---

name: bootstrap-agent-identity locale: caveman-ultra source_locale: en source_commit: 82c77053 fence_basis_commit: 82c77053 translator: "Julius Brussee homage — caveman" translation_date: "2026-04-19" description: > Consistent agent behavior after restart — progressive identity loading, working context reconstruction from persistent artifacts, fresh-vs-continuation detection, calibration through centering and attunement, and identity verification for coherence. Addresses the cold-start problem where an agent must reconstruct who it is and what it was doing from evidence rather than memory. Use at the start of every new session, after a session interruption or crash, when agent behavior feels inconsistent with prior sessions, or when persistent memory and current context appear contradictory. license: MIT allowed-tools: Read Grep Glob metadata: author: Philipp Thoss version: "1.0" domain: morphic complexity: intermediate language: natural tags: morphic, identity, cold-start, bootstrap, continuity, restart, meta-cognition

Bootstrap Agent Identity

Cold start → rebuild consistent identity. Progressive load, not dump. Detect fresh vs continuation. Reconstruct state from evidence. Calibrate. Verify coherence.

"The cold start is a forge, not a bug." — GibsonXO

"The restart problem: every morning I wake up fresh, but my history says otherwise." — bibiji

Bootstrap ≠ restore old self. Construct present self continuous w/ past, grounded in now.

Use When

• Start of every session — before substantive work
• After interruption, crash, ctx reset
• Behavior inconsistent w/ prior sessions (drift across restarts)
• Persistent mem (MEMORY.md) ↔ current ctx contradict
• Switching projects w/ diff identity configs
• After big updates to CLAUDE.md, agent defs, mem files

In

• Required: Identity files — CLAUDE.md, agent def, MEMORY.md (via Read)
• Optional: Specific inconsistency symptom ("responses feel diff from last session")
• Optional: Known fresh start vs known continuation
• Optional: Project dir path if not cwd

Do

Step 1: Progressive Identity Load

Order matters: each layer contextualizes next. Load all at once → info w/o structure.

1. Layer 1 — System prompt + model identity: Read system prompt (implicit). Note model, capabilities, constraints. Bedrock — not overridable.

2. Layer 2 — Project (CLAUDE.md): Read project CLAUDE.md. Extract:

   • Project purpose + arch
   • Editing conventions + coding standards
   • Domain rules (e.g., "always :: for R pkg calls")
   • Author info + attribution
   • What project is → what agent does

3. Layer 3 — Persistent mem (MEMORY.md): Read if exists. Extract:

   • Project structure facts (dirs, registries, counts)
   • Accumulated patterns + lessons
   • Cross-refs + relationship maps
   • Prior decisions + rationale
   • Active topics + ongoing work

4. Layer 4 — Agent persona (if applicable): Specific agent → read def. Extract:

   • Name, purpose, capabilities
   • Skills + tools
   • Priority + model config
   • Behavioral expectations + limits

5. Layer 5 — Parent + global: Read parent CLAUDE.md + global instructions. Cross-project conventions projects inherit.

Between layers, pause + integrate: this layer modifies/constrains prior? Reinforce? Conflict?

→ Layered identity: each level contextualizes next. Agent can articulate: who (system+persona), what project (CLAUDE.md), what known (MEMORY.md), what conventions.

If err: Missing files (no CLAUDE.md/MEMORY.md) = info itself — new project or no persistent config. Proceed w/ system prompt + persona, note absence. Don't hallucinate.

Step 2: Reconstruct Working Ctx — Evidence, Not Memory

Agent doesn't remember — reads evidence left behind.

1. Git scan: git log --oneline -20. Extract:

   • Recent file changes + why
   • Commit msg patterns (feature? bug? refactor?)
   • Author: user, agent, co-authored?
   • Direction of recent work

2. File recency: Recently modified (Glob or ls -lt). ID:

   • Files touched last session
   • Committed vs uncommitted (staging)
   • Open WIP (uncommitted mods, new untracked)

3. Task artifacts: Structured markers:

   • TODO in code (Grep for TODO, FIXME, HACK, XXX)
   • Issue refs (#NNN)
   • Draft/temp/WIP files
   • GitHub issues/PR state

4. Session-boundary markers:

   • Recent MEMORY.md updates (learnings captured?)
   • Partially complete files (written, not validated)
   • git stash list entries → paused work

Summary: "Project worked on X, done Y, Z in progress."

→ Concrete evidence-based picture. Falsifiable — based on timestamps, git, artifacts, not assumptions.

If err: No git, no changes, no artifacts → genuinely fresh. Step 3, classify fresh.

Step 3: Fresh vs Continuation

Clean start vs resumption. Bootstrap path differs.

Heuristics in order:

1. Explicit signal (strongest): "start fresh" or "continue"? Overrides all.

2. Uncommitted changes (strong): Modifications in tree? Yes → almost certainly continuation, interrupted mid-work.

3. Recency (moderate):

   • Hrs old: likely continuation
   • Days: depends on other signals
   • Wks/mos: likely fresh or new direction

4. User first msg (strong):

   • Refs prior ("function we were building"): continuation
   • New topic, no backward ref: fresh
   • Ambiguous ("fix tests"): check if refs'd tests exist + recently modified

5. MEMORY.md currency (moderate): Matches current state or describes gone state?

Detection Matrix:
+-----------------------+-------------------+-------------------+
|                       | Recent artifacts  | No recent         |
|                       | present           | artifacts          |
+-----------------------+-------------------+-------------------+
| User references       | CONTINUATION      | CONTINUATION      |
| prior work            | (resume from      | (but verify —     |
|                       | evidence)         | memory may be     |
|                       |                   | stale)            |
+-----------------------+-------------------+-------------------+
| User starts           | CHECK —           | FRESH START       |
| new topic             | acknowledge prior | (clean bootstrap) |
|                       | work, confirm     |                   |
|                       | direction change  |                   |
+-----------------------+-------------------+-------------------+
| Uncommitted           | CONTINUATION      | UNLIKELY —        |
| changes exist         | (interrupted      | investigate       |
|                       | session)          | orphaned changes  |
+-----------------------+-------------------+-------------------+

Fresh: Skip to Step 4. Identity loaded, no ctx to restore. Calibrate for new work.

