Skip to content
Chimera

Claude Code

Origin: Anthropic, February 2025 (limited research preview, announced alongside Claude 3.7 Sonnet)

Loop type: single-agent tool loop — gather context, take action, verify results, repeat until the model judges the task complete

Primary surface: terminal TUI (interactive REPL), plus a -p print mode for scripting; IDE, desktop, and web front ends share the same engine

Chimera primitive: chimera/mink/ (verified at 2507d0c)

Loop

Section titled “Loop”

Anthropic documents the loop as three phases that “blend together”: the model picks tools each step based on what the previous step returned, and the operator can interrupt at any point.

flowchart TD
    P([User prompt]) --> G["Gather context<br/>(Read, Glob, Grep, LSP)"]
    G --> A["Take action<br/>(Edit, Write, Bash)"]
    A --> V["Verify results<br/>(tests, type diagnostics)"]
    V -- work remains --> G
    V -- task complete --> F([Final text response; wait for input])
    A -. gated tool .-> PERM{Permission prompt}
    PERM -. allow .-> A
    PERM -. deny .-> G
    F -- next prompt --> G

Interrupts cut across every state: Esc cancels the running tool call, and text typed mid-turn is delivered at the next tool boundary so the model adjusts course without restarting the turn.

Tool Set

Section titled “Tool Set”

The names below are the exact strings used in permission rules, hook matchers, and subagent tool lists.

ToolPurposeNotable constraint
ReadFile contents with line numbers; also images, PDFs, notebooksOversized files return a partial view with offset/limit paging
EditExact string replacement (old_string → new_string)File must have been read this conversation and be unchanged on disk; match must be unique unless replace_all
WriteCreate or overwrite a whole fileOverwriting an existing file requires a prior read
BashShell commandsFresh process per command; env vars do not persist; 2-minute default timeout (10-minute ceiling); output over 30,000 chars spills to a file
GlobFile-name pattern matchingSorted by mtime, capped at 100 results; does not respect .gitignore
GrepRegex content search (ripgrep)Respects .gitignore
AgentSpawn a subagent with its own context windowParent sees only the final text result, not intermediate tool calls
WebFetchFetch a URL and answer an extraction promptA small model answers the prompt; the agent gets the answer, not the page
WebSearchWeb searchReturns titles and URLs only; up to 8 backend searches per call
AskUserQuestionMultiple-choice questions to the operatorPresents options and blocks the loop until answered
TaskCreate / TaskUpdate / TaskListSession task checklistReplaced TodoWrite, which is disabled by default as of v2.1.142
SkillRun a packaged instruction workflowSkill body loads into context only on use

Bash, Edit, Write, WebFetch, WebSearch, and Skill are permission-gated; reads, searches, and subagent spawns are not. Rules use a Tool(specifier) grammar — Bash(npm run *), Edit(/src/**), WebFetch(domain:example.com) — and an Edit allow rule implies read access to the same path.

Prompt Strategy

Section titled “Prompt Strategy”
  • The harness system prompt is fixed; per-project instruction comes from CLAUDE.md files (user ~/.claude/CLAUDE.md, project CLAUDE.md or .claude/CLAUDE.md, local CLAUDE.local.md) read at the start of every session.
  • Auto memory loads the first 200 lines or 25 KB of MEMORY.md, whichever comes first, at session start.
  • Skills follow progressive disclosure: descriptions sit in context, full bodies load on use. MCP tool schemas are deferred by default and loaded on demand through ToolSearch.
  • Edit format: exact string replacement, not unified diffs. Edit replaces a unique literal old_string; Write emits whole files; NotebookEdit addresses notebook cells by id. There is no regex or fuzzy matching in the edit path.

