-plan.md): the SDK sells verifiably smarter and cheaper — 'smarter' only with its gate+CI+n attached. This slice makes no efficacy claim (Observe + passthrough only).

Verify

pnpm run build clean (dist/intelligence.{js,d.ts}), tsc --noEmit clean, pnpm run lint clean, full suite 837/837 + 21 new intelligence tests (independently re-run: 21/21). Example verified end-to-end against a real local OTLP collector.

…t/billing boundary The 2-week go-live slice of docs/intelligence-sdk.md: a thin best-effort layer over createOtelExporter that wraps a generic agent, ships one trace per call, and proves inference-only billing at the OFF tier. - src/intelligence/effort.ts — EffortPolicy as pure data. resolveEffort(tier, overrides) compiles 'off'|'eco'|'standard'|'thorough'|'max' into a flat {analysts,corpus,fanout,loops,intelligenceBudgetUsd} settings object; 'standard' is the default; every field individually overridable; 'off' ⇒ every intelligence knob off + intelligenceBudgetUsd:0. - src/intelligence/index.ts — createIntelligenceClient, withTangleIntelligence (passthrough wrapper, swallows export failures, propagates agent errors), traceRun(meta, fn) with recordOutput/recordOutcome, and doctor() (network-free mode readiness). Usage is classified per-trace as {inferenceUsd, intelligenceUsd}; at OFF intelligenceUsd is clamped to 0 — the mechanism that proves an OFF customer paid inference-only (a trace classification, not a budget-pool split, so budget.ts is untouched). - src/intelligence/redact.ts — value-shaped PII/secret scrubber (default + caller override + explicit opt-out), cycle-safe and total. - examples/intelligence-drop-in/ — <150 LOC drop-in over a local OTLP collector: trace lands, survives a dead endpoint, OFF passthrough. - ./intelligence subpath wired into package.json exports, tsup, and the package-export verifier. Behavior-changing modes (improvement loop, Gated PRs) are out of scope for this slice; the wrapper only Observes and passes through, so there is no abort path.

tangletools

✅ Auto-approved PR — `d690add1`

Blanket team auto-approval is enabled for this reviewer service.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.

_{tangletools · auto-approval · reason: blanket_auto_approve · 2026-06-13T20:56:13Z}

drewstone · 2026-06-13T20:56:06Z

What

The 2-week go-live slice of docs/intelligence-sdk.md — the drop-in wrapper a product wraps its agent with, plus the intelligence-OFF billing boundary Drew asked for. Thin best-effort layer over the shipped createOtelExporter; no improvement-loop/Gated-PR modes (later phases).

src/intelligence/effort.ts — EffortPolicy as pure data. resolveEffort(tier, overrides?) compiles off | eco | standard | thorough | max → {analysts, corpus, fanout, loops, intelligenceBudgetUsd}. off ⇒ every intelligence axis off + budget 0; standard default; every field overridable.
src/intelligence/index.ts — createIntelligenceClient, withTangleIntelligence(agent, config), traceRun(meta, fn), doctor() (network-free mode readiness), flush().
src/intelligence/redact.ts — value-shaped PII/secret scrubber.
examples/intelligence-drop-in/ (~93 LOC) — wrap an agent, trace lands, survives a dead endpoint, OFF passthrough, doctor().
./intelligence subpath wired into exports + tsup + the packed-tarball import smoke.

The OFF billing boundary (Drew's ask)

effort: 'off' ⇒ the agent runs as pure passthrough (zero intelligence spawns) with best-effort telemetry still on (OFF and Observe are orthogonal — the upsell funnel stays lit). Usage is classified{inferenceUsd, intelligenceUsd} per trace (not a budget-pool split), and at OFF intelligenceUsd is clamped to 0 on export — so even a mis-reporting caller can't fabricate intelligence spend. Inference is still billed; OFF is inference-only, not free. The architecture: the billing line falls on the spawn line.

Honesty firewall

Aligned to the post-POWER-16 thesis (docs/go-live-plan.md): the SDK sells verifiably smarter and cheaper — 'smarter' only with its gate+CI+n attached. This slice makes no efficacy claim (Observe + passthrough only).

Verify

pnpm run build clean (dist/intelligence.{js,d.ts}), tsc --noEmit clean, pnpm run lint clean, full suite 837/837 + 21 new intelligence tests (independently re-run: 21/21). Example verified end-to-end against a real local OTLP collector.

