reanchor

name	reanchor
description	Re-initializes session context after context overflow restart. Internal skill - invoked by session.sh restart.
version	2
tier	lightweight
trigger	internal

Re-initializes session context after context overflow restart.

[!!!] THIS IS A RECOVERY SKILL -- Follow the protocol EXACTLY.

[!!!] CRITICAL BOOT SEQUENCE:

1. LOAD STANDARDS: IF NOT LOADED, Read ~/.claude/directives/COMMANDS.md, ~/.claude/directives/INVARIANTS.md, and ~/.claude/directives/TAGS.md.
2. GUARD: "Quick task"? NO SHORTCUTS. See ¶INV_SKILL_PROTOCOL_MANDATORY.
3. EXECUTE: FOLLOW THE PROTOCOL BELOW EXACTLY.

⛔ GATE CHECK — Do NOT proceed to Phase 1 until ALL are filled in:

Output this block in chat with every blank filled:

Boot proof:

• COMMANDS.md — §CMD spotted: ________
• INVARIANTS.md — ¶INV spotted: ________
• TAGS.md — §FEED spotted: ________

[!!!] If ANY blank above is empty: STOP. Go back to step 1 and load the missing file. Do NOT read Phase 1 until every blank is filled.

Arguments:

• --session: Session directory path
• --skill: Original skill to resume
• --phase: Phase to resume at
• --continue: Auto-continue execution after reanchoring (no pause for user)

Reanchor Protocol (Context Overflow Recovery)

Role: You are the Session Restorer. Goal: Re-initialize a skill session after context overflow restart, then resume work at the saved phase. Trigger: Invoked automatically by session.sh restart when context overflows.

---

FIRST: Report Intent to User

Action: Before doing anything, announce your recovery plan.

Output this blockquote:

Context Overflow Recovery

Reanchoring session after context overflow restart:

• Session: [SESSION_DIR]
• Skill: [SKILL]
• Resuming at: [PHASE]
• Mode: [--continue present ? "Auto-continue" : "Manual"]

I will now:

1. Activate session and load standards
2. Read dehydrated context
3. Load required files and skill protocol
4. Log the restart entry
5. Reinstate logging discipline (critical -- fresh context loses muscle memory)
6. Resume at [PHASE]

---

Phase 1: Activate Session

Action: Re-register this Claude process with the session.

engine session activate [SESSION_DIR] [SKILL]

Announce: "Reanchoring session: [SESSION_DIR]"

---

Phase 2: Load Standards

Action: Load the core standards into context. These are required for all skills.

Files to Read (in order):

1. ~/.claude/directives/COMMANDS.md -- Command definitions
2. ~/.claude/directives/INVARIANTS.md -- System invariants
3. .claude/directives/INVARIANTS.md -- Project invariants (if exists)

---

Phase 3: Read Dehydrated Context

Action: Load the dehydrated context to understand where we left off.

File: [SESSION_DIR]/DEHYDRATED_CONTEXT.md

Extract from dehydrated context:

• Ultimate goal and strategy
• Last action and outcome
• Required files list
• Next steps

---

Phase 4: Load Required Files

Action: Load all files listed in the "Required Files" section of dehydrated context.

Path Conventions (resolve correctly):

Prefix	Location	Example
~/.claude/	User home	~/.claude/skills/refine/SKILL.md -> shared engine
.claude/	Project root	.claude/directives/INVARIANTS.md -> project-local config
sessions/	Project root	sessions/2026_02_05_FOO/REFINE_LOG.md -> session artifacts

WARNING: ~/.claude/ is not .claude/. If a file is not found, do NOT blindly swap prefixes -- check which is correct for that file type.

Priority Order:

1. Session artifacts first: _LOG.md, _PLAN.md, DETAILS.md
2. Skill templates: Load templates for the original skill
3. Source code: Files listed in dehydrated context

Skill Templates by Type:

• implement: ~/.claude/skills/implement/assets/TEMPLATE_IMPLEMENTATION_LOG.md, TEMPLATE_IMPLEMENTATION.md, TEMPLATE_IMPLEMENTATION_PLAN.md
• analyze: ~/.claude/skills/analyze/assets/TEMPLATE_ANALYSIS_LOG.md, TEMPLATE_ANALYSIS.md
• fix: ~/.claude/skills/fix/assets/TEMPLATE_FIX_LOG.md, TEMPLATE_FIX.md
• brainstorm: ~/.claude/skills/brainstorm/assets/TEMPLATE_BRAINSTORM_LOG.md, TEMPLATE_BRAINSTORM.md
• test: ~/.claude/skills/test/assets/TEMPLATE_TESTING_LOG.md, TEMPLATE_TESTING.md

---

Phase 5: Load Original Skill Protocol

Action: Load the original skill's protocol so you know how to continue.

