[ranxianglei]

DCP Plugin: Comprehensive Bug Fix — Persistence, Token Heuristics, and Compression Stability

This PR fixes 18 bugs across the Dynamic Context Pruning plugin that collectively rendered DCP unreliable in production. The issues span three failure domains: persistence (compression state lost across restarts and compaction), token measurement (inflated counts causing false warnings on non-Anthropic providers), and compression block lifecycle (blocks deactivating spuriously, summaries silently dropped, re-compression impossible). Together, these fixes make DCP's compression behavior stable across long sessions, restarts, and multi-provider setups.

Key themes:

• State that should survive restarts (messageIds, compression blocks, lastCompaction) now does.
• Token usage calculation is now provider-aware instead of Anthropic-only.
• Compression blocks persist through opencode compaction and message windowing — they are only deactivated by explicit user action.

---

Test Environment

Item	Value
Models tested	Claude Sonnet 4, GLM-5.1 (204K context)
Session types	Fresh start, restart mid-session, post-compaction, long-running (50+ turns)
Providers	Anthropic, ZhipuAI
DCP modes	message and replace

---

🔴 CRITICAL Fixes

Bug	Files	What Happens	Root Cause	Fix
#3	`sync.ts`	Compression blocks deactivate at random over time. Summaries vanish, original messages return to full context.	`syncCompressionBlocks()` checks if `compressMessageId` (the assistant msg containing the compress tool call) still exists in the current message list. When opencode truncates/removes old assistant messages — normal behavior — the ID disappears and the block is deactivated, even though the compression data is intact.	Removed the `compressMessageId` presence check entirely. Blocks are only deactivated by explicit user action (`deactivatedByUser`) or if the anchor message is confirmed absent from both current messages AND persisted data (see Bug #19).
#6	`prune.ts`	Messages silently disappear from context — compressed range is pruned but no summary is injected. Complete data loss.	`filterCompressedRanges()` calls `getLastUserMessage()` to find a user message to attach the synthetic summary to. If no user message exists at or before the anchor (conversation start, post-compaction), it returns `null`. The warning branch logs but does not inject a summary, while the subsequent code still prunes the range. Messages removed, no summary → silent data loss.	Fallback: when `getLastUserMessage()` returns null, use the anchor message itself as the base for `createSyntheticUserMessage()`. Summary is always injected.
#17	`inject/utils.ts`, `token-utils.ts`	DCP reports context at 105% (triggers CRITICAL WARNING) when actual usage is ~36%. Forces constant, premature compression on non-Anthropic providers.	`getCurrentTokenUsage()` returns `input + output + reasoning + cacheRead + cacheWrite`. For Anthropic, `input` includes cached tokens so adding `cacheRead` double-counts a small amount. For other providers, `input` is only new non-cached tokens while `cacheRead` is the cumulative cached count ≈ actual context size. Adding them produces roughly 2× the real usage.	Provider-aware heuristic: if `cacheRead > input && input > 0`, return `cacheRead` (non-Anthropic: `cacheRead ≈ actual context size`). Otherwise return `input + output + reasoning` (Anthropic: `input` already includes the full context). See Design Decisions section.