…t/billing boundary The 2-week go-live slice of docs/intelligence-sdk.md: a thin best-effort layer over createOtelExporter that wraps a generic agent, ships one trace per call, and proves inference-only billing at the OFF tier. - src/intelligence/effort.ts — EffortPolicy as pure data. resolveEffort(tier, overrides) compiles 'off'|'eco'|'standard'|'thorough'|'max' into a flat {analysts,corpus,fanout,loops,intelligenceBudgetUsd} settings object; 'standard' is the default; every field individually overridable; 'off' ⇒ every intelligence knob off + intelligenceBudgetUsd:0. - src/intelligence/index.ts — createIntelligenceClient, withTangleIntelligence (passthrough wrapper, swallows export failures, propagates agent errors), traceRun(meta, fn) with recordOutput/recordOutcome, and doctor() (network-free mode readiness). Usage is classified per-trace as {inferenceUsd, intelligenceUsd}; at OFF intelligenceUsd is clamped to 0 — the mechanism that proves an OFF customer paid inference-only (a trace classification, not a budget-pool split, so budget.ts is untouched). - src/intelligence/redact.ts — value-shaped PII/secret scrubber (default + caller override + explicit opt-out), cycle-safe and total. - examples/intelligence-drop-in/ — <150 LOC drop-in over a local OTLP collector: trace lands, survives a dead endpoint, OFF passthrough. - ./intelligence subpath wired into package.json exports, tsup, and the package-export verifier. Behavior-changing modes (improvement loop, Gated PRs) are out of scope for this slice; the wrapper only Observes and passes through, so there is no abort path.

tangletools

✅ Auto-approved PR — `d690add1`

Blanket team auto-approval is enabled for this reviewer service.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.

_{tangletools · auto-approval · reason: blanket_auto_approve · 2026-06-13T20:56:13Z}

drewstone · 2026-06-13T20:56:06Z

What

The 2-week go-live slice of docs/intelligence-sdk.md — the drop-in wrapper a product wraps its agent with, plus the intelligence-OFF billing boundary Drew asked for. Thin best-effort layer over the shipped createOtelExporter; no improvement-loop/Gated-PR modes (later phases).

src/intelligence/effort.ts — EffortPolicy as pure data. resolveEffort(tier, overrides?) compiles off | eco | standard | thorough | max → {analysts, corpus, fanout, loops, intelligenceBudgetUsd}. off ⇒ every intelligence axis off + budget 0; standard default; every field overridable.
src/intelligence/index.ts — createIntelligenceClient, withTangleIntelligence(agent, config), traceRun(meta, fn), doctor() (network-free mode readiness), flush().
src/intelligence/redact.ts — value-shaped PII/secret scrubber.
examples/intelligence-drop-in/ (~93 LOC) — wrap an agent, trace lands, survives a dead endpoint, OFF passthrough, doctor().
./intelligence subpath wired into exports + tsup + the packed-tarball import smoke.

The OFF billing boundary (Drew's ask)

effort: 'off' ⇒ the agent runs as pure passthrough (zero intelligence spawns) with best-effort telemetry still on (OFF and Observe are orthogonal — the upsell funnel stays lit). Usage is classified{inferenceUsd, intelligenceUsd} per trace (not a budget-pool split), and at OFF intelligenceUsd is clamped to 0 on export — so even a mis-reporting caller can't fabricate intelligence spend. Inference is still billed; OFF is inference-only, not free. The architecture: the billing line falls on the spawn line.

Honesty firewall

Aligned to the post-POWER-16 thesis (docs/go-live-plan.md): the SDK sells verifiably smarter and cheaper — 'smarter' only with its gate+CI+n attached. This slice makes no efficacy claim (Observe + passthrough only).

Verify

pnpm run build clean (dist/intelligence.{js,d.ts}), tsc --noEmit clean, pnpm run lint clean, full suite 837/837 + 21 new intelligence tests (independently re-run: 21/21). Example verified end-to-end against a real local OTLP collector.

