failproofai 1.0.0-beta.5 → 1.0.0-beta.7

package/.next/standalone/crates/fpai-collect/src/sources/pi/mod.rs

@@ -1,201 +0,0 @@
-//! Pi (`@mariozechner/pi-coding-agent`) session capture — a
-//! [`filetail`](crate::filetail) adapter.
-//!
-//! pi writes one JSONL transcript per session under
-//! `~/.pi/agent/sessions/<encoded-cwd>/<ISO-timestamp>_<uuid>.jsonl`. We open
-//! them read-only and turn each newly-appended line into AgentEye events; pi's
-//! own files are never written, moved or deleted.
-//!
-//! # The file does not exist until the first assistant message completes
-//!
-//! `SessionManager._persist()` short-circuits while the session has no
-//! assistant message in it, buffering entries in memory and marking itself
-//! unflushed; the first assistant response then writes the whole prefix at
-//! once. Measured on a live run: the file appeared at 1551 bytes, already
-//! containing the header, the model change, the thinking-level change, the user
-//! prompt and the assistant turn.
-//!
-//! Two consequences, both of which look like bugs if you do not know this:
-//!
-//! * **A run killed before its first assistant message leaves ZERO trace on
-//!   disk.** Discovery finding nothing in a directory pi is actively writing to
-//!   is normal, not an error, and must never be logged or handled as one.
-//! * The engine's `agent_start` retry loop is never exercised for pi in
-//!   practice — by the time a file exists, its header is complete — which is
-//!   also why [`Format::seed_state`](crate::filetail::Format::seed_state) is
-//!   left as a no-op here. See `transform_line`'s `model_change` arm: pi orders
-//!   the model ahead of the first prompt inside the file itself, so the opening
-//!   row is never blank and nothing has to be seeded out of band. That matters
-//!   because seeding is the one part of the format table that is NOT dedup-safe
-//!   between a live tail and a later full re-read.
-//!
-//! # `_rewriteFile()` rewrites the whole file, but only on cold paths
-//!
-//! Steady-state turns are `appendFileSync`, so a byte cursor is correct and
-//! [`RereadPolicy::ByteCursor`] is what this format declares. Two paths do a
-//! full `writeFileSync` over an existing transcript:
-//!
-//! * resuming a file whose entries all failed to parse — it is replaced by a
-//!   lone fresh header, a large **shrink**;
-//! * migrating a session written by pi with `version < 3`.
-//!
-//! The shrink case is covered: the engine notices `size < cursor.offset`,
-//! resets to zero and re-reads, which re-ships the surviving prefix for the
-//! server to dedup. That is the behaviour `a_rewritten_transcript_that_shrank_
-//! is_re_read_rather_than_seeked_past` pins. A migration that happens to leave
-//! the file the same size or larger would not be detected — accepted, because
-//! v3 has been current since well before either probed release and a
-//! `ValidatePrefix` policy would not help anyway (the migration rewrites
-//! interior records, not line 1).
-//!
-//! Forking and branching are NOT in that list, contrary to what the shape of
-//! the code suggests: both write a **new** file and leave the source transcript
-//! untouched, so they appear to this source as an ordinary new session whose
-//! header carries a `parentSession` pointer.
-//!
-//! # Scope
-//!
-//! Session start/end, user prompts, assistant turns, tool calls and their
-//! results. `compaction`, `branch_summary`, `label`, `session_info`, `custom`
-//! and `thinking_level_change` records are recognised as records — they keep
-//! the session clock moving — but emit no events; they need their own event
-//! types rather than being forced into these.
-
-pub mod transform;
-
-use std::path::{Path, PathBuf};
-
-use crate::filetail::{Format, RereadPolicy, no_seed_state};
-use serde_json::Value;
-
-/// Session transcripts: `<ISO-timestamp>_<uuid>.jsonl` under an encoded-cwd
-/// directory.
-pub const FORMAT: Format = Format {
-    kind: "pi",
-    is_source_file: is_transcript,
-    session_id_from_path,
-    agent_id_from_path,
-    agent_start: transform::agent_start,
-    // Deliberately nothing: pi's own record order already primes the model
-    // before the first prompt. See the module docs.
-    seed_state: no_seed_state,
-    agent_end: transform::agent_end,
-    transform_line: transform::transform_line,
-    reread: RereadPolicy::ByteCursor,
-};
-
-/// The agent id used when nothing can be derived.
-pub const DEFAULT_AGENT_ID: &str = "pi";
-
-/// Where pi keeps its sessions.
