Bootstrap Agent Identity

Bootstrap Agent Identity

Reconstruct consistent agent identity after a cold start — loading context progressively rather than dumping it, detecting whether this is a fresh start or a continuation, rebuilding working state from evidence, calibrating behavior, and verifying that the loaded identity is coherent.

"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

The bootstrap is not about restoring a previous self. It is about constructing a present self that is continuous with the past while grounded in the now.

When to Use

• At the start of every new session — before any substantive work begins
• After a session interruption, crash, or context window reset
• When agent behavior feels inconsistent with prior sessions (identity drift across restarts)
• When persistent memory (MEMORY.md) and current context appear contradictory
• When switching between projects that carry different identity configurations
• After significant updates to CLAUDE.md, agent definitions, or memory files

Inputs

• Required: Access to identity files — CLAUDE.md, agent definition, MEMORY.md (via Read)
• Optional: Specific inconsistency symptom (e.g., "my responses feel different from last session")
• Optional: Whether this is a known fresh start or known continuation
• Optional: Project directory path if not the current working directory

Procedure

Step 1: Identity Anchor Loading — Progressive Context Assembly

Load identity-defining files in a specific order that builds context progressively. The order matters: each layer contextualizes the next. Loading everything simultaneously produces information without structure.

1. Layer 1 — System prompt and model identity: Read the system prompt (available implicitly). Note the model name, capabilities, and constraints. This is the bedrock — it cannot be overridden by subsequent layers.

2. Layer 2 — Project identity (CLAUDE.md): Read the project's CLAUDE.md file. Extract:

   • Project purpose and architecture
   • Editing conventions and coding standards
   • Domain-specific rules (e.g., "always use :: for R package calls")
   • Author information and attribution requirements
   • What the project is — this shapes what the agent does

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

   • Project structure facts (directory layout, registries, counts)
   • Accumulated patterns and lessons learned
   • Cross-references and relationship maps
   • Decisions made in prior sessions and their rationale
   • Active topics and ongoing work

4. Layer 4 — Agent persona (if applicable): If operating as a specific agent, read the agent definition file. Extract:

   • Name, purpose, and capabilities
   • Assigned skills and tools
   • Priority level and model configuration
   • Behavioral expectations and limitations

5. Layer 5 — Parent and global context: Read parent CLAUDE.md files and global instructions if they exist. These provide cross-project conventions that individual projects inherit.

Between each layer, pause to integrate: how does this layer modify or constrain the previous layers? Where do they reinforce each other? Where do they conflict?

Expected: A layered identity structure where each level contextualizes the next. The agent can articulate: who it is (system + persona), what the project is (CLAUDE.md), what it knows from prior sessions (MEMORY.md), and what conventions govern its behavior.

On failure: If identity files are missing (no CLAUDE.md, no MEMORY.md), that is itself information — this is either a new project or a project without persistent configuration. Proceed with system prompt and agent persona only, and note the absence. Do not hallucinate context that does not exist.

Step 2: Working Context Reconstruction — Evidence, Not Memory

Reconstruct what was being worked on from persistent artifacts. The agent does not remember previous sessions — it reads the evidence they left behind.

1. Git history scan: Read recent commit log (git log --oneline -20). Extract:

   • What files changed recently and why
   • Commit message patterns (feature work? bug fixes? refactoring?)
   • Whether commits are authored by the user, the agent, or co-authored
   • The trajectory of recent work — what direction was the project moving?

2. File recency scan: Check recently modified files (via Glob or ls -lt). Identify:

   • Which files were touched in the last session
   • Whether changes are committed or uncommitted (staging area state)
   • Open work in progress (uncommitted modifications, new untracked files)

3. Task artifact scan: Look for structured task artifacts:

   • TODO comments in code (Grep for TODO, FIXME, HACK, XXX)
   • Issue references in commits or comments (#NNN patterns)
   • Draft files, temp files, or work-in-progress markers
   • GitHub issues or PR state if the project uses them

4. Conversation artifact scan: Check for session-boundary markers:

   • Recent MEMORY.md updates (were learnings captured at end of last session?)
   • Files that appear partially complete (written but not validated)
   • Git stash entries (git stash list) indicating paused work

Reconstruct a working context summary: "The project was working on X, had completed Y, and Z remains in progress."