…t/billing boundary The 2-week go-live slice of docs/intelligence-sdk.md: a thin best-effort layer over createOtelExporter that wraps a generic agent, ships one trace per call, and proves inference-only billing at the OFF tier. - src/intelligence/effort.ts — EffortPolicy as pure data. resolveEffort(tier, overrides) compiles 'off'|'eco'|'standard'|'thorough'|'max' into a flat {analysts,corpus,fanout,loops,intelligenceBudgetUsd} settings object; 'standard' is the default; every field individually overridable; 'off' ⇒ every intelligence knob off + intelligenceBudgetUsd:0. - src/intelligence/index.ts — createIntelligenceClient, withTangleIntelligence (passthrough wrapper, swallows export failures, propagates agent errors), traceRun(meta, fn) with recordOutput/recordOutcome, and doctor() (network-free mode readiness). Usage is classified per-trace as {inferenceUsd, intelligenceUsd}; at OFF intelligenceUsd is clamped to 0 — the mechanism that proves an OFF customer paid inference-only (a trace classification, not a budget-pool split, so budget.ts is untouched). - src/intelligence/redact.ts — value-shaped PII/secret scrubber (default + caller override + explicit opt-out), cycle-safe and total. - examples/intelligence-drop-in/ — <150 LOC drop-in over a local OTLP collector: trace lands, survives a dead endpoint, OFF passthrough. - ./intelligence subpath wired into package.json exports, tsup, and the package-export verifier. Behavior-changing modes (improvement loop, Gated PRs) are out of scope for this slice; the wrapper only Observes and passes through, so there is no abort path.

tangletools

✅ Auto-approved PR — `d690add1`

Blanket team auto-approval is enabled for this reviewer service.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.

_{tangletools · auto-approval · reason: blanket_auto_approve · 2026-06-13T20:56:13Z}

drewstone · 2026-06-13T20:56:06Z

What

The 2-week go-live slice of docs/intelligence-sdk.md — the drop-in wrapper a product wraps its agent with, plus the intelligence-OFF billing boundary Drew asked for. Thin best-effort layer over the shipped createOtelExporter; no improvement-loop/Gated-PR modes (later phases).

src/intelligence/effort.ts — EffortPolicy as pure data. resolveEffort(tier, overrides?) compiles off | eco | standard | thorough | max → {analysts, corpus, fanout, loops, intelligenceBudgetUsd}. off ⇒ every intelligence axis off + budget 0; standard default; every field overridable.
src/intelligence/index.ts — createIntelligenceClient, withTangleIntelligence(agent, config), traceRun(meta, fn), doctor() (network-free mode readiness), flush().
src/intelligence/redact.ts — value-shaped PII/secret scrubber.
examples/intelligence-drop-in/ (~93 LOC) — wrap an agent, trace lands, survives a dead endpoint, OFF passthrough, doctor().
./intelligence subpath wired into exports + tsup + the packed-tarball import smoke.

The OFF billing boundary (Drew's ask)

effort: 'off' ⇒ the agent runs as pure passthrough (zero intelligence spawns) with best-effort telemetry still on (OFF and Observe are orthogonal — the upsell funnel stays lit). Usage is classified{inferenceUsd, intelligenceUsd} per trace (not a budget-pool split), and at OFF intelligenceUsd is clamped to 0 on export — so even a mis-reporting caller can't fabricate intelligence spend. Inference is still billed; OFF is inference-only, not free. The architecture: the billing line falls on the spawn line.

Honesty firewall

Aligned to the post-POWER-16 thesis (docs/go-live-plan.md): the SDK sells verifiably smarter and cheaper — 'smarter' only with its gate+CI+n attached. This slice makes no efficacy claim (Observe + passthrough only).

Verify

pnpm run build clean (dist/intelligence.{js,d.ts}), tsc --noEmit clean, pnpm run lint clean, full suite 837/837 + 21 new intelligence tests (independently re-run: 21/21). Example verified end-to-end against a real local OTLP collector.

