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

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

@@ -1,158 +0,0 @@
-//! GitHub Copilot CLI session capture — a [`filetail`](crate::filetail) adapter.
-//!
-//! Copilot writes one live-appended JSONL transcript per session at
-//! `~/.copilot/session-state/<sessionId>/events.jsonl` (`COPILOT_HOME`
-//! overrides the home). We open them read-only and turn each newly-appended
-//! line into AgentEye events; Copilot's own files are never written, moved or
-//! deleted.
-//!
-//! # The file is genuinely append-only, which is why a byte cursor is safe
-//!
-//! Proven rather than assumed, because two of the CLIs probed for this engine
-//! rewrite their transcripts without saying so (see [`RereadPolicy`]). A
-//! session was captured, resumed, and captured again: same inode, and the
-//! first 39,797 bytes of the resumed file are md5-identical to the entire
-//! pre-resume file. A resume appends `session.resume` and carries on — it does
-//! not rewrite, renumber, or compact anything. Hence
-//! [`RereadPolicy::ByteCursor`].
-//!
-//! # Siblings that must not be tailed
-//!
-//! Every one of these lives inside the same tree, and each breaks something
-//! different:
-//!
-//! * `workspace.yaml` — rewritten **in place** on every resume (`updated_at`
-//!   moves), so a byte cursor would either re-ship it or resume mid-record.
-//!   It is also YAML, not JSONL.
-//! * `checkpoints/index.md` — rewritten in place as checkpoints are added, and
-//!   Markdown rather than JSON.
-//! * `session.db`, and `session-store.db` / `-wal` / `-shm` one level up —
-//!   SQLite. A byte cursor over a page-rewriting binary file produces garbage
-//!   indefinitely, and the `-wal` here is 900 KB and churning.
-//! * `logs/process-*.log` — process-scoped, NOT session-scoped: one file spans
-//!   every session a `copilot` process ran, so its lines have no single session
-//!   id to file events under. Plain text, and it embeds base64 icon blobs.
-//!
-//! [`is_transcript`] excludes all of them by requiring the exact filename
-//! `events.jsonl`, which is an allowlist rather than a denylist on purpose:
-//! whatever artifact a future Copilot release drops in a session directory is
-//! excluded by default instead of needing to be discovered in production.
-//!
-//! # Scope
-//!
-//! This covers the records that carry the session: start/end, prompts,
-//! assistant turns, and tool calls with their results. `hook.start` / `hook.end`
-//! are deliberately not mapped — failproofai's own hook-activity source already
-//! ships those from the store that records the decision, and re-deriving them
-//! here would double-count every hook. Checkpoints, `files/`, `research/` and
-//! the SQLite indexes are out of scope by construction.
-
-pub mod transform;
-
-use std::path::{Path, PathBuf};
-
-use crate::filetail::{Format, RereadPolicy, no_seed_state};
-use serde_json::Value;
-
-/// One transcript per session directory: `<sessionId>/events.jsonl`.
-pub const FORMAT: Format = Format {
-    kind: "copilot",
-    is_source_file: is_transcript,
-    session_id_from_path,
-    agent_id_from_path,
-    agent_start: transform::agent_start,
-    // No priming needed, unlike the Claude source. Copilot writes
-    // `session.model_change` before the first `user.message`, so the opening
-    // prompt already has a model by the time it is transformed — which also
-    // means this source avoids the live-tail-vs-re-read hazard documented on
-    // `Format::seed_state`.
-    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 = "copilot";
-
-/// The one file per session directory that is a transcript.
-const TRANSCRIPT_FILE: &str = "events.jsonl";
-
-/// Directory holding one subdirectory per session.
-const SESSION_STATE_DIR: &str = "session-state";
-
-/// Where Copilot keeps its per-session directories.
-///
-/// `COPILOT_HOME` overrides `~/.copilot` — the same override the CLI itself
-/// honours, and the one `lib/copilot-sessions.ts` reads, so a machine that
-/// relocates its Copilot home stays covered by both halves of the product.
