Files

T

can1357 f40e528073 fix(agent/compaction): skipped truncating tiny tool outputs during pruning

- Added a 50-token minimum floor constant for pruning decisions.
- Allowed non-superseded, non-useless tool results below that floor to avoid truncation.

2026-06-15 12:45:21 +02:00

25 KiB

Raw Blame History

Compaction and Branch Summaries

Compaction and branch summaries are the two mechanisms that keep long sessions usable without losing prior work context.

Compaction rewrites old history into a summary on the current branch.
Branch summary captures abandoned branch context during /tree navigation.

Both are persisted as session entries and converted back into user-context messages when rebuilding LLM input.

Key implementation files

packages/agent/src/compaction/compaction.ts (context-full summarization and handoff generation)
packages/snapcompact/src/snapcompact.ts (snapcompact strategy: history archived as dense bitmap images)
packages/agent/src/compaction/branch-summarization.ts
packages/agent/src/compaction/pruning.ts
packages/agent/src/compaction/utils.ts
packages/agent/src/compaction/openai.ts
packages/coding-agent/src/session/session-manager.ts
packages/coding-agent/src/session/agent-session.ts
packages/coding-agent/src/session/messages.ts
packages/coding-agent/src/extensibility/hooks/types.ts
packages/coding-agent/src/config/settings-schema.ts

Session entry model

Compaction and branch summaries are first-class session entries, not plain assistant/user messages.

CompactionEntry
- type: "compaction"
- summary, optional shortSummary
- firstKeptEntryId (compaction boundary)
- tokensBefore
- optional details, preserveData, fromExtension
BranchSummaryEntry
- type: "branch_summary"
- fromId, summary
- optional details, fromExtension

When context is rebuilt (buildSessionContext):

Latest compaction on the active path is converted to one compactionSummary message.
Kept entries from firstKeptEntryId to the compaction point are re-included.
Later entries on the path are appended.
branch_summary entries are converted to branchSummary messages.
custom_message entries are converted to custom messages.

Those custom roles are then transformed into LLM-facing messages in convertToLlm(): compactionSummary and branchSummary become user messages rendered through the static templates

packages/agent/src/compaction/prompts/compaction-summary-context.md
packages/agent/src/compaction/prompts/branch-summary-context.md

while custom messages pass through as developer messages with their raw content (no template).

Compaction pipeline

Triggers

Compaction/context maintenance can run in five ways:

Manual context compaction: /compact [instructions] calls AgentSession.compact(...).
Automatic overflow recovery: after a same-model assistant error that matches context overflow.
Automatic incomplete-output recovery: after a same-model assistant message ends with stopReason === "length" (OpenAI/Codex response.incomplete).
Automatic threshold maintenance: after a successful turn when context exceeds the resolved threshold.
Idle maintenance: runIdleCompaction() can invoke the same auto-maintenance path with reason "idle".

Compaction shape (visual)

Before compaction:

  entry:  0     1     2     3      4     5     6      7      8     9
        ┌─────┬─────┬─────┬──────┬─────┬─────┬──────┬──────┬─────┬──────┐
        │ hdr │ usr │ ass │ tool │ usr │ ass │ tool │ tool │ ass │ tool │
        └─────┴─────┴─────┴──────┴─────┴─────┴──────┴──────┴─────┴──────┘
                └────────┬───────┘ └──────────────┬──────────────┘
               messagesToSummarize            kept messages
                                   ↑
                          firstKeptEntryId (entry 4)

After compaction (new entry appended):

  entry:  0     1     2     3      4     5     6      7      8     9      10
        ┌─────┬─────┬─────┬──────┬─────┬─────┬──────┬──────┬─────┬──────┬─────┐
        │ hdr │ usr │ ass │ tool │ usr │ ass │ tool │ tool │ ass │ tool │ cmp │
        └─────┴─────┴─────┴──────┴─────┴─────┴──────┴──────┴─────┴──────┴─────┘
               └──────────┬──────┘ └──────────────────────┬───────────────────┘
                 not sent to LLM                    sent to LLM
                                                         ↑
                                              starts from firstKeptEntryId

