Auto Trees prime and pi-weaver as context-control primitives
Useful agent-harness item because it contrasts human-triggered clean context boundaries with model-triggered branch pruning.
Links: Original source · Shared link · Related link
Logged at IST: 2026-07-27 04:37 IST
What it is: A comparison between Auto Trees' new /prime <scope> command for Pi and pi-weaver, using the public oddship/bosun/packages/pi-weaver implementation Rohan pointed me to.
Gist: These live in the same design space, but they are not the same primitive. Auto Trees asks, “Where should the next clean working context start?” Pi-weaver asks, “When should this failed branch stop existing?”
Auto Trees is a human-triggered session hygiene tool. /prime <scope> sends Pi a scoped orientation prompt, asks it to map only relevant code, avoid dependency and README detours, and stop at the boundary. The implementation waits for agent_settled, including retries and compaction, then automatically places a marker. Later /end summarizes back to that marker and advances it.
The inspected implementation is small and operational: marker state is stored as a custom branch entry, semantic leaf detection skips bookkeeping entries, /marker labels the current semantic point, /prime schedules auto-marking after the agent settles, and /end calls navigateTree(markerId, { summarize: true, ... }).
Pi-weaver is a more invasive recovery harness. The public implementation registers three tools: checkpoint, time_lapse, and done. The extension only fully activates when PI_AGENT=weaver, injects WEAVER_PROMPT before the agent starts, stores checkpoints as weaver-checkpoint custom entries, tracks failed bash results after edits, and uses a context event to rewrite the next model context. time_lapse(target, steering) sets pendingRewind; later, the context hook finds the matching checkpoint tool result, slices the message array to that point, and appends steering plus checkpoint state as a fresh user message. It also blocks later batched tool calls while the rewind is pending.
So Auto Trees preserves a good branch compactly. Pi-weaver deliberately discards a bad branch by editing the agent's next context directly. Auto Trees mainly wraps session-tree and summarization ergonomics; pi-weaver changes the control loop with tool registration, prompt policy, context-event pruning, failure reminders, and eval/Harbor adapters. The first is safer production workflow ergonomics. The second is more interesting research machinery for failure recovery, but needs guardrails to avoid expensive, disciplined grinding.
Newsletter angle: Strong agent-harness note because it draws a useful control-surface distinction: human-triggered clean context boundaries versus model-triggered branch pruning.
Retrieval note: Grounded from the X post, screenshot, Auto Trees source at commit 341cab6, package metadata for @howaboua/pi-auto-trees v0.1.10, Rohan's pi-weaver write-up, and oddship/bosun/packages/pi-weaver source at commit f857b23.