docs(prompts): extract all in-repo system prompts into docs/prompts

[qiaone]

Summary

This PR adds a new docs/prompts/ corpus that captures every LLM-facing system / instruction prompt living inside the Reasonix repository today, organised into 14 focused chapters with a top-level index.

The goal is to give contributors and reviewers a single place to:

• read the exact string the model sees, with zero paraphrasing;
• locate the source constant / file / line so any divergence can be fixed at the source;
• understand when and why each prompt is injected (compose path, retry path, subagent boundary, output-style overlay, ...);
• cross-reference the cache-stable prefix design that the rest of the codebase already revolves around.

Nothing executable is touched — this is documentation only.

What's inside docs/prompts/

File	Coverage
README.md	Index, reading order, source-code map
01-default-system-prompt.md	DefaultSystemPrompt
02-policies.md	UserDecisionPolicy, LanguagePolicy, ReasoningLanguageBlock
03-plan-mode.md	Plan-mode marker, auto-plan classifier, planApprovedMessage
04-goal-mode-and-autoresearch.md	activeGoalBlock, autoResearchGoalInstructions, goalContinueTurn, goalSelfCheckTurn, transient compose-time blocks
05-coordinator-handoff.md	DefaultPlannerPrompt, formatHandoff, executor-handoff retry
06-task-subagents.md	DefaultTaskSystemPrompt, DefaultReadOnlyTaskSystemPrompt, subagentToolBoundarySummary
07-builtin-skills.md	explore / research / review / security-review / test / init / install-capability
08-compaction.md	summarySystemPrompt and the three wrap-text variants
09-retry-and-recovery.md	streamRecoveryMessage, emptyFinalRetryMessage, finalReadinessRetryMessage, executorHandoffRetryMessage + the keyword tables that gate them
10-token-economy.md	tokenEconomyPrompt
11-server-and-utility.md	titlePrompt, prometheusPrompt, conductor template, e2e/benchmark prompts
12-custom-commands.md	.reasonix/commands/*.md user-level slash templates
13-output-styles.md	Built-in explanatory / learning / concise and KeepCoding overlay

Each chapter follows the same template: source file + symbol → role → injection trigger → verbatim original text → Chinese translation → design notes / cross-refs.

How this was produced

Every string was lifted directly from the Go source (internal/agent/agent.go, internal/config/config.go, internal/skill/builtins.go, internal/agent/task.go, internal/agent/compact.go, internal/control/*.go, internal/outputstyle/outputstyle.go, internal/serve/serve.go, ...) — never paraphrased. Where a prompt is assembled at runtime (e.g. the /plan-exec conductor template, formatHandoff), the doc shows the actual strings.Builder skeleton with the same \n\n separators.

Why this is worth landing

1. Reviewer ergonomics. Today, locating a particular instruction means grepping across internal/agent, internal/config, internal/skill, internal/control, internal/outputstyle, etc. The index makes the surface area auditable in one read.
2. Cache-prefix discipline. Several chapters explicitly call out the cache-stable layering (e.g. how policy blocks attach after the persona, why DefaultReadOnlyTaskSystemPrompt is the very first message of a read-only subagent transcript, how transient <memory-update> / Referenced context: blocks are stripped before they reach the persisted prefix). This complements REASONIX.md § Cache-impact PR metadata and scripts/check-cache-impact.sh.
3. Onboarding & external integrations. Anyone wanting to write a custom skill, a /command, or an MCP integration that respects the existing prompt contract can now read one document instead of reverse-engineering the agent loop.
4. Drift detection. Because every section names the source symbol, future prompt edits become trivially diff-able against the doc. (Happy to follow up with a small CI check that re-extracts the constants and diffs them against these chapters if maintainers want it.)

What this PR does not do

• No internal/** source file is modified.
• No prompt text is altered or re-worded — only documented.
• No new dependency, no new build target, no new CI job.
• No change to the model-visible byte stream, anywhere.

Verification

• git diff --name-only upstream/main-v2...HEAD — every entry is under docs/prompts/.
• bash scripts/check-cache-impact.sh against the changed file list prints No cache-sensitive prompt/tool files changed. and exits 0 (none of the documented chapters touch any path in the script's cache_sensitive allow-list).
• go build ./... and go test ./... are unaffected by this change set.

Cache metadata (per CONTRIBUTING.md § Cache-first review gate)

Even though the cache-impact script auto-passes for docs-only diffs, I'm filling these in explicitly so reviewers don't have to double-check:

Cache-impact: none — documentation-only change under docs/prompts/; no Go source, no system-prompt construction, no tool schema, no provider serialization, and no compaction code is modified. The model-visible byte stream is byte-identical before and after this PR.

Cache-guard: not applicable — scripts/check-cache-impact.sh exits 0 with No cache-sensitive prompt/tool files changed. for the file list in this PR; existing scripts/cache-guard.sh release-level coverage is unchanged.

System-prompt-review: none required — no file under internal/agent/task.go, internal/boot/, internal/config/, internal/memory/, internal/outputstyle/, or internal/skill/ is modified; the documentation only quotes existing constants verbatim.

Follow-ups I'm happy to do in separate PRs

• Add a tiny scripts/check-prompt-doc-drift.sh that re-greps the documented constants and fails CI if a quote drifts from the source.
• Translate the chapters into English-only mirrors if the maintainers prefer a single-language reference.

Thanks for considering — happy to split, restructure, or shrink the corpus per reviewer preference.