中文 | English
DCP folds older successful tool outputs before OpenCode serializes a model request. It operates on a request copy; the host retains ownership of stored history, tool execution, and native compaction.
The plugin resolves the current model from explicit message identity and the host's read-only provider catalog. It never guesses a session from timestamps or reuses a previous request's context window. Missing identity, model limits, or unsupported content causes an unchanged request.
The engine preserves at least the latest four complete tool execution steps and 16,000 estimated tokens of recent steps. Both conditions apply. One user request can contain many separate steps; parallel tools in one step remain protected together.
When history exceeds 70% of the conservative input budget, redundant old read outputs are folded first, followed by other eligible old successful outputs in chronological order until the target is reached or no safe candidates remain. A fold only sets the native state.time.compacted marker. The host renders [Old tool result content cleared]; tool inputs and call/result identities remain intact.
Eligible tools are known read, grep, glob, and bash with an explicit zero exit status. Errors, running tools, unknown tools, attachments, skill/task results, and instruction-bearing reads (including dynamically loaded instructions) remain protected. Direct reads of AGENTS.md, AGENTS.override.md, CLAUDE.md, CONTEXT.md, and SKILL.md are also protected without dynamic-loading metadata. Additional tools can be protected in configuration.
Redundancy priority applies only to read, grep, and glob with identical tool names, JSON-serialized complete inputs, and byte-identical complete outputs, backed by a later eligible copy that is still visible. Different read pages and changed file contents stay independent. The latest copy survives the redundancy pass, but may still be folded by the subsequent lossy budget pass. This does not activate below budget or bypass savings thresholds and recent-step protections.
User instructions, assistant text, reasoning signatures, tool inputs, errors, message/part counts, identities and ordering remain unchanged. There is no topic inference, synthetic digest, input reduction, or structural tool-call merging. Prioritizing redundant outputs does not turn multiple executions into a single call. Projection is prepared independently and committed only on success.
Host markers for ordinary file, directory, and agent references do not disable pruning. Their expanded text is counted and the markers remain intact. Already-compacted tools are estimated using the host's cleared output, even when stored history retains attachments. Media that still reaches the model and unfamiliar content keep the request unchanged rather than receiving a guessed token cost.
Folding is lossy output cleanup. Original outputs remain in stored history. Protected steps, long inputs and system instructions may themselves exceed the budget; DCP then leaves the protection rules intact and lets the host handle native compaction.
The host may start automatic compaction from the previous response's reported usage before the next pruning hook runs. A small context window or large system/tool definitions can therefore trigger a native summary before any old steps become eligible. targetRatio limits estimated history, not the final provider request, and insufficient capacity never lowers recent-step protection automatically. Successful summary continuation and settlement of running tools remain host execution contracts.
Compression does not control the agent execution state machine. DCP only handles an already-scheduled request copy. It does not pause or cancel tools, wait for stopped work to compress, trigger summaries, or restart tasks. Projection failure retains the original request and returns to the existing flow. Tool settlement, continuation after native compaction, and explicit user cancellation remain host responsibilities; host execution defects must be ruled out for the specific runtime version.
The model-facing dcp_prune tool requests one fold on the next ordinary request, subject to the same protections. It returns immediately and does not permanently change policy.
Native /compact keeps its own prompt and checkpoint behavior. DCP never calls summarize, writes session history, or changes host compaction defaults. Its compacting hook arms one skip for the following transform in that session.
No configuration file is needed. Automatic pruning and redundancy priority are enabled by default, retaining at least 4 complete recent steps and 16,000 estimated tokens, with a history budget ratio of 0.7 and minimum per-output savings of 512 tokens. These retain the verified protection and savings thresholds; redundancy priority adds no switch to tune. Regression tests keep the runtime defaults, schema, and both README examples aligned. Existing valid user settings still override defaults through the documented layers.
Add "@lexwdex-org/opencode-dcp@^6" to OpenCode's plugin array. The V1 plugin peer range is >=1.4.3 <2; the CI matrix checks minimum/latest V1 types and a pinned real-host contract separately. See Development for test setup.