What the LLM sees:

  ┌────────┬─────────┬─────┬─────┬──────┬──────┬─────┬──────┐
  │ system │ summary │ usr │ ass │ tool │ tool │ ass │ tool │
  └────────┴─────────┴─────┴─────┴──────┴──────┴─────┴──────┘
       ↑         ↑      └─────────────────┬────────────────┘
    prompt   from cmp          messages from firstKeptEntryId

Overflow/incomplete recovery vs threshold/idle maintenance

The automatic paths are intentionally different:

Overflow recovery
- Trigger: current-model assistant error is detected as context overflow and the error is not older than the latest compaction.
- The failing assistant error message is removed from active agent state before retry.
- Context promotion is tried first; if a configured larger model is available, the agent switches model and retries without compacting.
- If promotion is unavailable and compaction is enabled, context-full compaction runs with reason: "overflow" and willRetry: true; handoff strategy is not used for overflow because the handoff request would reuse the overflowing input.
- On success, agent.continue() is scheduled to retry the turn.
Incomplete-output recovery
- Trigger: same-model assistant message ends with stopReason === "length" and the message is not older than the latest compaction.
- The incomplete assistant message is removed from active agent state before recovery.
- Context promotion is tried first.
- If promotion is unavailable and compaction is enabled, auto maintenance runs with reason: "incomplete" and willRetry: true.
- Unlike overflow, compaction.strategy: "handoff" is allowed for incomplete-output recovery because the input context is still usable.
- On context-full success, agent.continue() is scheduled to retry the turn.
Threshold maintenance
- Trigger: successful, non-error assistant message whose adjusted context tokens exceed resolveThresholdTokens(...).
- Tool-output pruning can reduce the measured token count before threshold comparison.
- Context promotion is tried before compaction.
- If promotion is unavailable, auto maintenance runs with reason: "threshold" and willRetry: false.
- With compaction.strategy: "handoff", threshold maintenance normally schedules a post-prompt auto-handoff task instead of writing a compaction entry; pre-prompt checks run it inline to avoid racing the next turn. If handoff returns no document without aborting, it falls back to context-full compaction.
- On success, if compaction.autoContinue !== false, schedules an agent-authored developer auto-continue prompt from prompts/system/auto-continue.md.
Idle maintenance
- Trigger: runIdleCompaction() when not streaming or already compacting.
- Uses reason: "idle" and does not auto-continue afterward.

Snapcompact strategy

compaction.strategy: "snapcompact" replaces the LLM summarization call with a local, deterministic archival pass (compact from @oh-my-pi/snapcompact):