…t/billing boundary The 2-week go-live slice of docs/intelligence-sdk.md: a thin best-effort layer over createOtelExporter that wraps a generic agent, ships one trace per call, and proves inference-only billing at the OFF tier. - src/intelligence/effort.ts — EffortPolicy as pure data. resolveEffort(tier, overrides) compiles 'off'|'eco'|'standard'|'thorough'|'max' into a flat {analysts,corpus,fanout,loops,intelligenceBudgetUsd} settings object; 'standard' is the default; every field individually overridable; 'off' ⇒ every intelligence knob off + intelligenceBudgetUsd:0. - src/intelligence/index.ts — createIntelligenceClient, withTangleIntelligence (passthrough wrapper, swallows export failures, propagates agent errors), traceRun(meta, fn) with recordOutput/recordOutcome, and doctor() (network-free mode readiness). Usage is classified per-trace as {inferenceUsd, intelligenceUsd}; at OFF intelligenceUsd is clamped to 0 — the mechanism that proves an OFF customer paid inference-only (a trace classification, not a budget-pool split, so budget.ts is untouched). - src/intelligence/redact.ts — value-shaped PII/secret scrubber (default + caller override + explicit opt-out), cycle-safe and total. - examples/intelligence-drop-in/ — <150 LOC drop-in over a local OTLP collector: trace lands, survives a dead endpoint, OFF passthrough. - ./intelligence subpath wired into package.json exports, tsup, and the package-export verifier. Behavior-changing modes (improvement loop, Gated PRs) are out of scope for this slice; the wrapper only Observes and passes through, so there is no abort path.

tangletools

✅ Auto-approved PR — `d690add1`

Blanket team auto-approval is enabled for this reviewer service.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.

_{tangletools · auto-approval · reason: blanket_auto_approve · 2026-06-13T20:56:13Z}

drewstone · 2026-06-13T20:56:06Z

What

The 2-week go-live slice of docs/intelligence-sdk.md — the drop-in wrapper a product wraps its agent with, plus the intelligence-OFF billing boundary Drew asked for. Thin best-effort layer over the shipped createOtelExporter; no improvement-loop/Gated-PR modes (later phases).

src/intelligence/effort.ts — EffortPolicy as pure data. resolveEffort(tier, overrides?) compiles off | eco | standard | thorough | max → {analysts, corpus, fanout, loops, intelligenceBudgetUsd}. off ⇒ every intelligence axis off + budget 0; standard default; every field overridable.
src/intelligence/index.ts — createIntelligenceClient, withTangleIntelligence(agent, config), traceRun(meta, fn), doctor() (network-free mode readiness), flush().
src/intelligence/redact.ts — value-shaped PII/secret scrubber.
examples/intelligence-drop-in/ (~93 LOC) — wrap an agent, trace lands, survives a dead endpoint, OFF passthrough, doctor().
./intelligence subpath wired into exports + tsup + the packed-tarball import smoke.

The OFF billing boundary (Drew's ask)

effort: 'off' ⇒ the agent runs as pure passthrough (zero intelligence spawns) with best-effort telemetry still on (OFF and Observe are orthogonal — the upsell funnel stays lit). Usage is classified{inferenceUsd, intelligenceUsd} per trace (not a budget-pool split), and at OFF intelligenceUsd is clamped to 0 on export — so even a mis-reporting caller can't fabricate intelligence spend. Inference is still billed; OFF is inference-only, not free. The architecture: the billing line falls on the spawn line.

Honesty firewall

Aligned to the post-POWER-16 thesis (docs/go-live-plan.md): the SDK sells verifiably smarter and cheaper — 'smarter' only with its gate+CI+n attached. This slice makes no efficacy claim (Observe + passthrough only).

Verify

pnpm run build clean (dist/intelligence.{js,d.ts}), tsc --noEmit clean, pnpm run lint clean, full suite 837/837 + 21 new intelligence tests (independently re-run: 21/21). Example verified end-to-end against a real local OTLP collector.

…t/billing boundary The 2-week go-live slice of docs/intelligence-sdk.md: a thin best-effort layer over createOtelExporter that wraps a generic agent, ships one trace per call, and proves inference-only billing at the OFF tier. - src/intelligence/effort.ts — EffortPolicy as pure data. resolveEffort(tier, overrides) compiles 'off'|'eco'|'standard'|'thorough'|'max' into a flat {analysts,corpus,fanout,loops,intelligenceBudgetUsd} settings object; 'standard' is the default; every field individually overridable; 'off' ⇒ every intelligence knob off + intelligenceBudgetUsd:0. - src/intelligence/index.ts — createIntelligenceClient, withTangleIntelligence (passthrough wrapper, swallows export failures, propagates agent errors), traceRun(meta, fn) with recordOutput/recordOutcome, and doctor() (network-free mode readiness). Usage is classified per-trace as {inferenceUsd, intelligenceUsd}; at OFF intelligenceUsd is clamped to 0 — the mechanism that proves an OFF customer paid inference-only (a trace classification, not a budget-pool split, so budget.ts is untouched). - src/intelligence/redact.ts — value-shaped PII/secret scrubber (default + caller override + explicit opt-out), cycle-safe and total. - examples/intelligence-drop-in/ — <150 LOC drop-in over a local OTLP collector: trace lands, survives a dead endpoint, OFF passthrough. - ./intelligence subpath wired into package.json exports, tsup, and the package-export verifier. Behavior-changing modes (improvement loop, Gated PRs) are out of scope for this slice; the wrapper only Observes and passes through, so there is no abort path.

tangletools

✅ Auto-approved PR — `d690add1`

Blanket team auto-approval is enabled for this reviewer service.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.

