@bastani/atomic 0.9.18 → 0.9.19-alpha.10

package/docs/compaction.md

@@ -1,9 +1,18 @@
+---
+title: "Context and compaction"
+description: "Verbatim line compaction, when it runs, planning rungs, and branch summarization."
+---
+
 # Compaction & Branch Summarization
 
 LLMs have finite context windows. Atomic reduces transcript context with **verbatim line compaction** while preserving an exact count of recent context-visible messages as ordinary messages. Branch summarization is a separate, intentionally lossy feature used only when navigating away from a branch.
 
 Compaction runs entirely locally; no external compaction service is involved. It normally uses the active session model. If that model cannot rank the lines — a rate limit, a quota exhaustion, a provider error, a context overflow, or an empty plan — Atomic *borrows* the next model from your configured `fallbackModels` for that one planner request. **A configured fallback model may therefore receive the compaction transcript**, and it is sent with that provider's own credentials. Borrowing never changes the session's model or thinking level. The model only selects which lines to delete — Atomic reconstructs the retained text mechanically, so surviving lines are never rewritten.
 
+## On this page and its reference
+
+This page covers the concepts and normal use of compaction and branch summarization. Parameters, persistence, extension hooks, formats, settings, and historical formats live in the [Compaction reference](/compaction/reference).
+
 ## Overview
 
 | Mechanism | Trigger | Model output | Durable result |
@@ -17,6 +26,8 @@ There is one context-compaction door: `compact`.
 
 ## Verbatim Line Compaction
 
+<a id="what-verbatim-means"></a>
+
 ### What "verbatim" means
 
 Atomic serializes the compactable part of the conversation into role-tagged lines:
@@ -78,35 +89,11 @@ Use it for role constraints, invariants, and anything whose loss would silently
 
 ## Parameters
 
-The effective parameters appear in extension events and successful results:
-
-| Parameter | Default | Meaning |
-|---|---:|---|
-| `compression_ratio` | `0.5` | Fraction of compactable **lines to keep**, not a token ratio |
-| `preserve_recent` | `2` | Exact number of newest context-visible messages protected client-side |
-| `query` | Last visible user message | Relevance focus for deciding which older lines to retain |
-
-`preserve_recent` counts context-visible messages without aligning the boundary to a user turn. An assistant message or tool result may therefore begin the kept tail. Because such a tail can start or end mid-turn, the kept messages are not replayed as structured message blocks: they are serialized with the same transcript grammar as the compacted region and appended to the end of the boundary string, so the whole boundary reaches the provider as one message. Serialization of the kept tail is lossless — tool results keep their full text instead of being truncated at 16k characters, and images stay attached as image blocks rather than becoming `[image]` markers — so protected content is preserved, not merely summarized. A value of `0` protects no messages and makes the entire active transcript compactable. If `query` is absent, Atomic derives it from the last visible user message.
-
-One consequence is worth stating for Claude models that sign their reasoning. Because the kept tail is serialized into the boundary string rather than replayed as structured assistant messages, no `thinking` or `redacted_thinking` block survives a compaction boundary. Compaction therefore **intentionally resets the signed reasoning chain**: reasoning produced before a boundary is not carried across it. This is deliberate, and it is the first of the two remedies Anthropic documents for keep-tail compaction — carry the text and tool calls across, leave the thinking blocks behind — reached structurally rather than by a stripping pass. The tail's own content is unaffected: text, tool calls, and tool results cross the boundary losslessly. See [Preserved thinking and model switches](/models#preserved-thinking-and-model-switches) for how Atomic handles prefix changes *between* boundaries, which is a separate mechanism.
-
-The query is used whole and is never truncated. This matters for structured prompts: a truncated query would make section order the retention policy, because only the leading section could influence what the planner kept, and a constraint stated later in the prompt could not. Long queries are safe — an oversized planner request surfaces as an explicit provider-overflow failure rather than silent truncation — but `keepContext` tags, not query length, are the way to guarantee a span survives.
+Moved to [Compaction reference](/compaction/reference#parameters).
 
-Configure defaults in `~/.atomic/agent/settings.json` or `.atomic/settings.json`:
+### Per-model budgets
 
-```json
-{
-  "compaction": {
-    "enabled": true,
-    "reserveTokens": 16384,
-    "compression_ratio": 0.5,
-    "preserve_recent": 2,
-    "query": "optional focus"
-  }
-}
-```
-
-`reserveTokens` controls the automatic threshold that decides when compaction runs; it is not converted into a classifier line ratio. Manual calls can pass parameter overrides through the SDK.
+Moved to [Compaction reference](/compaction/reference#per-model-budgets).
 