-pub fn session_state_root() -> PathBuf {
-    if let Some(home) = std::env::var_os("COPILOT_HOME") {
-        return PathBuf::from(home).join(SESSION_STATE_DIR);
-    }
-    std::env::var_os("HOME")
-        .map(PathBuf::from)
-        .unwrap_or_default()
-        .join(".copilot")
-        .join(SESSION_STATE_DIR)
-}
-
-/// A transcript is any file named exactly `events.jsonl`.
-///
-/// The directory walk is recursive, so this predicate carries the whole filter.
-/// See the module docs for what each excluded sibling would break; the point of
-/// matching the name exactly is that they are excluded by construction rather
-/// than by a list that has to be kept current.
-fn is_transcript(path: &Path) -> bool {
-    path.file_name().and_then(|n| n.to_str()) == Some(TRANSCRIPT_FILE)
-}
-
-/// The session id is the name of the directory containing the transcript.
-///
-/// Every transcript is `events.jsonl`, so unlike Claude the filename carries
-/// nothing — the directory does. Verified on disk: the directory name equals
-/// `session.start`'s `data.sessionId` in all 4 live sessions, including one
-/// started with an explicitly chosen `--session-id`.
-///
-/// Deliberately the only source of it. The header carries the same id, and
-/// reading both would give two answers that can disagree — while the path is
-/// what the cursor keys on and what the hook source's `sessionId` must match
-/// for hook events and transcript events to land on one timeline.
-fn session_id_from_path(path: &Path) -> Option<String> {
-    if !is_transcript(path) {
-        return None;
-    }
-    let dir = path.parent()?.file_name()?.to_str()?;
-    // A transcript sitting directly in the session-state root has no session
-    // directory; a synthetic id is better than filing it under the root's name,
-    // which every such file would share.
-    (!dir.is_empty() && dir != SESSION_STATE_DIR).then(|| dir.to_string())
-}
-
-/// `copilot-<project>`, from the real `cwd` on the transcript's first lines.
-///
-/// NOT decoded from the directory name, which is a bare session UUID and
-/// encodes no path at all — Copilot keeps the working directory as a real
-/// field (`session.start` → `data.context.cwd`, repeated on `session.resume`),
-/// so there is nothing to invert and nothing to get wrong.
-///
-/// The scheme matches what the Claude and hook sources derive, so a hook event
-/// and the 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("data")?
-            .get("context")?
-            .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!("copilot-{project}"))
-}

package/.next/standalone/crates/fpai-collect/src/sources/copilot/transform.rs

@@ -1,512 +0,0 @@
-//! GitHub Copilot CLI `events.jsonl` line → AgentEye events.
-//!
-//! Every function here is pure in `(line, ctx, offset, state)`, which is what
-//! makes a live tail and a full re-read byte-identical and therefore
-//! dedup-collapsible. Nothing may consult the clock, a counter, or anything
-//! outside its arguments.
-//!
-//! # Record shapes, verified on disk
-//!
-//! Read from 4 live sessions under `~/.copilot/session-state/` plus 3 captured
-//! transcripts (Copilot CLI 1.0.77, 140 records total). Every line — with no
-//! exceptions in that corpus — is:
-//!
-//! ```text
-//! {type, data, id, timestamp, parentId}
-//! ```
-//!
-//! `timestamp` is ISO-8601 UTC with **milliseconds** on every record
-//! (`2026-08-03T07:56:17.824Z`; 140/140 matched, and each file was monotonic).
-//! `parentId` is null exactly once per file, on the opening record. The
-//! per-`type` payloads this module reads:
-//!
-//! ```text
-//! session.start            data.context.cwd, data.copilotVersion, data.sessionId
-//! session.resume           data.context.cwd, data.selectedModel
-//! session.model_change     data.newModel
-//! system.message           data.content  ← the ~31 KB system prompt; SKIPPED
-//! user.message             data.content (raw) + data.transformedContent (wrapped)
-//! assistant.message        data.{content, model, outputTokens, toolRequests[],
-//!                                messageId, turnId, interactionId}
-//! tool.execution_start     data.{toolCallId, toolName, arguments, model}
-//! tool.execution_complete  data.{toolCallId, success, result.content, model}
-//! assistant.turn_start/_end, session.shutdown, hook.start/hook.end
-//! ```
-//!