_{tangletools · auto-approval · reason: blanket_auto_approve · 2026-06-13T20:56:13Z}

drewstone · 2026-06-13T20:56:06Z

What

The 2-week go-live slice of docs/intelligence-sdk.md — the drop-in wrapper a product wraps its agent with, plus the intelligence-OFF billing boundary Drew asked for. Thin best-effort layer over the shipped createOtelExporter; no improvement-loop/Gated-PR modes (later phases).

src/intelligence/effort.ts — EffortPolicy as pure data. resolveEffort(tier, overrides?) compiles off | eco | standard | thorough | max → {analysts, corpus, fanout, loops, intelligenceBudgetUsd}. off ⇒ every intelligence axis off + budget 0; standard default; every field overridable.
src/intelligence/index.ts — createIntelligenceClient, withTangleIntelligence(agent, config), traceRun(meta, fn), doctor() (network-free mode readiness), flush().
src/intelligence/redact.ts — value-shaped PII/secret scrubber.
examples/intelligence-drop-in/ (~93 LOC) — wrap an agent, trace lands, survives a dead endpoint, OFF passthrough, doctor().
./intelligence subpath wired into exports + tsup + the packed-tarball import smoke.

The OFF billing boundary (Drew's ask)

effort: 'off' ⇒ the agent runs as pure passthrough (zero intelligence spawns) with best-effort telemetry still on (OFF and Observe are orthogonal — the upsell funnel stays lit). Usage is classified{inferenceUsd, intelligenceUsd} per trace (not a budget-pool split), and at OFF intelligenceUsd is clamped to 0 on export — so even a mis-reporting caller can't fabricate intelligence spend. Inference is still billed; OFF is inference-only, not free. The architecture: the billing line falls on the spawn line.

Honesty firewall

Aligned to the post-POWER-16 thesis (docs/go-live-plan.md): the SDK sells verifiably smarter and cheaper — 'smarter' only with its gate+CI+n attached. This slice makes no efficacy claim (Observe + passthrough only).

Verify

pnpm run build clean (dist/intelligence.{js,d.ts}), tsc --noEmit clean, pnpm run lint clean, full suite 837/837 + 21 new intelligence tests (independently re-run: 21/21). Example verified end-to-end against a real local OTLP collector.

…t/billing boundary The 2-week go-live slice of docs/intelligence-sdk.md: a thin best-effort layer over createOtelExporter that wraps a generic agent, ships one trace per call, and proves inference-only billing at the OFF tier. - src/intelligence/effort.ts — EffortPolicy as pure data. resolveEffort(tier, overrides) compiles 'off'|'eco'|'standard'|'thorough'|'max' into a flat {analysts,corpus,fanout,loops,intelligenceBudgetUsd} settings object; 'standard' is the default; every field individually overridable; 'off' ⇒ every intelligence knob off + intelligenceBudgetUsd:0. - src/intelligence/index.ts — createIntelligenceClient, withTangleIntelligence (passthrough wrapper, swallows export failures, propagates agent errors), traceRun(meta, fn) with recordOutput/recordOutcome, and doctor() (network-free mode readiness). Usage is classified per-trace as {inferenceUsd, intelligenceUsd}; at OFF intelligenceUsd is clamped to 0 — the mechanism that proves an OFF customer paid inference-only (a trace classification, not a budget-pool split, so budget.ts is untouched). - src/intelligence/redact.ts — value-shaped PII/secret scrubber (default + caller override + explicit opt-out), cycle-safe and total. - examples/intelligence-drop-in/ — <150 LOC drop-in over a local OTLP collector: trace lands, survives a dead endpoint, OFF passthrough. - ./intelligence subpath wired into package.json exports, tsup, and the package-export verifier. Behavior-changing modes (improvement loop, Gated PRs) are out of scope for this slice; the wrapper only Observes and passes through, so there is no abort path.

tangletools

✅ Auto-approved PR — `d690add1`

Blanket team auto-approval is enabled for this reviewer service.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.