 ## When compaction runs
 
@@ -117,6 +104,8 @@ Configure defaults in `~/.atomic/agent/settings.json` or `.atomic/settings.json`
 
 Exactly the configured recent-message tail is outside the compactable region; Atomic does not force the final logical turn to remain outside it. After successful automatic compaction, input queued before compaction started resumes automatically once the active run becomes idle; pressing Escape is not required to release it. Pressing Escape while compaction is active cancels it like other session operations. In isolated interactive mode, cancellation and host UI response frames use an independent RPC control lane, so they can reach the engine while the ordinary `compact` request is still pending instead of waiting behind it. Atomic writes a backup snapshot immediately before appending a compaction boundary.
 
+Automatic continuations belong to the original prompt, including repeated output-cap continuations. The prompt and `agent_settled` lifecycle finish only after the entire chain drains, so status integrations such as [Herdr](/herdr) stay working through quiet response waits and continuation gaps.
+
 Manual compaction is single-flight. A manual request made while another manual compaction is still in flight joins that run and receives its result instead of starting a second model call, so it emits no extra `compaction_start`/`compaction_end` pair and appends no extra boundary; `abortCompaction()` (Escape) still cancels the single owning run for every waiter. A manual request made while automatic compaction is active takes ownership: it cancels unfinished automatic work, abandons any pending automatic continuation, waits for the active turn and automatic state to settle, then runs the requested manual compaction. If the automatic boundary committed first, the manual request runs after it rather than reusing that result or writing concurrently. In the interactive TUI, `/compact` uses this takeover only for automatic compaction; a second manual compaction or branch summary is still refused. Ordinary text typed during compaction remains queued and flushes after its non-mid-turn completion, including cancellation or failure.
 
 The post-tool check stays inside the active Pi loop: it runs the rung ladder once for that completed tool turn, returns the rebuilt context to the loop, and never calls or schedules `agent.continue()`. If the caller installed `prepareNextTurn` or `prepareNextTurnWithContext`, Atomic invokes it after compaction with that rebuilt context and sends the caller's complete replacement context through the normal context transform to the next provider request. A `shouldStopAfterTurn` result of `true` takes precedence over queued steering and follow-up messages, including messages queued while the callback is running: Atomic leaves that work pending until a later explicit prompt or continuation instead of starting another provider request. Normal context reconstruction preserves provider tool-call/result protocol validity. Below-threshold tool turns follow the unchanged request path. Because the same active run resumes without emitting another `agent_start`, the interactive TUI replaces the compaction loader with its working spinner and keeps terminal progress active as soon as successful mid-turn compaction ends; streaming feedback therefore resumes immediately without waiting for another user interaction.
@@ -222,91 +211,23 @@ Context thresholds and persisted token-reduction statistics use API-aware normal
 
 ## Persistence and resume
 
-A successful run appends the existing pi-style `type:"compaction"` entry shape:
-
-```json
-{
-  "type": "compaction",
-  "id": "c1",
-  "parentId": "m9",
-  "timestamp": "2026-07-13T10:00:00.000Z",
-  "summary": "[User]: fix the failing test\n(filtered 42 lines)\n[Assistant]: Fixed.",
-  "firstKeptEntryId": "m7",
-  "tokensBefore": 51234,
-  "details": {
-    "strategy": "verbatim-lines",
-    "promptVersion": 3,
-    "rung": "planned",
-    "parameters": {"compression_ratio": 0.5, "preserve_recent": 2, "query": "fix the failing test"},
-    "stats": {"linesBefore": 812, "linesDeleted": 417, "linesKept": 395, "rangeCount": 63, "tokensBefore": 51234, "tokensAfter": 24980, "percentReduction": 51.2}
-  }
-}
-```
-
-`details.rung` is one of `"planned"` (a model ranked the lines — the session model **or** a borrowed fallback, including silent partial recovery), `"extension"` (a `session_before_compact` override), or `"fresh"` (the compactable conversation was discarded and a new context window started). `details.plannerModel` is present **only** when a borrowed fallback model ranked the lines:
-
-```json
-"details": {
-  "strategy": "verbatim-lines",
-  "promptVersion": 3,
-  "rung": "planned",
-  "plannerModel": {"provider": "openai", "id": "gpt-5.1", "thinkingLevel": "high"}
-}
-```
-
-There is no format-version bump and no new entry type. Both `"fresh"` and `plannerModel` are additive: they are absent on every existing entry and on any compaction that used the session model, so old readers are unaffected. A `"fresh"` boundary that had to drop the `preserve_recent` tail persists `firstKeptEntryId: null`.
-
-A `compaction` entry is active only when `details.strategy === "verbatim-lines"`. On rebuild, Atomic emits one visible custom-role boundary message: the durable `summary` with the kept tail—the entries from `firstKeptEntryId` up to the boundary—serialized and concatenated onto its end. The tail is never restored as separate assistant/tool-result blocks, so a tail that starts or ends mid-turn cannot produce out-of-order provider blocks; images inside the tail ride along as image blocks on that same boundary message. When no pre-boundary context-visible message is retained—such as with `preserve_recent: 0`—`firstKeptEntryId` is `null` and the boundary carries the `summary` alone. Messages appended after the boundary are always replayed as real messages. The boundary is converted to a user-role provider message and shown in the TUI as a collapsible compaction card.
-
-Resume does not rerun planning or re-derive deletions: the exact compacted string and nullable tail boundary are already in JSONL. Existing records with a string `firstKeptEntryId` keep their original resume behavior. Legacy `context_compaction` logical-deletion records and old `compaction` summary records without the discriminator are inert archival data. Their historical omissions are not reapplied when an old session resumes.
+Moved to [Compaction reference](/compaction/reference#persistence-and-resume).
 
