Cross-harness context-compaction provenance (kind v1.2.0)

CLAUDE.md

@@ -194,15 +194,20 @@ Format references for the agent on-disk formats live at `docs/agents/formats/` 
 
 - `PathOrRef::Path` is `Box<Path>` to avoid a large enum variant size difference.
 - Shared derivation: `toolpath-convo` provides the provider-agnostic `ConversationView → Path` mapping (`toolpath_convo::derive_path`). New conversation providers build on it rather than re-implementing the mapping.
-- Provider extras: `Turn.extra` and `WatcherEvent::Progress.data` use provider-namespaced keys (e.g. `extra["claude"]`, `extra["gemini"]`) so trait-only consumers can reach provider metadata without importing provider types.
-- Path kinds: `PathMeta.kind` is an optional URI naming a hosted kind spec; URIs are immutable and semver-versioned. The only one defined is `https://toolpath.net/kinds/agent-coding-session/v1.1.0` (`toolpath::v1::PATH_KIND_AGENT_CODING_SESSION`); every conversation → `Path` derivation sets it. Spec sources: `site/kinds/<name>/<version>/{index.md,schema.json}` (schema.json symlinks into `crates/path-cli/kinds/`, which `p validate` bundles). RFC section: "Document Kind".
+- The conversation IR (`toolpath-convo`) is a **closed typed set** — no catch-all `extra`/passthrough fields on `Turn`, `Compaction`, or any other IR type; provider detail either earns a typed optional field or is accepted as lossy. Native-layer catch-alls (e.g. `toolpath-claude`'s `ConversationEntry.extra` serde-flatten) are fine — they represent the *source format*, not the IR. `WatcherEvent::Progress.data` (a live event stream, not the IR) still uses provider-namespaced keys (e.g. `data["claude"]`).
+- Path kinds: `PathMeta.kind` is an optional URI naming a hosted kind spec; URIs are immutable and semver-versioned. The only kind defined is `agent-coding-session`, currently at `https://toolpath.net/kinds/agent-coding-session/v1.2.0` (`toolpath::v1::PATH_KIND_AGENT_CODING_SESSION`; `…_V1_1_0`/`…_V1_0_0` name the superseded URIs, still registered and bundled); v1.2.0 adds the `conversation.compact` step type on top of v1.1.0's message-level token accounting. Every conversation → `Path` derivation sets it. Spec sources: `site/kinds/<name>/<version>/{index.md,schema.json}` (schema.json symlinks into `crates/path-cli/kinds/`, which `p validate` bundles — all versions). RFC section: "Document Kind".
 
 ### Token accounting (kind v1.1.0)
 
 - Two optional keys on `conversation.append`/`Turn`: `token_usage` is the total for a provider message, stamped on the message group's final step (`Σ` over a path = session total); `attributed_token_usage` is a step's own attributed spend, populated only where the source genuinely reports per-step spend. `Turn.group_id` groups the steps of one message (remainder = group total − Σ attributed, computed not stored).
 - Hard rules: never stamp a cumulative counter, a repeated message total, or zero-filled placeholders onto a step. Claude's per-line `usage` is a cumulative streaming snapshot, not a per-block cost — take the field-wise-max group total and never derive attribution from it. Codex must difference the cumulative `total_token_usage`, never sum `last_token_usage` (re-emitted stale). pi/opencode decode all-zero wire counters as `None`.
 - `breakdowns` is an optional decomposition of a top-level class into sub-classes (e.g. `breakdowns["output"]["reasoning"]`). Informational only — **never summed into any total**; invariant `Σ(inner) ≤ parent`; omitted when empty. Gemini and OpenCode fold reasoning tokens into `output_tokens` and record this breakdown (projectors un-fold on the reverse path); Codex differences `reasoning_output_tokens`; Claude reports no breakdown.
 
+### Compaction (kind v1.2.0)
+
+- The `conversation.compact` step type records a context-compaction boundary as its own step between the turns it separates (turns after the boundary parent on it, so the `head`-ancestry walk crosses it in order). `structural` fields are all optional but `type`: `trigger` (`auto`|`manual`), `summary`, `pre_tokens`, and `kept` (ids of prior turns surviving verbatim — always a contiguous parent-chain run, oldest first, whose first element is the anchor; empty = wholesale). It is not a turn (no `text`/`role`/`tool_uses`); `step.actor` is `tool:<provider>`.
+- Compaction provenance is itself a closed typed set — no catch-all `extra`; native detail richer than the contiguous run (e.g. Claude's replay-pinned messages) is deliberately not carried, and round-trips are lossy beyond these fields. The IR carries the anchor as `Compaction.kept_from`; `toolpath_convo::expand_kept` expands it to the run. Steps whose parents were rewired by derive-splicing also carry `source_parent` (string|null) on their structural change. Each provider's `Item::Compaction` derivation populates it and projects it back to harness markers.
+
 ### Providers
 