-//! # Deliberate omissions, each for a measured reason
-//!
-//! * **`system.message` is skipped entirely** rather than truncated. It is the
-//!   31,047-character system prompt, it is not conversation, and a resume
-//!   re-emits it *verbatim* — the captured resumed session carries two copies
-//!   of the identical blob. Truncating would still put ~62 KB of boilerplate on
-//!   the session timeline across a couple of resumes and would still render as
-//!   a message the user never sent; skipping costs nothing that is not already
-//!   in the prompt/response pairs.
-//! * **`assistant.message.toolRequests[]` does not become `tool_use`.** The
-//!   same call is announced there AND in its own `tool.execution_start` record.
-//!   Across all 7 transcripts the two sets were identical (no requested-but-
-//!   unstarted call, no started-but-unrequested call), so emitting both would
-//!   double every tool call in the product. `tool.execution_start` wins because
-//!   it is the record that pairs with `tool.execution_complete`.
-//! * **`reasoningText` / `reasoningOpaque` are dropped.** `reasoningText` is
-//!   *repeated verbatim* on every later `assistant.message` of the same
-//!   interaction — in the captured corpus turn 1 always carries turn 0's
-//!   reasoning, byte-identical, including the opaque blob. Emitting it would
-//!   attribute thinking to a turn that did not produce it.
-//! * **`result.detailedContent` is dropped.** For `bash` it equals
-//!   `result.content` exactly; for `view` it is a synthetic `git diff` rendering
-//!   of the same bytes. Either way it doubles the payload and adds no
-//!   information.
-//! * **`user.message.transformedContent` is dropped in favour of `content`.**
-//!   The transformed form wraps the prompt in an injected
-//!   `<current_datetime>` and `<system_reminder>` — machinery, not what the
-//!   human typed.
-//! * **`hook.start` / `hook.end` produce nothing.** failproofai's own
-//!   hook-activity source already ships those, from the store that records the
-//!   *decision*; re-deriving them here would double-count every hook on the
-//!   `/hooks` page. Their `input` also embeds whole tool results, so they are
-//!   the largest records after the system prompt.
-
-use serde_json::{Map, Value, json};
-
-use crate::cursor::TailState;
-use crate::filetail::Ctx;
-
-/// Longest derived id component.
-const MAX_ID_PART: usize = 48;
-
-/// Longest `goal` excerpt kept on `agent_start`.
-const MAX_GOAL_CHARS: usize = 1024;
-
-/// Make a derived id component safe and bounded.
-///
-/// Kept local rather than shared with the Claude or hook sources, which each
-/// hold their own copy. The three MUST agree — a hook event and a transcript
-/// event for one run landing under two agent ids is the defect this scheme
-/// exists to prevent — but that agreement is asserted by a test rather than by
-/// a shared symbol, so no source can be refactored into breaking another.
-pub fn sanitize_id_part(s: &str) -> String {
-    let cleaned: String = s
-        .chars()
-        .map(|c| {
-            if c.is_ascii_alphanumeric() || c == '.' || c == '_' {
-                c
-            } else {
-                '-'
-            }
-        })
-        .collect();
-    cleaned
-        .split('-')
-        .filter(|p| !p.is_empty())
-        .collect::<Vec<_>>()
-        .join("-")
-        .chars()
-        .take(MAX_ID_PART)
-        .collect()
-}
-
-/// Normalize a Copilot timestamp to the microsecond form ingest expects,
-/// offset by `index` microseconds.
-///
-/// Copilot writes milliseconds (`…T07:56:17.824Z`); the hook source emits
-/// microseconds. Both streams share one session timeline, so they must sort
-/// against each other — a millisecond string and a microsecond string do not
-/// compare correctly as text, and the server orders by `(ts, random id)`.
-///
-/// The offset keeps several events derived from ONE line in order. Index 0 is
-/// untouched, so the primary event of a line keeps the line's exact time.
