Metadata Evolution

kiki may eventually help keep thread names, checkpoint bookmark names, and revision descriptions — jj's commit messages — current as work changes. Both the ownership foundation and the execution loop are v1.x work. The acceptance slice writes no AI-authored metadata, so it does not need to carry a ledger before that behavior exists. When the feature ships, kiki must know which prose it owns and which belongs to the human.

Ownership ledger

kiki tracks ownership in sqlite. It does not infer ownership by guessing from content, and it does not depend on jj trailers.

The ledger works like a sign-out sheet: kiki may only edit prose it can prove it wrote last. For every kiki-authored revision description or bookmark name, kiki records a content hash — a fingerprint of the text — in metadata_writes. On the next read:

• if the current content still matches the stored hash, kiki may update it;
• if the current content differs, the artifact becomes human-owned;
• once human-owned, kiki does not rewrite it unless the user explicitly opts back in.

This conservative rule applies to human edits and to agent-authored edits alike. From kiki's point of view, anything outside the ledger is user territory.

Revision descriptions

Auto-describe, when enabled, may run after:

• thread creation;
• an agent quiescence window — a quiet stretch with no agent tool calls — defaulting to 30 seconds;
• jj split;
• jj squash;
• explicit kk thread describe --refresh.

kiki stamps each job with an input hash. If the diff, ancestry, bookmark name, or relevant metadata changes while the model is running, kiki discards the result and queues a fresh job from the current state.

The auto-describe prompt may use:

• the revision diff;
• ancestor descriptions;
• the current bookmark name;
• kiki-owned metadata state.

It must not read the thread transcript.

Bookmark names

Auto-rename, when enabled, may propose slugs — short, dash-separated names — for kiki-owned thread bookmarks. (A jj bookmark is roughly a git branch name.) The prompt includes sibling bookmark names so the model can choose a distinct slug.

An external jj bookmark rename marks the bookmark human-owned for the rest of the thread's life. kiki does not rename it again unless an explicit future command opts it back in.

Squash and merge ownership

jj squash folds one revision's changes into another. When a squash combines a kiki-owned description with non-kiki content, the resulting description becomes human-owned. Ownership defaults toward preservation.

This rule is intentionally simple. If kiki cannot prove the resulting prose is still its own, it steps aside.

Trailers

Setting [autorename] trailer = true may add an opt-in Kk-Auto: true trailer — an extra labeled line at the end of a commit message — for transparency. The trailer is audit decoration only. The sqlite content-hash ledger still decides ownership.

Race control

Metadata writes require a per-thread advisory lock, a lock that cooperating writers agree to check before writing. Auto-describe and auto-rename wait for snapshot quiescence: no in-flight agent tool call touching files, and no in-progress jj operation.

If kiki cannot acquire the lock within 5 seconds, it abandons the attempt and re-queues it. kiki may spend a model call and still discard the result. That waste is better than writing stale prose.

Configuration

[autorename] controls the feature:

• enabled: true | false
• cadence: idle-only | events | always | off
• idle_ms: how long things must stay quiet before a job fires
• trailer: whether to write the transparency trailer (default false; opt-in)
• event toggles for creation, split, squash, and manual refresh

Cost management remains outside v1. kiki does not cap model spend, concurrent agents, CPU, RAM, or token usage.

Provider

Auto-describe and auto-rename call the model through AICompose, a provider-agnostic seam — an interface that hides which AI vendor is in use. [ai] selects the backend:

• provider: anthropic | openai | local-ollama | ...
• model: provider-specific model identifier (defaults to a Haiku-class model for Anthropic).
• api_key_env: env var holding the API key (default ANTHROPIC_API_KEY for Anthropic).
• api_key_path: file path holding the API key (alternative to api_key_env).

The first provider implementation is Anthropic. The trait surface is generic, so additional providers can be added without touching the metadata ledger or the prompt-assembly logic.

The auto-AI loop is opt-in. [autorename] enabled = false is the default, so a fresh install never calls a provider until the user configures one.