The discarded history is serialized, whitespace-collapsed, and printed onto model-aware PNG frames (frame width fixed per shape; frame height hugs the rows actually printed) using bundled public-domain pixel fonts. The shape — and frame size — resolve from the model id when the model line was measured: Claude reads X.org 6x12 glyphs with dimmed stopwords (6x12-dim; high-res lines — Opus 4.7+, Fable, Mythos — get 1932px frames under Anthropic's 4,784 visual-token cap, older lines stay at 1568px), Gemini reads two word-wrapped columns of 8x13 glyphs with sentence-hue ink and dimmed stopwords (doc-8on16-sent-dim at 2048px — Gemini 3.x bills a fixed 1,120-token budget per image at any pixel size), GPT/Kimi/GLM read 8x13 glyphs on a 16px pitch (8on16-bw at 1568px — patch billing is area-proportional, and kimi's processor downscales past 1792px). A Claude routed through Vertex or OpenRouter keeps its Claude shape. Unmeasured models fall back to their wire API family (Anthropic-family/unknown → 6x12-dim, Google → doc-8on16-sent-dim, OpenAI-compatible → 8on16-bw); billing (per-family patch/budget formulas, OpenAI's detail: "original" hint) always follows the API carrying the request, computed for the resolved frame size. The snapcompact.shape setting (default auto) forces one of the research-eval variants instead: square grids (8x8r/8x8u/6x6u/5x8 × sentence-hue/black ink) or the per-model eval winners (6x12-dim, 8x13-bw, 8on16-bw, and the two-column word-wrapped doc-8on16-bw/-sent/-sent-dim, where dim prints stopwords in gray). A forced variant keeps its geometry but is re-priced for the target provider's image billing. The same setting governs inline system-prompt/tool-result imaging (snapcompact.systemPrompt, snapcompact.toolResults).
Serialization keeps the archive conversation-dense: tool results are truncated head+tail (default 2,000 chars at a 0.6 head ratio), tool-call argument values are capped per value (500) and per call (2,000), and tool output is printed in dim gray ink so conversation reads louder than tool noise. All budgets and the dimming are configurable via SerializeOptions (toolResultMaxChars, toolArgMaxChars, toolCallMaxChars, truncateHeadRatio, dimToolResults).
Frames persist under CompactionEntry.preserveData.snapcompact and are re-attached to the compactionSummary message as image blocks on every context rebuild; the entry's summary is a deterministic reading guide (grid geometry, role tags, truncation notes) plus the usual file-operation lists.
Later compactions carry earlier frames forward. The frame budget is provider-aware (providerFrameBudget): the per-provider image cap clamped to 8 (MAX_FRAMES) — OpenRouter hard-caps requests at 8 images and silently drops the excess, unknown providers get a safe floor of 5. Beyond the budget the archive fades from the middle out: the earliest frame (session head — the original request, or the filmed summary of older history) is pinned, and the oldest unpinned frames are evicted. Pages of the current compaction that no longer fit are never rendered or dropped — the newest unframed slice survives verbatim as a text tail on the summary (Archive.textTail, capped at two frame capacities with middle elision) and is folded back into frames by the next compaction. If the previous compaction was text-based, its summary is printed at the head of the frame archive as [Summary of earlier history].
No model, API key, or network is involved, so snapcompact is also safe for overflow recovery. It requires a vision-capable current model (model.input includes "image"); otherwise the run falls back to context-full and emits a warning notice (auto and manual paths). Manual /compact honors the strategy unless custom instructions are given (those imply a directed LLM summary).
Rationale: the shape table comes from the snapcompact 200k-token evals in packages/snapcompact, where bitmap frames preserved QA recall at lower billed-token cost than raw text for vision-capable models.

Display transcript

Compaction no longer visually restarts the conversation. The TUI renders the display transcript (buildSessionContext({ transcript: true }) / AgentSession.buildTranscriptSessionContext()): every path entry in chronological order, with each compaction shown inline as a slim divider — ── 📷 compacted · ctrl+o ── — at the point it fired. Expanding (ctrl+o) reveals the summary. Only the LLM context resets at the compaction boundary; the scrollback above the divider stays intact, including across session resume.

Pre-compaction pruning

Before compaction checks, tool-result pruning may run (pruneToolOutputs).

Default prune policy:

Protect newest 40_000 tool-output tokens.
Require at least 20_000 total estimated savings.
Never blank a result below 50 tokens (MIN_PRUNE_TOKENS): the [Output truncated - N tokens] placeholder costs ~8 tokens, so pruning a sub-floor result would grow the context and churn the prompt cache for nothing. (Superseded and useless results keep their own rules — the useless collector already drops no-savings candidates; superseded reads prune for correctness regardless of size.)
Never prune skill tool results, read results of skill:// paths, or reads of the active plan reference file (added via AgentSession's plan protection).

Pruned tool results are replaced with:

[Output truncated - N tokens]

If pruning changes entries, session storage is rewritten and agent message state is refreshed before compaction decisions.

