continue

<Purpose> Reads .forge/ state and resumes from the most actionable point. The goal is zero re-explanation — the user should be able to type "forge continue" and immediately be back in context. </Purpose> <Use_When> - User returns after a session break - SessionStart detects existing .forge/state.json - User says "forge continue", "resume", "pick up where we left off" - KO: "이어서 해줘", "계속해줘", "포지 계속" JA: "続けて", "フォージ続行" ZH: "继续", "锻造继续" </Use_When> <Steps>

1. Detect state

Read .forge/state.json. If missing → "No active Forge project. Use forge to start one."

2. Load context (minimal per phase)

Phase	Files to load
0–1	state.json only
2	+ spec.md
3	+ spec.md, code-rules.md, design/, contracts/, runtime.json
4–6	+ spec.md, holes/, runtime.json
7	+ spec.md, holes/, delivery-report/

If .forge/runtime.json exists, always load it — it has lane ownership, blockers, and handoff notes.

3. Show continuation summary

Keep it short. The user needs three things:

Forge: {{project_name}}
Phase {{N}}/{{8 for build, 7 for repair}} — {{phase_name}}
{{one-line summary of where things stand}}

Next: {{what to do now}}
{{if blockers: list them}}

Examples:

Forge: my-saas-app
Phase 3/8 — develop
3 lanes active, 1 blocked (auth waiting on DB schema)

Next: Continue auth-api lane after resolving DB contract

Forge: my-saas-app
Phase 6/8 — fix
2 blockers remaining (login redirect loop, rate limit bypass)

Next: Diagnose login redirect — attempt 1/3

Do NOT dump every runtime field. Surface only what's actionable.

4. Determine continuation target

Evaluate in this exact order. Take the FIRST match — do not skip ahead.

Priority	Condition	Action
1	Customer blocker exists	Surface the question. Don't pretend work can continue without user input.
2	Internal blocker exists	Route to the owning team's phase skill with the blocker context.
3	Saved analysis is stale, or repair flow has no analysis yet	Route to forge:analyze first. Refresh analysis before phase work continues.
4	Lane has handoff notes and is in_progress	Load that lane's task file, worktree, and latest handoff note. Continue in-lane.
5	Runtime has a designated next lane (next_lane)	Load that lane's context. This is the lead dev's recommendation for where to pick up.
6	None of the above	Fall back to the current phase skill.

The goal: the user types forge continue and lands on the single most useful thing to work on, with full context loaded and zero re-explanation needed.

4a. AgentRecommendation routing

On resume, recommendedAgentsFor() provides layer-classified agent recommendations. Dispatch agents according to their layer classification (layer0, layer2_subagent, etc.).

Apply staleness-aware dispatch (from state-restore.mjs):

• Fresh (<1hr): full context available — dispatch per layer recommendation immediately
• Warm (1-24hr): abbreviated context — confirm direction with user before dispatching agents
• Stale (>24hr): minimal context — suggest forge cancel or require explicit forge continue confirmation before proceeding

5. Phase → skill mapping

Standard mode (build, N/8):

• 0 → forge:intake
• 1 → forge:discovery
• 2 → forge:design
• 3 → forge:develop
• 4 → forge:qa
• 5 → forge:security
• 6 → forge:fix
• 7 → forge:deliver

Repair mode (N/7):

• 0 → forge:intake
• 1 → forge:troubleshoot # focus on reproduction (guidance hint, not a sub-mode parameter)
• 2 → forge:troubleshoot # focus on isolation and root cause (guidance hint, not a sub-mode parameter)
• 3 → forge:fix
• 4 → forge:qa (regress)
• 5 → forge:security (verify)
• 6 → forge:deliver

Express mode:

• plan → forge:discovery (compressed)
• build → forge:develop (inline QA)
• ship → forge:deliver (compressed)

6. Hand off

Invoke the target skill with loaded context. The skill takes over from here.

</Steps> <Tool_Usage> - Read: .forge/state.json, .forge/runtime.json, phase-specific files per table above - Skill: invoke mapped phase skill </Tool_Usage>