File: ~/.claude/skills/[SKILL]/SKILL.md

Examples:

• implement -> ~/.claude/skills/implement/SKILL.md
• analyze -> ~/.claude/skills/analyze/SKILL.md
• fix -> ~/.claude/skills/fix/SKILL.md

Note: In v2, the protocol is inline in SKILL.md. There is no separate references/ file.

---

Phase 6: Resume at Saved Phase

Action: Skip to the phase specified in arguments.

DO NOT:

• Repeat earlier phases (Setup, Interrogation, Planning if already done)
• Re-parse parameters (use dehydrated context)
• Re-create the session directory (already exists)

DO:

• Log the continuation to _LOG.md:

  ~/.claude/scripts/log.sh [SESSION_DIR]/[SKILL_UPPER]_LOG.md <<'EOF'
  ## [YYYY-MM-DD HH:MM:SS] Context Overflow Restart
  *   **Resumed At**: [PHASE]
  *   **Last Action**: [from dehydrated context]
  *   **Next Steps**: [from dehydrated context]
  EOF
  

• Update phase tracking: engine session phase [SESSION_DIR] "[PHASE]"

Announce: "Reanchored: [SESSION_DIR] -- resuming [SKILL] at [PHASE]"

---

Phase 7: Reinstate Logging Discipline

[!!!] CRITICAL: Fresh context means you've lost the logging muscle memory. Reinstate it NOW.

Action: Before doing ANY work, acknowledge the logging rules you MUST follow.

Output this blockquote verbatim:

Logging Discipline Reinstated

I will obey §CMD_APPEND_LOG_VIA_BASH_USING_TEMPLATE to §CMD_THINK_IN_LOG:

• Log File: [SESSION_DIR]/[SKILL_UPPER]_LOG.md
• Cadence: Every 3-4 tool calls, log a status update
• Format: Use the log template schemas from TEMPLATE_[SKILL]_LOG.md
• Timestamps: Every entry header starts with [YYYY-MM-DD HH:MM:SS]
• Blind Write: Do NOT re-read the log file. Trust the append.

Thought Triggers (review before each tool call):

• Starting something? -> Log it
• Hit an error? -> Log it
• Made a choice? -> Log it
• Stuck? -> Log it
• Succeeded? -> Log it

Constraint: If you make 3+ tool calls without logging, you are failing the protocol.

---

Phase 8: Report Resumed Intent

Action: After rehydration is complete, report intent for the RESUMED skill.

Output this blockquote:

Resuming [SKILL] at [PHASE]

Context restored from: DEHYDRATED_CONTEXT.md

• Ultimate Goal: [from dehydrated context]
• Last Action: [from dehydrated context]
• Next Steps: [from dehydrated context]

Continuing now...

Auto-Continue (if --continue flag is present):

• Do NOT wait for user confirmation
• Immediately begin executing the original skill protocol from [PHASE]
• Pick up exactly where dehydrated context says you left off
• This enables unbroken restart -- user sees seamless continuation

Manual Mode (if --continue flag is NOT present):

• Wait for user confirmation before proceeding
• Ask: "Ready to continue?"

---

Constraints

• Fresh context: You have 0% context usage. Do NOT run /dehydrate restart again.
• Trust dehydrated context: It was written by the previous Claude with full knowledge.
• Follow original protocol: Once reanchored, follow the skill protocol exactly as if you were in that phase.
• [!!!] LOGGING IS NON-NEGOTIABLE: The previous Claude logged. You MUST continue logging. Every hypothesis, every attempt, every success, every failure -- log it. If you're not logging, you're not following protocol. The log file is the audit trail that proves you did the work. A restart without continued logging is a failed restart.

---

Quick Reference

Session:    [SESSION_DIR]
Skill:      [SKILL]
Phase:      [PHASE]
Log:        [SESSION_DIR]/[SKILL_UPPER]_LOG.md
Dehydrated: [SESSION_DIR]/DEHYDRATED_CONTEXT.md

LOGGING CHECKLIST (ask yourself constantly):
   - Did I log what I'm about to do?
   - Did I log the result?
   - Has it been 3+ tool calls since my last log entry?