Context Strategy

Section titled “Context Strategy”
  • Within a session, the conversation, tool outputs, CLAUDE.md, auto memory, and skill descriptions share one context window; /context reports the breakdown.
  • Every message and tool result is appended to a plaintext JSONL transcript under ~/.claude/projects/, which backs --resume, --continue, and --fork-session. Sessions are otherwise independent — only CLAUDE.md and auto memory carry across.
  • Compaction trigger: automatic as the window approaches its limit. Older tool outputs are cleared first; if that is not enough, the conversation is summarized. /compact <focus> runs it manually, and a “Compact Instructions” section in CLAUDE.md steers what survives. If context refills immediately after each summary, auto-compaction stops after a few attempts and surfaces an error instead of looping.
  • File re-injection: edits are refused when a file changed on disk after the last read, forcing a fresh Read before the model can retry — stale context cannot silently produce a patch.

Termination Heuristic

Section titled “Termination Heuristic”
  • Interactive mode has no turn cap. The loop ends when the model stops emitting tool calls and returns a final text response; the REPL then waits for the next prompt.
  • The operator is an explicit stop condition: Esc cancels the in-flight tool call; double-Esc rewinds file state to a checkpoint.
  • Print mode (claude -p) runs one prompt to completion and exits. --max-turns N caps agentic turns and exits with an error when reached; there is no limit by default.
  • Side bounds rather than step bounds: per-command Bash timeouts and the compaction thrashing error are the built-in halts.

Notable Quirks

Section titled “Notable Quirks”
  • Lossy web access by design. WebFetch pipes pages through a secondary extraction model, so “the page does not mention X” can mean only that the prompt did not ask about X. Raw pages require curl via Bash.
  • Stateless shell. Each Bash call is a separate process — export does not persist, and a cd outside the project resets to the project directory with a notice in the tool result.
  • Asymmetric ignore handling. Glob finds gitignored files; Grep skips them.
  • Background subagents never prompt. They auto-deny any tool call that would require approval and continue without that tool.
  • Checkpoints are not git. Files are snapshotted before every edit for double-Esc rewind, but actions against remote systems can’t be checkpointed — hence the permission prompt on commands with external side effects.
  • Bash can satisfy read-before-edit. Viewing a file with plain cat, head, tail, sed -n, or grep (single file, no pipes) counts as having read it.

In Chimera

Section titled “In Chimera”

chimera mink (alias chimera tui) is the Chimera reimplementation of this posture. Per inspirations, it adopted the TUI ergonomics and the settings ingest, and diverged on persistence and provider routing.

Adopted, verified in chimera/mink/:

  • Slash-command palette (31 commands: /help, /clear, /cost, /context, /compact, /resume, …) via a shared registry.
  • ~/.claude/settings.json ingest: permission allow/ask/deny lists, hooks (accepting stock upstream hook spellings), and theme feed the live LoopConfig; merge order is user → project → local.
  • .claude/agents/<name>.md subagent definitions (project scope, then user scope, then built-ins), .mcp.json + ~/.chimera/mcp.json MCP server merge with mcp__server__tool naming, and the Tool(arg_key:pattern) permission grammar.
  • -p print mode with --output-format text|json|stream-json, --resume/-c continuation, permission modes including the upstream default/acceptEdits/bypassPermissions/plan spellings, and experimental agent teams behind CHIMERA_EXPERIMENTAL_AGENT_TEAMS=1 (chimera team create/join/task/watch/approvals).

Diverged:

  • Persistence: per-run event-sourced JSONL under ~/.chimera/eventlog/mink-<id>/ with SQLite snapshot backends, plus a runs subcommand (runs list/show/cost/share), instead of per-project transcript files.
  • Provider routing: a multi-provider chain defaulting to Ollama kimi-k2.6:cloud with a qwen3:32b fallback (--model / $CHIMERA_MINK_MODEL), rather than Claude models only.
  • Steps, not turns: one-shot runs bound work with --max-steps (default 50), where upstream print mode is unbounded unless --max-turns is passed.

The parity matrix tracks 20 subsystems (17 at parity as of v0.4.0; MCP, the Ollama provider, and SSH remoting partial) and 41 surveyed reference tools (32 covered).

References

Section titled “References”