feat(multi_review): per-agent prompt template loader with fallback chain

[samestrin]

Summary

multi_review today sends ONE task message to every reviewer. The openclaw pool has six agents with distinct specializations (per their SOULs on nucleus) AND distinct failure modes (kai over-explores, otto under-reports, etc.) that warrant model-specific prompt overrides.

This PR externalizes the task message to disk-based markdown templates so per-model prompts can be tuned without rebuilding the binary. The templates themselves live in claude-prompts/.planning/.templates/multi-review-prompts/ (companion PR) and are synced to ~/.llm-tools/multi-review/prompts/ by update-prompts.sh.

Fallback chain

Per agent invocation, in resolution order:

1. --task-message CLI flag — explicit user override, wins for ALL agents (already existed; unchanged)
2. <dir>/<agent>.md — per-agent override (e.g. kai.md)
3. <dir>/_base.md — universal base (mirrors today's buildDefaultTaskMessage content as a template)
4. Hardcoded buildDefaultTaskMessage — today's behavior, last resort

If the prompts dir doesn't exist OR is empty, every agent falls through to the hardcoded default. Fresh installs that haven't run update-prompts.sh see no behavior change.

Architecture

internal/support/multireview/prompt_loader.go    # NEW
  func ResolvePromptDir() string                 # env var override or $HOME/...
  func LoadAgentPrompt(dir, agent, vars) (string, error)
  type PromptVars { DiffPath, DiffBytes, DiffLines, DiffMB, LargeDiff,
                    BaseRef, HeadRef, RemoteRepo, AgentName }

internal/support/commands/multi_review.go        # modified
  runMultiReview:
    promptDir := multireview.ResolvePromptDir()
    promptVarsBase := multireview.PromptVars{...}  # build once
    invokeOne(agent):
      msg := resolveTaskMessage(agent)           # apply fallback chain
      invokeReviewerFn(... TaskMessage: msg ...)

Resilience

• Missing dir, missing files → silently fall through to next layer
• Template parse error in any file → silently fall through (a single broken per-agent file does NOT crash the run)
• Template execution error (missing field, bad pipeline) → silently fall through
• Path-traversal in agent name → filepath.Base() collapses ../etc/passwd to passwd; only files inside the prompts dir are searched

Test plan

• go test ./internal/support/multireview/... — 12 new tests for the loader, 83.8% coverage on the package
• go test ./internal/support/commands/... — 4 new integration tests, 80.8% coverage on commands
• No real openclaw / SSH / file system outside t.TempDir() — all tests use the existing execCommandContext swap pattern + t.Setenv(LLM_TOOLS_MULTI_REVIEW_PROMPTS, ...) for filesystem isolation
• Concurrent map-write bug caught by the race detector during dev (fixed with mutex in the test mock)
• Companion PR in claude-prompts populates the template files and adds the sync step to update-prompts.sh (separate PR; merges after this one lands)

Adversarial review

Scenario	Verdict
Missing _base.md	Falls back to hardcoded. Tested.
Missing <agent>.md, has _base.md	Returns _base.md rendered. Tested.
Missing dir entirely	Returns "". Tested.
--task-message CLI flag set	Wins over all per-agent files. Tested.
Broken template in kai.md	Falls back to _base.md. Tested.
Broken template in BOTH	Returns "", caller uses hardcoded. Tested.
Backticks in template body	Pass through unchanged (text/template treats them as literal). Tested.
Path-traversal via agent name	filepath.Base sanitizes. Tested.
Concurrent parallel reviewers	No shared mutable state; loader is stateless.

Commits

SHA	TDD step
1b80326	feat(multireview): LoadAgentPrompt with fallback chain (RED + GREEN)
1eca92a	feat(multi_review): wire LoadAgentPrompt into runMultiReview
aa8a30f	docs(changelog): per-agent prompt template loader under Unreleased

Rollback path

Zero risk: every layer of the fallback chain returns to today's behavior. Revert at any commit is safe. Even if the loader is shipped without the companion claude-prompts PR (no template files on disk), the binary falls back to the hardcoded default.