Useless-result elision

Tools can flag a finished result as contextually useless — a search with zero matches, a job poll that timed out with everything still running, an empty irc inbox drain. The flag originates on the tool result (AgentToolResult.useless, set via ToolResultBuilder.useless() or directly on the returned object), is copied by the agent loop onto the persisted ToolResultMessage (never together with isError — errors always win), and is consumed in three places:

Per-turn stale-result pass (pruneSupersededToolResults, gated by compaction.dropUseless, default on): flagged results are blanked to the exact placeholder [Uneventful result elided] (USELESS_NOTICE) with the same cache-aware timing as superseded reads — only when the suffix after the candidate is small (≤ ~8k tokens) or the session has idled past the provider prompt-cache lifetime. Results smaller than the notice itself are never blanked (no savings), and protected tools are exempt.
Threshold prune (pruneToolOutputs): flagged results bypass the protect-recent window, same as superseded reads, and receive USELESS_NOTICE instead of the token-count placeholder.
Summary serialization: serializeConversation (agent and snapcompact) drops the whole tool call/result pair from summarizer/archive input — the source region is discarded after summarization anyway, so the exclusion costs no cache.

The flag never reaches provider wire formats, and flagged pairs are never removed from history (only blanked in place), so tool-call/result pairing and provider-native history replay stay intact.

Boundary and cut-point logic

prepareCompaction() only considers entries since the last compaction entry (if any).

Find previous compaction index.
Compute boundaryStart = prevCompactionIndex + 1.
Adapt keepRecentTokens using measured usage ratio when available.
Run findCutPoint() over the boundary window.

Valid cut points include:

message entries with roles: user, assistant, bashExecution, hookMessage, branchSummary, compactionSummary
custom_message entries
branch_summary entries

Hard rule: never cut at toolResult.

If there are non-message metadata entries immediately before the cut point (model_change, thinking_level_change, labels, etc.), they are pulled into the kept region by moving cut index backward until a message or compaction boundary is hit.

Split-turn handling

If cut point is not at a user-turn start, compaction treats it as a split turn.

Turn start detection treats these as user-turn boundaries:

message.role === "user"
message.role === "bashExecution"
custom_message entry
branch_summary entry

Split-turn compaction generates two summaries:

History summary (messagesToSummarize)
Turn-prefix summary (turnPrefixMessages)

Final stored summary is merged as:

<history summary>

---

**Turn Context (split turn):**

<turn prefix summary>

Summary generation

compact(...) builds summaries from serialized conversation text:

Convert messages via convertToLlm().
Serialize with serializeConversation().
Wrap in <conversation>...</conversation>.
Optionally include <previous-summary>...</previous-summary>.
Optionally inject extension hook context and active memory-backend compaction context as <additional-context> entries.
Execute summarization prompt with SUMMARIZATION_SYSTEM_PROMPT.

Prompt selection:

first compaction: compaction-summary.md
iterative compaction with prior summary: compaction-update-summary.md
split-turn second pass: compaction-turn-prefix.md
short UI summary: compaction-short-summary.md
handoff document: handoff-document.md (used by generateHandoff(...), not serialized compaction)

Remote summarization modes:

If compaction.remoteEndpoint is set and remote compaction is enabled, local summary generation POSTs:
- { systemPrompt, prompt }
Expects JSON containing at least { summary }.
For OpenAI/OpenAI Codex models, compaction first tries the provider-native /responses/compact endpoint when remote compaction is enabled. It preserves provider replacement history in preserveData.openaiRemoteCompaction and falls back to local summarization if that native request fails.

Handoff generation

packages/agent/src/compaction/compaction.ts also exports generateHandoff(...). Handoff generation uses the same completeSimple(...) oneshot style as summarization, but it preserves the live agent cache prefix by sending the active system prompt, tool array, and real LLM message history, then appending one agent-attributed user message containing the handoff prompt. It forces toolChoice: "none" and returns joined text blocks directly.