 ## Extension hooks
 
-### `session_before_compact`
-
-Extensions may cancel or provide a complete replacement for the prepared region:
+Moved to [Compaction reference](/compaction/reference#extension-hooks).
 
-```typescript
-pi.on("session_before_compact", async (event) => {
-  const { reason, parameters, preparation, branchEntries, signal } = event;
-  if (signal.aborted) return { cancel: true };
-
-  // Optional offline override. It must contain non-whitespace text.
-  if (reason === "manual" && branchEntries.length > 100) {
-    return { compactedText: preparation.region.lines.slice(0, 40).join("\n") };
-  }
-});
-```
+### `session_before_compact`
 
-`preparation` is a deep-frozen clone. An override changes only the compacted region text; Atomic retains the prepared boundary and persists the supplied text verbatim. Empty/whitespace text is rejected. The override path does not require provider credentials.
+Moved to [Compaction reference](/compaction/reference#session_before_compact).
 
 ### `session_compact`
 
-After persistence, Atomic emits an observe-only event:
-
-```typescript
-pi.on("session_compact", async (event) => {
-  console.log(event.result.rung, event.result.stats);   // rung: "planned" | "extension" | "fresh"
-  console.log(event.result.plannerModel);               // set only when a fallback model was borrowed
-  console.log(event.compactionEntry.details.strategy);  // "verbatim-lines"
-  console.log(event.fromExtension);
-});
-```
-
-Observer errors are isolated and cannot roll back the already-persisted boundary.
+Moved to [Compaction reference](/compaction/reference#session_compact).
 
 ### `session_compact_failed`
 
-A failed or cancelled manual, threshold, or overflow compaction emits an observe-only failure event:
-
-```typescript
-pi.on("session_compact_failed", async (event) => {
-  console.log(event.reason, event.errorMessage);
-  console.log(event.aborted, event.willRetry, event.fromExtension);
-});
-```
-
-`errorMessage` is absent for cancellation. `fromExtension` identifies failures after a `session_before_compact` handler supplied replacement text; no compaction boundary is persisted.
+Moved to [Compaction reference](/compaction/reference#session_compact_failed).
 
 ## Branch Summarization
 
@@ -393,92 +314,33 @@ See [`collectEntriesForBranchSummary()`](https://github.com/bastani-inc/atomic/b
 
 ## Branch Summary Format
 
-Branch summarization uses a structured format:
-
-```markdown
-## Goal
-[What the user is trying to accomplish]
-
-## Constraints & Preferences
-- [Requirements mentioned by user]
-
-## Progress
-### Done
-- [x] [Completed tasks]
-
-### In Progress
-- [ ] [Current work]
-
-### Blocked
-- [Issues, if any]
+Moved to [Compaction reference](/compaction/reference#branch-summary-format).
 
-## Key Decisions
-- **[Decision]**: [Rationale]
-
-## Next Steps
-1. [What should happen next]
-
-## Critical Context
-- [Data needed to continue]
+### Message Serialization for Branch Summaries
 
-<read-files>
-path/to/file1.ts
-path/to/file2.ts
-</read-files>
+Moved to [Compaction reference](/compaction/reference#message-serialization-for-branch-summaries).
 