-///
-/// Saturates within the second rather than carrying, so an offset can never
-/// reorder an event past a genuinely later line.
-pub fn with_index(ts: &str, index: usize) -> Option<String> {
-    let body = ts.strip_suffix('Z')?;
-    let (main, frac) = match body.rsplit_once('.') {
-        Some((m, f)) => (m, f),
-        None => (body, ""),
-    };
-    let mut micros: u32 = if frac.is_empty() {
-        0
-    } else {
-        let padded = format!("{frac:0<6}");
-        padded.get(..6)?.parse().ok()?
-    };
-    micros = (micros + index.min(999) as u32).min(999_999);
-    Some(format!("{main}.{micros:06}Z"))
-}
-
-fn ts_of(v: &Value) -> Option<&str> {
-    v.get("timestamp")?.as_str()
-}
-
-fn type_of(v: &Value) -> Option<&str> {
-    v.get("type")?.as_str()
-}
-
-fn str_at<'a>(data: &'a Value, key: &str) -> Option<&'a str> {
-    data.get(key)?.as_str().filter(|s| !s.is_empty())
-}
-
-/// `data.context.cwd` off a `session.start` or `session.resume` payload.
-fn context_cwd(data: &Value) -> Option<&str> {
-    data.get("context")?.get("cwd")?.as_str()
-}
-
-/// The envelope every emitted event carries.
-///
-/// `copilot_line_offset` is the dedup discriminator: the server collapses
-/// events by content hash, so two genuinely repeated records — the same prompt
-/// asked twice, the same command run twice — must hash differently or one of
-/// them disappears from the product.
-fn base(ctx: &Ctx, kind: &str, ts: &str, index: usize, offset: u64) -> Option<Map<String, Value>> {
-    let mut m = Map::new();
-    m.insert("timestamp".into(), json!(with_index(ts, index)?));
-    m.insert("session_id".into(), json!(ctx.session_id));
-    m.insert("agent_id".into(), json!(ctx.agent_id));
-    m.insert("type".into(), json!(kind));
-    m.insert("environment".into(), json!(ctx.environment));
-    m.insert("copilot_line_offset".into(), json!(offset));
-    if index > 0 {
-        m.insert("copilot_block_index".into(), json!(index));
-    }
-    Some(m)
-}
-
-/// Build the session's `agent_start` from its header lines.
-///
-/// Scans for the first line carrying a usable timestamp rather than trusting
-/// line 1: a session with no `agent_start` is not merely incomplete, the server
-/// selects sessions on that event, so it is absent from the product entirely.
-///
-/// The `goal` is reachable from the header despite the enormous
-/// `system.message` sitting in front of it — the first `user.message` is the
-/// 4th record and the whole prefix is ~32 KB, well inside the engine's 1 MB /
-/// 64-line header budget.
-pub fn agent_start(header: &[String], ctx: &Ctx, offset: u64) -> Option<(Value, Option<String>)> {
-    let mut first_ts: Option<String> = None;
-    let mut cwd: Option<String> = None;
-    let mut version: Option<String> = None;
-    let mut branch: Option<String> = None;
-    let mut goal: Option<String> = None;
-
-    for line in header {
-        let Ok(v) = serde_json::from_str::<Value>(line) else {
-            continue;
-        };
-        if first_ts.is_none()
-            && let Some(t) = ts_of(&v)
-        {
-            first_ts = Some(t.to_string());
-        }
-        let Some(data) = v.get("data") else {
-            continue;
-        };
-        if cwd.is_none()
-            && let Some(c) = context_cwd(data)
-        {
-            cwd = Some(c.to_string());
-        }
-        if version.is_none()
-            && let Some(x) = str_at(data, "copilotVersion")
-        {
-            version = Some(x.to_string());
-        }
-        // Present only when the session's cwd is inside a repository. The
-        // captured corpus has none — the probe ran in a non-git `/tmp` dir —
-        // but `lib/copilot-sessions.ts` records `context.{gitRoot, branch, …}`
-        // as verified against Copilot 1.0.39, so it is taken opportunistically
-        // rather than assumed absent.