_{tangletools · auto-approval · reason: blanket_auto_approve · 2026-06-13T20:56:13Z}

drewstone · 2026-06-13T20:56:06Z

What

The 2-week go-live slice of docs/intelligence-sdk.md — the drop-in wrapper a product wraps its agent with, plus the intelligence-OFF billing boundary Drew asked for. Thin best-effort layer over the shipped createOtelExporter; no improvement-loop/Gated-PR modes (later phases).

src/intelligence/effort.ts — EffortPolicy as pure data. resolveEffort(tier, overrides?) compiles off | eco | standard | thorough | max → {analysts, corpus, fanout, loops, intelligenceBudgetUsd}. off ⇒ every intelligence axis off + budget 0; standard default; every field overridable.
src/intelligence/index.ts — createIntelligenceClient, withTangleIntelligence(agent, config), traceRun(meta, fn), doctor() (network-free mode readiness), flush().
src/intelligence/redact.ts — value-shaped PII/secret scrubber.
examples/intelligence-drop-in/ (~93 LOC) — wrap an agent, trace lands, survives a dead endpoint, OFF passthrough, doctor().
./intelligence subpath wired into exports + tsup + the packed-tarball import smoke.

The OFF billing boundary (Drew's ask)

effort: 'off' ⇒ the agent runs as pure passthrough (zero intelligence spawns) with best-effort telemetry still on (OFF and Observe are orthogonal — the upsell funnel stays lit). Usage is classified{inferenceUsd, intelligenceUsd} per trace (not a budget-pool split), and at OFF intelligenceUsd is clamped to 0 on export — so even a mis-reporting caller can't fabricate intelligence spend. Inference is still billed; OFF is inference-only, not free. The architecture: the billing line falls on the spawn line.

Honesty firewall

Aligned to the post-POWER-16 thesis (docs/go-live-plan.md): the SDK sells verifiably smarter and cheaper — 'smarter' only with its gate+CI+n attached. This slice makes no efficacy claim (Observe + passthrough only).

Verify

pnpm run build clean (dist/intelligence.{js,d.ts}), tsc --noEmit clean, pnpm run lint clean, full suite 837/837 + 21 new intelligence tests (independently re-run: 21/21). Example verified end-to-end against a real local OTLP collector.

…t/billing boundary The 2-week go-live slice of docs/intelligence-sdk.md: a thin best-effort layer over createOtelExporter that wraps a generic agent, ships one trace per call, and proves inference-only billing at the OFF tier. - src/intelligence/effort.ts — EffortPolicy as pure data. resolveEffort(tier, overrides) compiles 'off'|'eco'|'standard'|'thorough'|'max' into a flat {analysts,corpus,fanout,loops,intelligenceBudgetUsd} settings object; 'standard' is the default; every field individually overridable; 'off' ⇒ every intelligence knob off + intelligenceBudgetUsd:0. - src/intelligence/index.ts — createIntelligenceClient, withTangleIntelligence (passthrough wrapper, swallows export failures, propagates agent errors), traceRun(meta, fn) with recordOutput/recordOutcome, and doctor() (network-free mode readiness). Usage is classified per-trace as {inferenceUsd, intelligenceUsd}; at OFF intelligenceUsd is clamped to 0 — the mechanism that proves an OFF customer paid inference-only (a trace classification, not a budget-pool split, so budget.ts is untouched). - src/intelligence/redact.ts — value-shaped PII/secret scrubber (default + caller override + explicit opt-out), cycle-safe and total. - examples/intelligence-drop-in/ — <150 LOC drop-in over a local OTLP collector: trace lands, survives a dead endpoint, OFF passthrough. - ./intelligence subpath wired into package.json exports, tsup, and the package-export verifier. Behavior-changing modes (improvement loop, Gated PRs) are out of scope for this slice; the wrapper only Observes and passes through, so there is no abort path.

tangletools

✅ Auto-approved PR — `d690add1`

Blanket team auto-approval is enabled for this reviewer service.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.

_{tangletools · auto-approval · reason: blanket_auto_approve · 2026-06-13T20:56:13Z}