 - Data locations: claude `~/.claude/projects/` (JSONL); gemini `~/.gemini/tmp/<project>/chats/` (project slot is a friendly name or a path SHA-256; sub-agents in sibling UUID dirs); codex `~/.codex/sessions/YYYY/MM/DD/rollout-*.jsonl` (date-bucketed, not project-keyed); copilot `~/.copilot/session-state/<id>/events.jsonl` (`COPILOT_HOME` overrides; global, not project-keyed); opencode `~/.local/share/opencode/opencode.db` (read-only SQLite); cursor `state.vscdb` global SQLite under the Cursor app-support dir; pi `~/.pi/agent/sessions/` (tree in a single file, preserved as a DAG).
@@ -211,6 +216,7 @@ Format references for the agent on-disk formats live at `docs/agents/formats/` 
 - `toolpath-copilot` is a **preview** provider over a reverse-engineered `events.jsonl` schema; the reader is deliberately tolerant of shape variations and preserves unknown events. It is wired both directions: forward (`p import/list/show copilot`, `share`) and reverse via `CopilotProjector` (`p export copilot`, `resume` — writes `session-state/<id>/` plus a `session-store.db` row, only ever INSERTing a fresh id). The real `copilot --resume` loader is strict; its writer contract is documented in `docs/agents/formats/copilot-cli/writing-compatible.md` — follow it exactly.
 - File-change fidelity varies: codex and copilot embed real diffs/file content in the session log, so derived paths get a true `raw` perspective; opencode diffs a sibling snapshot git repo via `git2`, falling back to structural-only changes for gitignored paths; cursor round-trips numeric tool-dispatch ids via `TOOL_TABLE` so projector-written composers render correctly in Cursor.app (the cursor-agent CLI's separate protobuf store is not parsed).
 - `ConversationMetadata`/`SessionMeta` for claude/gemini/pi expose `first_user_message: Option<String>` (cheap, populated during the metadata pass) — used by pickers and any "list sessions by topic" surface.
+- Compaction mappings: **claude** projects fully empty assistant turns as Claude's own empty thinking block (Claude 2.1.216's transcript renderer aborts on `[{"type":"text","text":""}]`). **codex** (2026-07 payload): `compacted.message` is empty — the summary lives encrypted in `replacement_history` and the surviving turns are a prefix-keep the suffix-anchored `kept_from` contract cannot represent, so the boundary derives wholesale with no summary; the projector re-emits the `event_msg`/`context_compacted` marker the TUI renders from, and places the opening `turn_context` before the first real prompt (codex titles a backfilled session from the first user message). **copilot** (observed at 1.0.68): a successful `session.compaction_complete` becomes the typed `Item::Compaction` (`summary` ← `summaryContent`, `pre_tokens` ← `preCompactionTokens`; wholesale — Copilot reports removed-message counts, not surviving ids); failed compactions and `session.compaction_start` stay generic events; the projector emits the observed pair back and mirrors the summary to `checkpoints/NNN-*.md`. **pi**: `model_change`/`thinking_level_change`/`label` entries map to typed `Item::Event`s (chain links — `expand_kept` passes through them) and project back; the projector threads the chain's model context into assistant `provider`/`api` so pi restores the resumed session's model (verified in pi 0.72); compaction `details`/`fromHook` are deliberate loss.
 
 ### CLI behaviors
 

Cargo.toml

@@ -24,18 +24,18 @@ edition = "2024"
 license = "Apache-2.0"
 
 [workspace.dependencies]
-toolpath = { version = "0.7.0", path = "crates/toolpath" }
+toolpath = { version = "0.8.0", path = "crates/toolpath" }
 toolpath-convo = { version = "0.12.0", path = "crates/toolpath-convo" }
-toolpath-git = { version = "0.6.0", path = "crates/toolpath-git" }
+toolpath-git = { version = "0.7.0", path = "crates/toolpath-git" }
 toolpath-claude = { version = "0.13.0", path = "crates/toolpath-claude", default-features = false }
 toolpath-gemini = { version = "0.7.0", path = "crates/toolpath-gemini", default-features = false }
 toolpath-codex = { version = "0.7.0", path = "crates/toolpath-codex" }
 toolpath-copilot = { version = "0.2.0", path = "crates/toolpath-copilot" }
 toolpath-opencode = { version = "0.6.0", path = "crates/toolpath-opencode" }
 toolpath-cursor = { version = "0.3.0", path = "crates/toolpath-cursor" }
-toolpath-github = { version = "0.6.0", path = "crates/toolpath-github" }
-toolpath-dot = { version = "0.5.0", path = "crates/toolpath-dot" }
-toolpath-md = { version = "0.7.0", path = "crates/toolpath-md" }
+toolpath-github = { version = "0.7.0", path = "crates/toolpath-github" }
+toolpath-dot = { version = "0.6.0", path = "crates/toolpath-dot" }
+toolpath-md = { version = "0.8.0", path = "crates/toolpath-md" }
 toolpath-pi = { version = "0.7.0", path = "crates/toolpath-pi" }
 path-cli = { version = "0.17.0", path = "crates/path-cli" }
 pathbase-client = { version = "0.2.0", path = "crates/pathbase-client" }

RFC.md

@@ -296,10 +296,16 @@ unrecognized URIs should be treated as a generic path. Kind URIs are
 immutable, semver-versioned, and revisions ship at a new version URI.
 
 Defined kinds are listed at <https://toolpath.net/kinds/>. The only one defined
-so far is `https://toolpath.net/kinds/agent-coding-session/v1.0.0` — a path
-recording an AI coding conversation, where each conversational-turn step
-carries a `"conversation.append"` structural change with the turn's role,
-text, and so on. See the linked spec for the full contract.
+so far is `agent-coding-session`, currently at
+`https://toolpath.net/kinds/agent-coding-session/v1.2.0` — a path recording an
+AI coding conversation, where each conversational-turn step carries a
+`"conversation.append"` structural change with the turn's role, text, and so
+on, and context-compaction boundaries carry a `"conversation.compact"` step
+whose `kept` list names the prior turns that survived into the post-compaction
+window — always a contiguous parent-chain run, oldest first, anchored at its
+first element. The earlier `v1.1.0` URI (message-level token accounting,
+without `conversation.compact`) and `v1.0.0` URI both remain valid and
+documented. See the linked spec for the full contract.
 
 #### Actor Definitions