TT-Wang/forge

Fixes six real orchestration bugs surfaced while using forge to ship memem v0.10.0. Every fix traces to a concrete failure mode observed in a production run.

Validator CWD fixes

• handleValidate accepts optional cwd to redirect file checks, syntax checks, contract checks, and command execution. Workers in worktrees can now pass their worktreePath so validation actually sees their changes. Pre-v0.4.0 the validator resolved everything against a fixed server-startup CWD — every worker self-verify was silently checking main, not the worktree. This is the root cause of the worktree-clobber trust problem.
• Nonexistent cwd paths return a cwd_check failure with recommendation: ESCALATE instead of a confusing ENOENT cascade.

Iteration state per-run scoping

• State now stored at iterations/<runId>/<moduleId>.json when runId provided, legacy path as fallback.
• Previously attempts accumulated across every run forever — a brand-new m1 would see attempt: 21 and trigger ESCALATE because prior plans had used the same module ID.
• Security: runId regex-whitelisted to /^[\\w.-]{1,128}$/ to prevent path traversal.

Agent prompts

• planner.md: unconditional failure-pattern recall via "forge workflow failure" query in addition to task keywords. Flag file overlaps. Prefer one file per module.
• worker.md: explicitly forbidden from calling mcp__forge__validate. Bash self-checks from worktree root instead. Report worktreePath and runId.
• reviewer.md: two-mode — per-module (Phase 2b) and final release (Phase 4.5). Final mode has an 8-item checklist matching the bug categories we missed in memem v0.10.0: field-name consistency, default values, hook stdin double-drain, lazy state races, transient vs permanent errors, subprocess cold-start, ARG_MAX, unbounded injection.

Orchestrator (skills/forge/SKILL.md)

• Phase 0: working-tree cleanliness check with warning
• Phase 0b: unconditional memory_recall for framework failure patterns
• Phase 1b: "File overlap risk" and "Known risks from memory" sections in plan approval
• Phase 2: MANDATORY auto-WIP-commit between tiers; per-tier validate-in-main before spawning next tier; workers receive runId in prompt
• Phase 4: all validate / iteration_state calls pass runId
• Phase 4.5 (NEW, MANDATORY): final release review with full cumulative diff
• Phase 5: save success_pattern memory with run-shape metadata; squash WIP commits into release commit
• Lite mode: ≤4 modules or --lite flag skips worktree isolation entirely

Meta

This release was reviewed by its own new Phase 4.5 final reviewer, which caught 4 error-severity bugs (2 path traversal, 1 phase ordering, 1 schema drift) before shipping. The review system reviewed itself and found real bugs.

See CHANGELOG.md for the full breakdown.