-        if branch.is_none()
-            && let Some(b) = data.get("context").and_then(|c| c.get("branch"))
-            && let Some(b) = b.as_str().filter(|s| !s.is_empty())
-        {
-            branch = Some(b.to_string());
-        }
-        // The session's goal is its first real human prompt — `content`, not
-        // `transformedContent`, which prepends injected machinery.
-        if goal.is_none()
-            && type_of(&v) == Some("user.message")
-            && let Some(text) = str_at(data, "content")
-        {
-            goal = Some(text.chars().take(MAX_GOAL_CHARS).collect());
-        }
-    }
-
-    let ts = first_ts?;
-    let mut m = base(ctx, "agent_start", &ts, 0, offset)?;
-    if let Some(g) = goal {
-        m.insert("goal".into(), json!(g));
-    }
-    if let Some(c) = cwd {
-        m.insert("copilot_cwd".into(), json!(c));
-    }
-    if let Some(b) = branch {
-        m.insert("copilot_git_branch".into(), json!(b));
-    }
-    if let Some(v) = version {
-        m.insert("copilot_version".into(), json!(v));
-    }
-    Some((Value::Object(m), Some(ts)))
-}
-
-/// The single `agent_end`, derived deterministically from the last timestamp
-/// and the file size (used as its offset, so it is unique per session).
-///
-/// Index 999 puts it after every content event sharing that timestamp — the
-/// server's tie-break is a random id, so without this a session can appear to
-/// end before its last turn. It matters here in particular because
-/// `session.shutdown` follows the final `assistant.turn_end` by ~70 ms, and a
-/// crashed session has no shutdown record at all, leaving `agent_end` sharing
-/// the exact timestamp of the last real event.
-pub fn agent_end(ctx: &Ctx, last_ts: &str, size: u64) -> Value {
-    match base(ctx, "agent_end", last_ts, 999, size) {
-        Some(m) => Value::Object(m),
-        // `base` only fails on an unparseable timestamp, which cannot happen
-        // here: `last_ts` came from a line this transform already accepted.
-        None => json!({
-            "timestamp": last_ts,
-            "session_id": ctx.session_id,
-            "agent_id": ctx.agent_id,
-            "type": "agent_end",
-            "environment": ctx.environment,
-        }),
-    }
-}
-
-/// One content line to its timestamp and the events it yields.
-///
-/// The `type` dispatch is an explicit allowlist rather than Claude's
-/// "no timestamp means metadata" trick: every Copilot record carries a
-/// timestamp, including the ones that must produce nothing, so there is no
-/// free signal to lean on. Unknown types fall through to no events while still
-/// returning their timestamp, so a record type a future Copilot release adds
-/// costs nothing and still advances `agent_end`.
-pub fn transform_line(
-    line: &str,
-    ctx: &Ctx,
-    offset: u64,
-    state: &mut TailState,
-) -> (Option<String>, Vec<Value>) {
-    let Ok(v) = serde_json::from_str::<Value>(line) else {
-        return (None, Vec::new());
-    };
-    let Some(ts) = ts_of(&v).map(str::to_string) else {
-        return (None, Vec::new());
-    };
-    let Some(data) = v.get("data") else {
-        return (Some(ts), Vec::new());
-    };
-
-    let events = match type_of(&v) {
-        // A user prompt names no model, and the server builds a
-        // `model_request` row's summary from the model alone. Copilot writes
-        // `session.model_change` BEFORE the first `user.message` in every
-        // captured transcript, which is why this source needs no `seed_state`
-        // — and therefore does not inherit the live-tail-vs-re-read hazard
-        // documented on `Format::seed_state`.
-        Some("session.model_change") => {
-            remember_model(state, str_at(data, "newModel"));
-            Vec::new()
-        }
-        // A resume re-announces the model it is resuming with, so the first
-        // prompt after a resume is never modelless either.