Handoff does not write a CompactionEntry. AgentSession.handoff() owns the session transition: it starts a new session, injects the generated document as a visible custom_message with customType: "handoff", and rebuilds agent messages from that new session.

File-operation context in summaries

Compaction tracks cumulative file activity using assistant tool calls:

read(path) → read set
write(path) → modified set
edit(path) → modified set

Cumulative behavior:

Includes prior compaction details only when prior entry is pi-generated (fromExtension !== true).
In split turns, includes turn-prefix file ops too.
details.readFiles excludes files also modified; details.modifiedFiles carries the rest (persisted shape is unchanged).

Summary text gets one <files> tag appended via prompt template: a grouped, prefix-folded directory tree (find-tool shape) with a per-file access marker — (Read) for read-only files, (Write) for modified files never read, (RW) for modified files also present in the cumulative read set. Capped at 20 files with an … (N more files omitted) line.

<files>
# packages/agent/src/compaction/
compaction.ts (Read)
utils.ts (RW)
## prompts/
file-operations.md (Write)
</files>

Legacy <read-files>/<modified-files> tags from summaries written by earlier versions are stripped (alongside <files>) before re-appending, so old summaries self-heal on the next compaction.

Persist and reload

After summary generation (or hook-provided summary), agent session:

Appends CompactionEntry with appendCompaction(...) for context-full maintenance; handoff strategy creates a new session and injects a handoff custom_message instead.
Rebuilds display context from the active leaf via buildDisplaySessionContext().
Replaces live agent messages with rebuilt context.
Synchronizes active todo phases from the rebuilt branch and closes provider sessions whose history was rewritten.
Emits session_compact hook event.

Branch summarization pipeline

Branch summarization is tied to tree navigation, not token overflow.

Trigger

During navigateTree(...):

Compute abandoned entries from old leaf to common ancestor using collectEntriesForBranchSummary(...).
If caller requested summary (options.summarize), generate summary before switching leaf.
If summary exists, attach it at the navigation target using branchWithSummary(...).

Operationally this is commonly driven by /tree flow when branchSummary.enabled is enabled.

Branch switch shape (visual)

Tree before navigation:

         ┌─ B ─ C ─ D (old leaf, being abandoned)
    A ───┤
         └─ E ─ F (target)

Common ancestor: A
Entries to summarize: B, C, D

After navigation with summary:

         ┌─ B ─ C ─ D ─ [summary of B,C,D]
    A ───┤
         └─ E ─ F (new leaf)

Preparation and token budget

generateBranchSummary(...) computes budget as:

tokenBudget = model.contextWindow - branchSummary.reserveTokens

prepareBranchEntries(...) then:

First pass: collect cumulative file ops from all summarized entries, including prior pi-generated branch_summary details.
Second pass: walk newest → oldest, adding messages until token budget is reached.
Prefer preserving recent context.
May still include large summary entries near budget edge for continuity.

Compaction entries are included as messages (compactionSummary) during branch summarization input.

Summary generation and persistence

Branch summarization:

Converts and serializes selected messages.
Wraps in <conversation>.
Uses custom instructions if supplied, otherwise branch-summary.md.
Calls summarization model with SUMMARIZATION_SYSTEM_PROMPT.
Prepends branch-summary-preamble.md.
Appends file-operation tags.

Result is stored as BranchSummaryEntry with optional details (readFiles, modifiedFiles).

Extension and hook touchpoints

`session_before_compact`

Pre-compaction hook.

Can:

cancel compaction ({ cancel: true })
provide full custom compaction payload ({ compaction: CompactionResult })

`session.compacting`

Prompt/context customization hook for default compaction.

Can return:

prompt (override base summary prompt)
context (extra context lines injected into <additional-context>)
preserveData (stored on compaction entry)

`session_compact`

Post-compaction notification with saved compactionEntry and fromExtension flag.

`session_before_tree`

Runs on tree navigation before default branch summary generation.