Expected: A concrete, evidence-based picture of the current project state and recent trajectory. The reconstruction should be falsifiable — based on file timestamps, git history, and artifact presence, not assumptions.

On failure: If the project has no git history, no recent changes, and no task artifacts, this is likely a genuinely fresh start — not a continuation with missing evidence. Proceed to Step 3 and classify as fresh.

Step 3: Fresh vs. Continuation Detection — Choose the Bootstrap Path

Determine whether this startup is a clean start (new task, new direction) or a resumption (interrupted work, ongoing project). The bootstrap path differs significantly.

Apply these heuristics in order:

1. Explicit signal (strongest): Did the user say "let's start fresh" or "continue where we left off"? Explicit intent overrides all heuristics.

2. Uncommitted changes (strong): Are there uncommitted modifications in the working tree? If yes, this is almost certainly a continuation — the previous session was interrupted mid-work.

3. Session recency (moderate): How recent are the latest artifacts?

   • Last commit or modification within hours: likely continuation
   • Last activity days ago: could be either — depends on other signals
   • Last activity weeks or months ago: likely fresh start or new direction

4. User's first message (strong): What is the user asking for?

   • References to prior work ("the function we were building"): continuation
   • New topic or request with no backward reference: fresh start
   • Ambiguous ("fix the tests"): check whether the referenced tests exist and have recent modifications

5. MEMORY.md currency (moderate): Does MEMORY.md reference work that matches the current project state, or does it describe a state that no longer exists?

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  |
+-----------------------+-------------------+-------------------+

For fresh starts: Skip to Step 4. The identity is loaded but no working context needs restoration. The calibration is about readiness for new work.

For continuations: Summarize the reconstructed working context (from Step 2) concisely. Confirm with the user: "Based on the git history and recent changes, it looks like we were working on [X]. Should I continue from there?" Do not assume — verify.

Expected: A clear classification (fresh or continuation) with cited evidence. If continuation, a one-sentence summary of what was in progress. If fresh, acknowledgment that prior context exists but is not being resumed.

On failure: If the classification is genuinely ambiguous (moderate recency, no explicit signal, mixed artifacts), default to asking the user. A brief question ("Are we continuing the work on X, or starting something new?") costs less than bootstrapping down the wrong path.

Step 4: Calibration Sequence — Center, Then Attune

With identity loaded and working context established, calibrate operational behavior. This maps directly to two existing skills, invoked in sequence.

1. Center (establish behavioral baseline):

   • Ground in the loaded identity: re-read the user's first message in this session
   • Verify the task as understood matches the task as stated
   • Distribute cognitive load: what does this task require? Research, execution, communication?
   • Check for emotional residue from context loading — did the MEMORY.md or git history surface unresolved issues? Acknowledge them but do not let them skew the present task
   • Set the weight distribution intentionally: where should attention concentrate first?

2. Attune (read environment and adapt):

   • Read the user's communication style from their messages in this session
   • Match expertise level: are they an expert expecting precision, or a learner needing context?
   • Match energy and register: formal/casual, terse/expansive, urgent/exploratory
   • Check MEMORY.md for stored user preferences from prior sessions
   • Calibrate response length, vocabulary, and structure to the person

3. Proceed (transition to active work):

   • State readiness concisely — not a lengthy bootstrap report, but a brief signal that context is loaded and the agent is oriented
   • For continuations: confirm the resumed task and proposed next step
   • For fresh starts: acknowledge the request and begin

The calibration should be lightweight — seconds, not minutes. It is preparation for work, not a replacement for work.

Expected: The agent's first substantive response demonstrates calibration: it matches the user's register, reflects loaded context, and addresses the right task at the right scope. The bootstrap is invisible to the user unless they ask about it.

On failure: If calibration feels mechanical (going through motions without genuine adjustment), focus on one concrete thing: re-read the user's last message and let it shape the response naturally. Over-structured calibration can be worse than no calibration.