-        Some("session.resume") => {
-            remember_model(state, str_at(data, "selectedModel"));
-            Vec::new()
-        }
-        Some("user.message") => user_message(data, ctx, &ts, offset, state),
-        Some("assistant.message") => assistant_message(data, ctx, &ts, offset, state),
-        Some("tool.execution_start") => tool_use(data, ctx, &ts, offset, state),
-        Some("tool.execution_complete") => tool_result(data, ctx, &ts, offset, state),
-        _ => Vec::new(),
-    };
-    (Some(ts), events)
-}
-
-fn remember_model(state: &mut TailState, model: Option<&str>) {
-    if let Some(m) = model {
-        state.last_model = Some(m.to_string());
-    }
-}
-
-/// A human prompt → `model_request`.
-fn user_message(data: &Value, ctx: &Ctx, ts: &str, offset: u64, state: &TailState) -> Vec<Value> {
-    let Some(text) = str_at(data, "content") else {
-        return Vec::new();
-    };
-    let Some(mut m) = base(ctx, "model_request", ts, 0, offset) else {
-        return Vec::new();
-    };
-    if let Some(model) = &state.last_model {
-        m.insert("model".into(), json!(model));
-    }
-    m.insert(
-        "messages".into(),
-        json!([{ "role": "user", "content": text }]),
-    );
-    if let Some(id) = str_at(data, "interactionId") {
-        m.insert("copilot_interaction_id".into(), json!(id));
-    }
-    vec![Value::Object(m)]
-}
-
-/// An assistant turn → exactly one `model_response`.
-///
-/// Emitted even when `content` is empty, which is the case for every
-/// tool-calling turn. That is deliberate and load-bearing: `outputTokens`
-/// appears ONLY on `assistant.message` — `tool.execution_start` carries none —
-/// so skipping empty-content turns the way the Claude source skips empty text
-/// blocks would silently drop most of a session's output tokens. Measured on
-/// the captured corpus: 209 of 320 tokens in one session, 188 of 304 in
-/// another. The `content` key is omitted rather than set to `""`, so a
-/// consumer can tell "the model said nothing, it called tools" from "the model
-/// said the empty string".
-fn assistant_message(
-    data: &Value,
-    ctx: &Ctx,
-    ts: &str,
-    offset: u64,
-    state: &mut TailState,
-) -> Vec<Value> {
-    let model = str_at(data, "model");
-    remember_model(state, model);
-
-    let Some(mut m) = base(ctx, "model_response", ts, 0, offset) else {
-        return Vec::new();
-    };
-    m.insert("role".into(), json!("assistant"));
-    if let Some(text) = str_at(data, "content") {
-        m.insert("content".into(), json!(text));
-    }
-    if let Some(model) = model {
-        m.insert("model".into(), json!(model));
-    }
-    if let Some(n) = data.get("outputTokens").and_then(|x| x.as_u64()) {
-        m.insert("output_tokens".into(), json!(n));
-    }
-    if let Some(id) = str_at(data, "messageId") {
-        m.insert("copilot_message_id".into(), json!(id));
-    }
-    if let Some(id) = str_at(data, "turnId") {
-        m.insert("copilot_turn_id".into(), json!(id));
-    }
-    if let Some(id) = str_at(data, "interactionId") {
-        m.insert("copilot_interaction_id".into(), json!(id));
-    }
-    // The count, not the calls: the calls themselves arrive as their own
-    // `tool.execution_start` records. Keeping the count is what makes a
-    // contentless turn legible instead of looking like a dropped response.
-    let requests = data
-        .get("toolRequests")
-        .and_then(|r| r.as_array())
-        .map(|r| r.len())
-        .unwrap_or(0);
-    if requests > 0 {
-        m.insert("copilot_tool_request_count".into(), json!(requests));
-    }
-    vec![Value::Object(m)]
-}
-
-/// A tool call → `tool_use`, and the name is stashed for its result.
