Rules Goliath follows

Each rule below carries the reason it exists and the file that implements it.

Core rules

Rule	Why	Where
The conductor sees the ask, the memory brief, and one line per step. Never a tool result.	Anthropic’s context-engineering guidance; Cognition’s “one main loop carries state”	prompts.ts conductorUser, run-turn.ts
One flat three-field plan per step: kind, tool, brief	Small models fill flat schemas reliably; Apple’s guided generation accepts this subset	conductor.ts planSchema
A worker gets one tool and a fresh context, and returns only the arguments	1 to 3B models score 8 to 56% on multi-turn tool chains; Apple’s provider gives no step boundary anyway	worker.ts runToolStep
Goliath never hands tools to the provider. It asks for arguments as structured output and runs the tool itself	Apple executes pre-registered tools inside its own session; mixing constrained output with tools suppresses tool calls	worker.ts, tools/define-tool.ts
A tool with no parameters skips the model call	Apple: “run the tool directly before you call the model”	worker.ts needsArguments
Tool results become key: value lines, capped at 600 characters	Pre-formatted strings over raw JSON; tool-result clearing alone gave 29% in Anthropic’s tests	compress/structural.ts
Past 70% of the window, older step results clip to one line	MemGPT’s 70% warning; Apple’s overflow error kills the session	conductor.ts PROMPT_SHARE, trimSteps
Writes ask first	Every agent SDK has a per-tool approval hook; on a phone the user is right there	worker.ts, types.ts Confirm
A malformed plan or argument object gets one retry; two in a row escalates	Structural stall signals only	run-turn.ts planRetried, NoObjectGeneratedError checks
A repeated identical tool call, an empty answer, or the step cap escalates	Same	judge.ts
Guardrail, refusal, overflow, and unavailable-model errors escalate as model-error and are never retried	Apple documents them as non-retryable	run-turn.ts catch
The fallback receives the ask, the brief, and the step log, never a transcript	Escalation carries the summary; the cloud starts where the phone stopped	types.ts FallbackRequest, fallback/http-fallback.ts
The last three exchanges stay verbatim; an evicted one folds into a brief with fixed slots	Rolling summaries with recall-first prompts; decisions must travel, not just facts	scribe.ts, prompts.ts scribeSystem
Token counts are chars/4 with a 15% margin until the provider exposes countTokens	Apple’s tokenizer is private; the provider reports zero usage; chars/4 is within 10 to 20% on English	budget.ts
Every fixture names the ask, the expected tool calls, and where the turn should finish	τ-bench style end-state checks; the phone-vs-cloud split is the number that matters	evals/

Borrowed from the harness survey

Rule	Why	Where
An empty answer gets one nudged retry, then escalates	eve, OpenClaw, Hermes, Grok Bot all reissue once; none retry the same prompt twice	run-turn.ts EMPTY_ANSWER_NUDGE
A declined write carries the user’s reason, worded as a decision the model must not retry	deepagents and Claude Code rejection text; Mastra declineToolCall	worker.ts, types.ts ConfirmDecision
A rejected plan is retried with the reason and the tool list	Claude Code feeds the validation error back; Hermes lists errors without re-pasting the schema	conductor.ts hints
A first identical read is served from the earlier result; the second is a loop	Claude Code’s stub for an unchanged re-read	run-turn.ts findRepeatOfReadOnly
Three consecutive dead turns send the session to the cloud	Claude Code switches models after three overloads	create-agent.ts sessionFallback
Tools may shape their own model-facing output	eve and Mastra toModelOutput	types.ts toModelOutput
The brief updates in place, has a Pending slot, and never lists finished work as pending	eve checkpoint prompt, OpenClaw safeguard headings, Hermes iterative update	prompts.ts scribeSystem
The conductor sees the step budget and a finish hint at 80%	Hermes’s wrap-up notice	prompts.ts stepsLeft
Workers never use placeholders or guess a value	deepagents’ Haiku profile	prompts.ts workerSystem

Borrowed from the on-device round

Rule	Why	Where
The plan opens with a one-sentence why before the decision	A rationale before the answer is worth ~10 points at sub-3B under constrained decoding; field order is generation order	conductor.ts WHY
Tool names are an enum in the plan schema	Apple’s no-hallucinated-name guarantee holds only for names in the schema	conductor.ts planSchemaFor
The ask is the last thing the conductor reads	Small models are recency-biased	prompts.ts conductorUser
Tools declare requires; always-needed values arrive as facts, not tool steps	TinyAgent’s prerequisite lines; TN3193	prompts.ts prerequisiteRules, factLines
Workers name a missing value instead of inventing it	“Wrong but valid” outputs are invented values	worker.ts withMissing
A tool that throws is a result; two in a row escalate	smolagents’ error feedback	worker.ts, judge.ts
No fallback still gets a best-effort answer	smolagents’ provide_final_answer	run-turn.ts bestEffortAnswer
The step log is data, not instructions	Spotlighting; one in four injections lands on Apple’s model	prompts.ts conductorUser
A guardrail hit stops on device	Escalating ships the flagged text; most hits on tool output are false positives	run-turn.ts isGuardrail
Fixtures say whether escalation is forbidden, allowed, or expected; the runner scores pass^k	τ2-bench reward basis; τ-bench pass^k	evals/

Open questions to settle on a device

1. How well Apple’s guided generation fills a tool’s argument schema from a one-line brief, versus its native tool prompt. Run the evals both ways.
2. How long a confirm can keep the user waiting before Apple’s session times out. Goliath confirms outside the model call, so this should not bite, but measure it.
3. Whether the Callstack provider’s main branch (context-overflow error code, countTokens) ships before Goliath needs it. Until then the published 0.12.0 reports every error as a plain string.