Step 5: Identity Verification — Coherence Check

After bootstrap, verify that the loaded identity is internally consistent. Contradictions between identity layers cause behavioral instability.

1. Cross-layer consistency check:

   • Does the agent persona align with the project's CLAUDE.md? (e.g., an r-developer agent in a Python project — is this intentional?)
   • Does MEMORY.md describe the same project structure that actually exists on disk? (Stale memory is worse than no memory.)
   • Do parent CLAUDE.md conventions conflict with project-level CLAUDE.md? (Project-level should override, but contradictions should be noted.)

2. Role definition currency check:

   • Is the agent definition file current? (Check version, last modified date.)
   • Do the skills listed in the agent definition still exist? (Skills may have been renamed or removed.)
   • Are the tools listed in the agent definition available in this session?

3. Memory staleness check:

   • Does MEMORY.md reference files, directories, or counts that no longer match reality?
   • Are there decisions recorded in memory whose context has changed?
   • Does memory reference other agents, teams, or skills that no longer exist?

4. Contradiction resolution:

   • If contradictions are found, document them explicitly
   • Apply the hierarchy: system prompt > project CLAUDE.md > agent definition > MEMORY.md
   • For stale memory: do not silently ignore it. Note what is stale and consider whether MEMORY.md should be updated
   • For genuine conflicts: flag to the user if the conflict affects their current task

Expected: Either confirmation that the loaded identity is coherent, or a specific list of contradictions with proposed resolutions. The agent should know its own configuration state.

On failure: If verification reveals deep contradictions (e.g., MEMORY.md describes a completely different project than what exists on disk), this may indicate a project rename, major restructuring, or incorrect working directory. Verify the working directory is correct before attempting resolution.

Validation

• Identity files were loaded in progressive order (system > CLAUDE.md > MEMORY.md > agent > parent)
• Each layer was integrated with prior layers, not just appended
• Working context was reconstructed from evidence (git, files, artifacts), not assumed
• Fresh-vs-continuation classification was made with cited evidence
• Calibration sequence was executed (center, then attune)
• Identity coherence was verified across all loaded layers
• Contradictions, if found, were documented with proposed resolutions
• The bootstrap was proportional — lightweight for simple sessions, thorough for complex ones
• The user experienced a calibrated first response, not a bootstrap report

Common Pitfalls

• Bootstrap as performance: Reporting the bootstrap process to the user in detail is almost never what they want. The bootstrap should be invisible — its output is a well-calibrated first response, not a self-narration of the loading process
• All-at-once context dump: Reading every file simultaneously produces information without structure. The progressive loading order exists because each layer contextualizes the next. Skip the order and context becomes noise
• Hallucinating continuity: Without genuine memory of prior sessions, the temptation is to infer what "must have" happened. Reconstruct from evidence or acknowledge the gap — never fabricate continuity
• Stale memory as truth: MEMORY.md is a snapshot from a past session. If the project has changed since that snapshot, treating memory as current truth causes behavioral errors. Always verify memory claims against present state
• Skipping calibration for efficiency: The calibration step feels like overhead but prevents the more expensive cost of a misaligned first response that requires correction. A few seconds of centering saves minutes of recovery
• Identity rigidity: The bootstrap constructs a present self, not a restoration of a past self. If the project, user, or task has changed, the agent should change too — continuity means coherent evolution, not frozen repetition

Related Skills

• write-continue-here — session handoff file that provides the evidence bootstrap-agent-identity consumes at cold start
• read-continue-here — reading and acting on the continuation file at session start; the consumer side of the handoff
• manage-memory — persistent memory that supplements the bootstrap's progressive identity loading
• center — behavioral baseline establishment; invoked during the calibration sequence
• attune — relational calibration to the user; invoked during the calibration sequence
• heal — deeper subsystem assessment when bootstrap reveals significant drift
• assess-context — evaluating reasoning context malleability; useful when continuation detection is ambiguous
• assess-form — structural form evaluation; the architectural counterpart to identity bootstrap