The official GraphAgent 1.0.39 macOS ARM64 artifact can interrupt a running tool during automatic compaction in Native LLM mode, with DCP both enabled and disabled. That combination is unsupported for tasks requiring reliable slow-tool settlement. AI SDK mode on the same artifact passed the slow-tool and explicit-cancellation controls. A passing development-source test does not establish released-artifact compatibility; see published-host evidence for the tested scope and reproduction command.
Configuration layers are global $XDG_CONFIG_HOME/opencode/dcp.jsonc (default ~/.config/opencode/dcp.jsonc), $OPENCODE_CONFIG_DIR/dcp.jsonc, then the nearest project .opencode directory containing a DCP configuration file. Each also accepts .json, with JSONC preferred. Directories without a DCP file are skipped during the upward search; only the nearest project layer is used, without merging more distant ancestors. The plugin does not create configuration files.
protectRecentSteps is a positive integer; protectRecentTokens is a nonnegative integer; targetRatio is in (0, 1]; minimumSavingsTokens is a positive integer. All three integer settings must be at most JavaScript's maximum safe integer, 9007199254740991. Every protectedTools name must contain a non-whitespace character; these tools add protections without disabling built-in ones. Invalid JSONC rejects the entire layer; invalid or out-of-range fields retain previous valid values. autoUpdate only notifies. Debug logs contain operational metadata, not conversation dumps.
Projection statistics report all folded outputs as foldedTools; redundantTools counts the subset prioritized because a later identical read result remained visible. This is not a count of merged calls or a promise to permanently retain the last copy through further budget cleanup.
This version fixes the selection order that discarded unique evidence before redundant evidence. Safe edit-chain consolidation needs complete version snapshots; failed-retry cleanup needs root-cause and partial-side-effect evidence; distant context and completed branches need traceable task summaries. Their dependencies and acceptance criteria are tracked in #61, #62, and #63. These capabilities are not enabled. See the design review.
The old engine is removed, without a legacy mode. The retired tailTurns, lowWatermarkRatio, driftThreshold, toolOutputKeepChars, mergeRuns and commands.* settings produce migration notices and are ignored.
The /dcp fold|status commands are removed because the V1 command hook has no supported cancellation result. Use dcp_prune for manual folding and debug logs for diagnostics. Manual folding now affects one request, not every future task.
DCP no longer injects compaction.tail_turns or preserve_recent_tokens. Existing user configuration is still owned by the host. Older compress, summarize, autoPrune, manualMode, strategies, turnProtection, language and experimental settings remain retired. No topic-threshold migration is performed.
Restart OpenCode after upgrading. Stored history requires no migration.
Use npm and Node's node:test: npm test, npm run typecheck, npm run format:check, and npm run check:package.
Real-host validation requires a clean, isolated OpenCode checkout pinned to the commit in scripts/test-host.mjs. Install Bun and host dependencies following .github/workflows/pr-checks.yml, set OPENCODE_SOURCE_ROOT to that checkout, then run npm run test:host.
The development toolchain uses Node.js 26.8.1 and npm 12.0.2. Install with npm ci --no-audit --no-fund, then run npm audit --audit-level=high separately. See the upgrade record for versions, install-script permissions and upstream dependency constraints.
npm publication uses GitHub Actions trusted publishing. Ordinary branch pushes do not publish. After the reviewed version is merged, push its v<version> tag matching package.json. Manual workflow dispatch must also select that version tag. The workflow checks tag identity, formatting, types, tests, package contents, and dependency audit before publishing; do not publish the same version locally.
License: AGPL-3.0-or-later.
{ "$schema": "https://raw.githubusercontent.com/LeXwDeX/opencode-dynamic-context-pruning/master/dcp.schema.json", "enabled": true, "autoUpdate": true, "debug": false, "dtc": { "enabled": true, "protectRecentSteps": 4, "protectRecentTokens": 16000, "targetRatio": 0.7, "minimumSavingsTokens": 512, "protectedTools": [], }, "tool": { "enabled": true }, }