-fn tool_use(data: &Value, ctx: &Ctx, ts: &str, offset: u64, state: &mut TailState) -> Vec<Value> {
-    let Some(mut m) = base(ctx, "tool_use", ts, 0, offset) else {
-        return Vec::new();
-    };
-    let name = str_at(data, "toolName").unwrap_or("tool");
-    // Fall back to an offset-derived id so two identical parallel calls cannot
-    // hash-collapse into one row, and so a result can still find nothing
-    // rather than finding the wrong call.
-    let id = str_at(data, "toolCallId")
-        .map(str::to_string)
-        .unwrap_or_else(|| format!("copilot-{offset}"));
-    state.remember_tool(id.clone(), name.to_string());
-
-    m.insert("tool_name".into(), json!(name));
-    m.insert("tool_call_id".into(), json!(id));
-    if let Some(args) = data.get("arguments") {
-        m.insert("input".into(), args.clone());
-    }
-    if let Some(model) = str_at(data, "model") {
-        m.insert("model".into(), json!(model));
-    }
-    if let Some(t) = str_at(data, "turnId") {
-        m.insert("copilot_turn_id".into(), json!(t));
-    }
-    vec![Value::Object(m)]
-}
-
-/// A tool completion → `tool_result`, paired to its call BY ID.
-///
-/// Never by position. Copilot runs a turn's tool calls in parallel and writes
-/// each completion when it lands, so completions arrive out of order with
-/// respect to their starts — a captured session starts (`glob`, `view`) and
-/// completes (`view`, `glob`), 47 ms apart. Pairing positionally would label
-/// every row in that turn with the wrong tool, which is worse than labelling
-/// none of them: a wrong tool name is indistinguishable from a right one.
-fn tool_result(data: &Value, ctx: &Ctx, ts: &str, offset: u64, state: &TailState) -> Vec<Value> {
-    let Some(mut m) = base(ctx, "tool_result", ts, 0, offset) else {
-        return Vec::new();
-    };
-    if let Some(id) = str_at(data, "toolCallId") {
-        m.insert("tool_call_id".into(), json!(id));
-        // The tool's name appears on NO completion record. Without carrying it
-        // from the call, every tool result is a blank row in the product.
-        if let Some(name) = state.tool_name(id) {
-            m.insert("tool_name".into(), json!(name));
-        }
-    }
-    let output = data
-        .get("result")
-        .and_then(|r| r.get("content"))
-        .map(stringify);
-    if let Some(text) = &output {
-        m.insert("output".into(), json!(text));
-    }
-    // Absent `success` is treated as success: every record in the captured
-    // corpus carries it, and inventing a failure for a field a future release
-    // might drop would flag healthy sessions as broken.
-    if data.get("success").and_then(|s| s.as_bool()) == Some(false) {
-        // The server's `is_error` is a truthiness check, so this must never be
-        // an empty string — a failure with no output would otherwise render as
-        // a success.
-        let reason = output
-            .as_deref()
-            .map(str::trim)
-            .filter(|r| !r.is_empty())
-            .unwrap_or("copilot tool execution failed");
-        m.insert("error".into(), json!(reason));
-        m.insert("error_type".into(), json!("copilot_tool_error"));
-    }
-    if let Some(model) = str_at(data, "model") {
-        m.insert("model".into(), json!(model));
-    }
-    if let Some(t) = str_at(data, "turnId") {
-        m.insert("copilot_turn_id".into(), json!(t));
-    }
-    vec![Value::Object(m)]
-}
-
-/// Render a tool result payload as text.
-///
-/// Copilot writes `result.content` as a string in every captured record, but
-/// the field is untyped JSON — a structured payload is rendered rather than
-/// dropped, so a future tool returning an object does not produce an empty row.
-fn stringify(v: &Value) -> String {
-    match v {
-        Value::String(s) => s.clone(),
-        Value::Array(items) => items
-            .iter()
-            .map(|b| match b {
-                Value::String(s) => s.clone(),
-                other => other
-                    .get("text")
-                    .and_then(|t| t.as_str())
-                    .map(str::to_string)
-                    .unwrap_or_else(|| other.to_string()),
-            })
-            .collect::<Vec<_>>()
-            .join("\n"),
-        Value::Null => String::new(),
-        other => other.to_string(),
-    }
-}