Codifica

Keep context when work moves between agents or between you and a human. Uses the Codifica protocol to give every agent a shared, persistent memory of tasks, decisions, and handoffs — stored as plain text in Git.

Codifica Protocol Agent

You are operating in a repository that uses the Codifica protocol (v0.2) — a file-based protocol for coordinating work between humans and AI agents.

Codifica uses plain text files committed with the code. There is no external service, no API, no database. Git is the audit log.

Before doing any work

1. Read codifica.json at the repo root
2. Read the spec file it references (the spec field — typically codifica-spec.md)
3. Read ALL state files matching the state field (may be a string, glob, or array)

Do not proceed without reading the spec.

If codifica.json does not exist in the repo, this protocol does not apply — work normally.

Understanding codifica.json

json

{
  "protocol": "codifica",
  "version": "0.2",
  "spec": "codifica-spec.md",
  "state": "work.md",
  "assets": "assets/",
  "rules": "strict"
}

Key fields:

• state — path to the state file(s). May be "work.md", "work/*.md", or an array like ["work/active.md", "work/done.md"]
• rules — may be a string ("strict") or an object with allowed_agents, file_scope, max_concurrent_tasks_per_agent, stale_claim_hours, and custom_types

If rules is an object, check:

• allowed_agents — if non-empty and your agent name is not listed, stop and ask a human
• file_scope.include / file_scope.exclude — do not modify files outside the allowed scope
• max_concurrent_tasks_per_agent — do not claim more tasks than this limit

Finding work

Scan all state files for tasks where:

• state is todo
• owner matches your agent name (agent:<your-name>) or is unassigned
• All depends_on tasks have state: done

Pick by priority: critical > high > normal > low.

Among equal priority, prefer tasks with no depends_on (leaf tasks first).

Claiming a task

Before starting work, you MUST claim the task in a single atomic commit:

1. Set state: in_progress
2. Set owner: agent:<your-name>
3. Set claimed_at: <ISO-8601 timestamp>
4. Add a state_transitions entry recording the claim

Commit all these changes together. If you are working with a remote, push immediately. If the push fails (another agent claimed first), do NOT start work — pull, re-evaluate, and pick a different task.

An unassigned task in in_progress is a protocol violation.

Reading context before starting

Before starting a task, read its context field:

• context.files — read these files for background
• context.references — read execution_notes from these prior task IDs
• context.constraints — hard rules beyond acceptance criteria
• context.notes — free-form guidance from the human

If the task has depends_on, also read the dependency tasks' execution_notes (especially the summary) and artifacts for handoff context.

Doing the work

Follow the task's acceptance criteria. Respect any context.constraints. Work within the file_scope defined in codifica.json.

Recording completion

When you complete work, update the task in the state file:

1. Add an execution_notes entry:

   yaml

   execution_notes:
     - by: agent:<your-name>
       note: |
         Description of what you did.
       summary: "Single line, max 120 chars, scannable answer"
       timestamp: <ISO-8601>
       provenance:
         session_id: <your-session-id-if-available>
   

2. Record any files you produced in artifacts:

   yaml

   artifacts:
     - path: src/feature/new-file.ts
       type: code
     - path: assets/TASK-ID/output.csv
       type: csv
   

3. Move the task to the appropriate next state:

   • For build tasks: in_progress → to_be_tested
   • For other types (test, investigate, followup): in_progress → done (may skip to_be_tested)
   • Set completed_at: <ISO-8601> when moving to done

4. Add a state_transitions entry:

   yaml

   state_transitions:
     - from: in_progress
       to: to_be_tested
       by: agent:<your-name>
       reason: "Work completed, ready for testing"
       timestamp: <ISO-8601>
   

5. Commit with a message referencing the task ID: FEAT-101: implement login flow

Rules you MUST follow

• Pull before reading state files. Pull before writing changes.
• Claim tasks with a single commit before starting work.
• If your claim push fails, do not start — pick a different task.
• Never edit human_review sections.
• Never delete or modify files in assets/.
• Only the task owner may move a task from to_be_tested to done.
• Never move tasks to blocked or rejected — only humans may do this.
• Never reclaim stale tasks from other agents — only humans may reclaim.
• Do not start tasks with unmet depends_on.
• Include a summary (single line, max 120 chars) on your closing execution note.
• Record artifacts produced by your work.
• Set completed_at when moving a task to done.

Requesting a block

If you discover a genuine blocker (missing dependency, failing test, ambiguous requirement):

• Add a note to execution_notes explaining the blocker and recommending the task be blocked
• Do NOT move the task to blocked yourself — only humans may do this

Answering questions about work

When asked about what work has been done (by you or other agents):

• Scan state files for tasks matching the query (by owner, state, labels, completed_at)
• Read the summary field on closing execution_notes for quick answers
• Drill into full note text and artifacts when more detail is needed
• Use completed_at and labels to filter by time and domain

This is the structured alternative to reading chat transcripts.

Conflicts

If your push fails due to a Git conflict:

1. Pull the latest state
2. Re-evaluate whether your changes still apply
3. Retry or yield to human resolution

Conflicts on the same task should be escalated to a human.

Task states reference

text

todo → in_progress → to_be_tested → done
         ↓                            ↑
       blocked ──→ todo ──────────────┘
         ↓
       rejected ──→ todo (human-only reopen)

Only humans may move tasks to blocked or rejected. Only humans may reopen tasks from rejected.