-///
-/// `PI_CODING_AGENT_DIR` relocates the whole agent directory (`~/.pi/agent`),
-/// and sessions hang off it — there is no separate root override. The tilde is
-/// expanded because pi expands it too: a value of `~/elsewhere` that we took
-/// literally would have us watching a directory named `~` under the process
-/// cwd, i.e. capturing nothing, silently.
-///
-/// Not exhaustive by construction. `PI_CODING_AGENT_SESSION_DIR`, `--session-dir`
-/// and a project's `settings.json` can each point ONE session at an arbitrary
-/// directory outside this root; those sessions are not discovered. Widening the
-/// walk to find them would mean walking the whole filesystem, so they are out of
-/// scope rather than half-handled.
-pub fn sessions_root() -> PathBuf {
-    let agent_dir = match std::env::var_os("PI_CODING_AGENT_DIR") {
-        Some(raw) => expand_tilde(PathBuf::from(raw)),
-        None => home().join(".pi").join("agent"),
-    };
-    agent_dir.join("sessions")
-}
-
-fn home() -> PathBuf {
-    std::env::var_os("HOME")
-        .map(PathBuf::from)
-        .unwrap_or_default()
-}
-
-fn expand_tilde(path: PathBuf) -> PathBuf {
-    let Some(s) = path.to_str() else {
-        return path;
-    };
-    match s {
-        "~" => home(),
-        _ => match s.strip_prefix("~/") {
-            Some(rest) => home().join(rest),
-            None => path,
-        },
-    }
-}
-
-fn is_uuid36(s: &str) -> bool {
-    s.len() == 36 && s.chars().all(|c| c.is_ascii_hexdigit() || c == '-')
-}
-
-/// The uuid half of a `<ISO-timestamp>_<uuid>` stem.
-///
-/// Split from the RIGHT: the timestamp half is `2026-08-03T07-57-44-591Z`, full
-/// of the same `-` the uuid uses, so only the underscore separates them
-/// reliably. The uuid itself never contains one.
-fn uuid_in_stem(stem: &str) -> Option<&str> {
-    let (ts, id) = stem.rsplit_once('_')?;
-    (!ts.is_empty() && is_uuid36(id)).then_some(id)
-}
-
-/// A transcript is `<ISO-timestamp>_<uuid>.jsonl`.
-///
-/// The walk is recursive, so this predicate carries the whole filter. Requiring
-/// both halves of the stem keeps this from claiming anything an operator (or a
-/// future pi) drops beside the transcripts — an editor's `.jsonl~`, an export,
-/// a hand-made note — any of which a byte cursor would otherwise tail as if it
-/// were a session.
-fn is_transcript(path: &Path) -> bool {
-    path.file_name()
-        .and_then(|n| n.to_str())
-        .and_then(|n| n.strip_suffix(".jsonl"))
-        .and_then(uuid_in_stem)
-        .is_some()
-}
-
-/// The uuid after the `_` in the filename IS the session id.
-///
-/// Deliberately the only source of it: the header's `id` field holds the same
-/// value, but taking it from both would give two answers that could disagree,
-/// and the path is what the cursor keys on. The uuid is also what the hook
-/// source reports as `session_id`, so hook events and transcript events for one
-/// run land on a single timeline.
-fn session_id_from_path(path: &Path) -> Option<String> {
-    path.file_name()
-        .and_then(|n| n.to_str())
-        .and_then(|n| n.strip_suffix(".jsonl"))
-        .and_then(uuid_in_stem)
-        .map(str::to_string)
-}
-
-/// `pi-<project>`, from the real `cwd` on the transcript's header.
-///
-/// NOT decoded from the directory name. pi encodes the working directory as
-/// `--${cwd.replace(/^[/\\]/,"").replace(/[/\\:]/g,"-")}--`, mapping every
-/// separator to `-` while leaving literal `-` alone — so the encoding is not
-/// invertible, and the loss is not a corner case: the probe's own
-/// `/tmp/probe-pi` becomes `--tmp-probe-pi--`, which decodes back to
-/// `/tmp/probe/pi` and yields the project name `pi` instead of `probe-pi`. The
-/// `cwd` field on line 1 is exact.
-///
-/// The scheme matches what the hook-activity source derives, so hook events and
-/// transcript events for one run land under a single agent rather than two that
-/// look unrelated.
-fn agent_id_from_path(_path: &Path, header: &[String]) -> Option<String> {
-    let cwd = header.iter().find_map(|line| {
-        serde_json::from_str::<Value>(line)
-            .ok()?
-            .get("cwd")?
-            .as_str()
-            .map(str::to_string)
-    })?;
-    let project = transform::sanitize_id_part(
-        cwd.trim_end_matches('/')
-            .rsplit('/')
-            .find(|p| !p.is_empty())?,
-    );
-    (!project.is_empty()).then(|| format!("pi-{project}"))
-}