-<modified-files>
-path/to/changed.ts
-</modified-files>
-```
-
-### Message Serialization for Branch Summaries
+## Extension Hooks for Branch Summarization
 
-Before branch summarization, messages are serialized to text via [`serializeConversation()`](https://github.com/bastani-inc/atomic/blob/main/packages/coding-agent/src/core/compaction/utils.ts):
+Moved to [Compaction reference](/compaction/reference#extension-hooks-for-branch-summarization).
 
-```text
-[User]: What they said
-[Assistant thinking]: Internal reasoning
-[Assistant]: Response text
-[Assistant tool calls]: read(path="foo.ts"); edit(path="bar.ts", ...)
-[Tool result]: Output from tool
-```
+### session_before_tree
 
-This prevents the model from treating it as a conversation to continue.
+Moved to [Compaction reference](/compaction/reference#session_before_tree).
 
-Tool results are truncated to 2000 characters during serialization. Content beyond that limit is replaced with a marker indicating how many characters were truncated.
+## Branch Summaries
 
-## Extension Hooks for Branch Summarization
+When `/tree` switches away from one branch to another, Atomic can summarize the abandoned branch and attach that summary at the new position. This preserves important context from the path you left without replaying the whole branch.
 
-### session_before_tree
+When prompted, choose one of:
 
-Fired before `/tree` navigation. Always fires regardless of whether user chose to summarize. Can cancel navigation or provide custom summary.
+1. no summary
+2. summarize with the default prompt
+3. summarize with custom focus instructions
 
-```typescript
-pi.on("session_before_tree", async (event, ctx) => {
-  const { preparation, signal } = event;
-
-  // preparation.targetId - where we're navigating to
-  // preparation.oldLeafId - current position (being abandoned)
-  // preparation.commonAncestorId - shared ancestor
-  // preparation.entriesToSummarize - entries that would be summarized
-  // preparation.userWantsSummary - whether user chose to summarize
-
-  // Cancel navigation entirely:
-  return { cancel: true };
-
-  // Provide custom summary (only used if userWantsSummary is true):
-  if (preparation.userWantsSummary) {
-    return {
-      summary: {
-        summary: "Your summary...",
-        details: { /* custom data */ },
-      }
-    };
-  }
-});
-```
+Branch summaries are separate from `/compact`: branch navigation can generate summary prose (optionally with focus instructions), while Verbatim Compaction lets a model select numbered line ranges and reconstructs retained text mechanically.
 
-See `SessionBeforeTreeEvent` and `TreePreparation` in the types file.
+See [Compaction](/compaction) for Verbatim Compaction, branch summarization internals, and extension hooks.
 
 ## Summary request isolation
 
@@ -488,31 +350,8 @@ Verbatim planning and branch summarization are standalone provider requests. Eac
 
 ## Settings
 
-Configure compaction in `~/.atomic/agent/settings.json` or `<project-dir>/.atomic/settings.json` (legacy `.pi` paths are also supported):
-
-```json
-{
-  "compaction": {
-    "enabled": true,
-    "reserveTokens": 16384
-  }
-}
-```
-
-| Setting | Default | Description |
-|---------|---------|-------------|
-| `enabled` | `true` | Enable automatic Verbatim Compaction. |
-| `reserveTokens` | `16384` | Tokens to reserve for the next LLM response; threshold auto-compaction starts when completed-response usage or a prospective post-tool context exceeds the model's effective input budget minus this reserve. It is an **input-side** reserve only and never caps planner output. |
-
-Compaction has no configuration key of its own for fallback borrowing: it reuses `settings.fallbackModels`, the same ordered `provider/model[:thinkingLevel]` list that main-chat model fallback walks. With no `fallbackModels` configured, compaction behaves as before: one planner model, then either an honest failure (recoverable) or a fresh context window (load-bearing).
-
-Disable auto-compaction with `"enabled": false`. You can still compact manually with `/compact`.
+Moved to [Compaction reference](/compaction/reference#settings).
 
 ## Historical formats
 
-Two old formats remain parseable but inactive:
-
-- `type:"context_compaction"` records store logical entry/content-block deletion targets from older versions. Those records are inert, so content they once hid can re-enter context when an old session resumes.
-- `type:"compaction"` without `details.strategy: "verbatim-lines"` stored generated summary prose. Those records also remain inert.
-
-Both are distinguished from active boundaries by the discriminated `details` on the shared `CompactionEntry` shape; the session format version is the same for all of them.
+Moved to [Compaction reference](/compaction/reference#historical-formats).