PAPI: Persistent Adaptive Project Intelligence

Give a project a memory, so every later session starts with its context instead of a blank slate. Setup reads the project, writes a Product Brief that says what it is for and who it serves, records the first Active Decisions and build conventions so settled answers outlive the session that found them, and installs the workflow instructions the coding assistant follows from then on. Everything after this (planning, building, reviewing) reads from what setup writes. It takes a few minutes and runs in two passes: the first returns prompts for you to execute, the second saves the results. Tell the user what each pass is doing while it runs. Run after configuring your MCP credentials (via `init` or manually from getpapi.ai). Only project_name is required — description and target_users are derived from README, package.json, and commit history when omitted. Set existing_project: true to adopt an existing codebase. ADOPTING AN EXISTING PROJECT OVER A REMOTE/HOSTED CONNECTOR (no local stdio install): PAPI cannot read your filesystem, so YOU (the client) must gather a `codebase_scan` and pass it in — list top-level dirs/files, the package manifest, the README (first ~3000 chars), and recent commit subjects. Without it, adoption falls back to asking for description/target_users. On a local stdio install PAPI scans the tree itself, so `codebase_scan` is optional there. First call returns prompts (prepare phase), then call again with mode "apply" and your outputs. After setup, run `plan` to start your first cycle.

Queue topics for the next strategy review. Topics surface as input in the next `strategy_review` prepare phase and are automatically marked as addressed after the review completes. Two modes: "add" to queue a topic, "list" to see pending topics. Use this when you spot a strategic question during a build — capture the topic now instead of losing it.

Apply a strategic shift to the project. Three modes: "capture" for lightweight mid-conversation decision capture (no LLM round-trip), "prepare" to get a change prompt for full analysis, "apply" to persist analysis output. Use "capture" when you detect a strategic decision in conversation and want to persist it quickly without disrupting the build flow. In "capture" mode, pass north_star to directly set/update the project North Star (no decision text needed).

Run a Strategy Review — assesses project direction, velocity, and Active Decisions. Produces recommendations and potential AD updates that feed into the next plan. Offered every 5 cycles; hard-blocked at 7+ overdue cycles. Run it in your current conversation — only start a fresh one if you are genuinely under context pressure (your host just compacted, you are near the context limit, or the session is heavy with build context), not just because a review is next. First call returns a review prompt for you to execute (prepare phase). Then call again with mode "apply" and your output. Pass `force: true` to run before the cadence gate.

Claim a task from the shared org Pool into your personal backlog (assignee = you). Atomic first-claim-wins — a concurrent double-claim is impossible. Cascades the DEPENDS ON chain: claiming a task also claims its not-Done prerequisites; if any prerequisite is already claimed by another member the whole claim is refused (a build unit is never split across owners). Does not call the Anthropic API.

Move a task from the current project to another project you own. Reassigns the task a fresh id in the target (collision-free) and carries its build reports, comments, and history with it; the cycle assignment is cleared so it lands in the target project's backlog. You must own (or have write access to) BOTH projects. Destructive-ish and cross-project, so it requires confirm=true — without it you get a preview only. Does not call the Anthropic API.

Release a task you claimed back to the shared Pool (clears assignee). Claimer-only and pre-review — you cannot unclaim another member's task or one that has reached In Review/Done. Does not cascade. Does not call the Anthropic API.

Run a Zoom-Out Retrospective — a higher-level meta-retrospective that sits above strategy reviews. Analyses the full project arc: every cycle, decision, and pivot. Use when you want to step back and see the big picture after many cycles. First call returns a prompt (prepare phase). Then call again with mode "apply" and your output.