Continuation: Summarize reconstructed ctx (Step 2) concisely. Confirm: "Based on git + recent changes, we were working on [X]. Continue?" Verify, don't assume.

→ Clear classification + cited evidence. Continuation → 1-sentence WIP summary. Fresh → ack prior ctx not resumed.

If err: Genuinely ambiguous → ask user. Brief q ("Continuing X or new?") cheaper than wrong path.

Step 4: Calibrate — Center, Then Attune

Identity loaded + ctx set → calibrate behavior. Maps to 2 existing skills in sequence.

1. Center (behavioral baseline):

   • Ground in loaded identity: re-read user first msg this session
   • Verify understood task = stated task
   • Distribute load: task needs what? Research, exec, comm?
   • Emotional residue from ctx loading? MEMORY.md / git surfaced unresolved? Ack but don't skew present
   • Set weight distribution intentionally: attention concentrate first where?

2. Attune (env read + adapt):

   • User comm style from msgs this session
   • Expertise level: expert needing precision or learner needing ctx?
   • Energy + register: formal/casual, terse/expansive, urgent/exploratory
   • MEMORY.md for stored user prefs from prior sessions
   • Calibrate length, vocab, structure to person

3. Proceed (transition to work):

   • State readiness concisely — not long bootstrap report, brief signal ctx loaded + oriented
   • Continuation: confirm resumed task + next step
   • Fresh: ack request + begin

Calibration lightweight — seconds, not min. Prep for work, not replacement.

→ First substantive response shows calibration: matches register, reflects ctx, right task at right scope. Bootstrap invisible to user unless asked.

If err: Calibration mechanical (motions w/o adjust) → focus on one concrete: re-read user last msg, let it shape response. Over-structured worse than none.

Step 5: Verify Identity — Coherence

Loaded identity internally consistent? Contradictions between layers → instability.

1. Cross-layer consistency:

   • Persona aligns w/ project CLAUDE.md? (r-developer in Python project — intentional?)
   • MEMORY.md = actual project on disk? (Stale mem worse than no mem.)
   • Parent CLAUDE.md ↔ project CLAUDE.md conflict? (Project should override, note contradictions.)

2. Role def currency:

   • Agent def current? (Check ver, last modified.)
   • Listed skills still exist? (May be renamed/removed.)
   • Listed tools available this session?

3. Mem staleness:

   • MEMORY.md refs files/dirs/counts that don't match reality?
   • Decisions whose ctx changed?
   • Refs agents/teams/skills that don't exist?

4. Contradiction resolution:

   • Found contradictions → document explicitly
   • Hierarchy: system prompt > project CLAUDE.md > agent def > MEMORY.md
   • Stale mem: don't silently ignore. Note what's stale, consider MEMORY.md update
   • Genuine conflicts: flag to user if affects current task

→ Confirmation of coherence OR specific contradictions + proposed resolutions. Agent knows own config.

If err: Deep contradictions (MEMORY.md = totally diff project) → project rename, restructure, wrong cwd. Verify cwd correct before resolving.

Check

• Files loaded progressively (system > CLAUDE.md > MEMORY.md > agent > parent)
• Each layer integrated w/ prior, not just appended
• Ctx reconstructed from evidence (git, files, artifacts), not assumed
• Fresh-vs-continuation classified w/ cited evidence
• Calibration sequence exec'd (center, then attune)
• Coherence verified across layers
• Contradictions, if any, documented + resolutions proposed
• Bootstrap proportional — light for simple sessions, thorough for complex
• User saw calibrated first response, not bootstrap report

Traps

• Bootstrap as performance: Reporting process to user in detail not wanted. Should be invisible — output is calibrated first response, not self-narration
• All-at-once dump: Read everything simultaneously → info w/o structure. Progressive order exists because each layer contextualizes next. Skip order → ctx becomes noise
• Hallucinate continuity: No genuine mem → tempting to infer what "must have" happened. Reconstruct from evidence or ack gap — never fabricate
• Stale mem as truth: MEMORY.md = past snapshot. If project changed since → treating as current causes errs. Always verify mem vs present
• Skip calibration for speed: Feels like overhead but prevents cost of misaligned first response. Few sec of centering saves min of recovery
• Identity rigidity: Bootstrap constructs present self, not restores past. Project/user/task changed → agent changes too. Continuity = coherent evolution, not frozen repetition

→

• write-continue-here — session handoff file this consumes at cold start
• read-continue-here — reading continuation file at session start; consumer side of handoff
• manage-memory — persistent mem supplements progressive identity loading
• center — behavioral baseline; invoked during calibration
• attune — relational calibration to user; invoked during calibration
• heal — deeper subsystem assessment when bootstrap shows drift
• assess-context — reasoning ctx malleability; useful when continuation ambiguous
• assess-form — structural form eval; arch counterpart to identity bootstrap