---

🟠 HIGH Fixes

Bug	Files	What Happens	Root Cause	Fix
#1	`persistence.ts`, `state.ts`	After restarting opencode, compress fails with `"startId m0043 is not available"`. Compression completely broken post-restart.	`saveSessionState()` only persists `prune`, `nudges`, `stats` — NOT `messageIds`. On restart, `assignMessageRefs()` starts from `m0001`. But the model references old mNNNN IDs from block summaries still in context. `resolveBoundaryIds()` can't find them.	Persist `messageIds` (`byRawId`, `byRef`, `nextRef`) and `lastCompaction` to the session JSON. Restore on load.
#2	`utils.ts`	When opencode's internal compaction runs, all DCP compression blocks vanish. Previously compressed context returns to full size.	`resetOnCompaction()` wipes everything including `prune.messages` (ALL compression blocks). Clearing it destroys all DCP compression summaries — effectively undoing all compression work.	`resetOnCompaction()` now only resets transient state. `prune.messages` (compression blocks) is preserved through compaction.
#7	`hooks.ts`	In message mode, compressed summaries accumulate but can never be re-compressed. Context fills with uncompressible summaries.	Synthetic summary messages are created AFTER the `assignMessageRefs` pass. They never get mNNNN IDs. In message mode, block IDs are replaced with "BLOCKED" — so the summary has no mNNNN AND no usable block ID. Completely unreferenceable.	Added a second `assignMessageRefs` pass after the prune step.
#18	`hooks.ts`	Percentage thresholds fire at wrong levels. A 45% config triggers at ~17% actual context.	`createSystemPromptHandler()` sets `modelContextLimit = computeInputBudget(limit)`. For models with context=204800 and output=131072: inputBudget = 73,728. So 45% of 73,728 = 33,178 tokens (17% of actual 204K). Related to #522.	Use `input.model.limit.context` directly instead of computed inputBudget.
#19	`sync.ts`	Compression blocks deactivate when anchor messages fall outside opencode's current message window.	`syncCompressionBlocks()` checks `messageIds.has(block.anchorMessageId)` where `messageIds` only contains IDs from the current API response. Historical anchors not in the window → false deactivation.	Secondary check against DCP's persisted `byMessageId` index. If anchor exists there, keep block active.

---

🟡 MEDIUM Fixes

Bug	Files	What Happens	Root Cause	Fix
#5	`inject/utils.ts`	DCP forces compression CRITICAL WARNING even at low context (e.g., 36%).	`overMinLimit` defaults to `true` when `modelContextLimit` is undefined. "Always over limit" when the limit can't be calculated.	Default to `false` — fail open, not closed.
#8	`token-utils.ts`	Context nudges never fire on the first turn or after full compaction.	`getCurrentTokenUsage()` returns 0 when no assistant message has `output > 0`. `0 > maxLimit = false`.	Fallback estimation from message content length.
#9	`persistence.ts`	Compress tool reports success, but on restart the compression is gone.	`saveSessionState()` wraps write in try/catch that only logs without re-throwing. Model sees no error.	Removed try/catch — errors propagate to callers.
#10	`hooks.ts`	After crash/restart, previously deactivated blocks become active again.	`syncCompressionBlocks()` deactivates blocks but no `saveSessionState()` follows. Next save only on nudge/compress.	Save state immediately after sync if blocks were deactivated.
#11	`decompress.ts`	User decompresses a block, sees "Also restored nested compression A", but A stays compressed.	Target block gets `deactivatedByUser=true` but consumed inner blocks don't. During sync, inner blocks pass the guard and get re-activated.	Recursively mark all consumed inner blocks as `deactivatedByUser=true`.
#20	`inject/utils.ts`	After compressing, model immediately gets another CRITICAL WARNING and compresses again in a loop.	API response token count doesn't immediately reflect reduced context. DCP still sees `overMax=true`.	Cooldown: if last 3 messages contain a compress tool call, suppress `overMax`.
#21	`compress-range.ts`	Model uses old mNNNN IDs from block summaries/nudges → "startId not available".	Tool description wasn't strong enough about using only visible IDs.	Strengthened prompt with explicit stale-ID warnings.

---

🟢 LOW Fixes

Bug	Files	What Happens	Root Cause	Fix
#12	`utils.ts`	`isMessageCompacted()` and `getCurrentTokenUsage()` disagree on compaction boundary.	Strict `<` vs `<=` comparison for the compaction summary message.	Added `summary === true` check to `isMessageCompacted()`.
#13	`utils.ts`	Stale `activeBlockIds` loaded from persistence for deactivated blocks.	Dual loading from persistence and `blocksById` without clearing.	Skip persisted loading — derive from `blocksById` only.
#14	`reasoning-strip.ts`	Metadata stripped from all assistant messages when `modelID`/`providerID` undefined.	`undefined === "id"` is always `false`.	Added undefined guard.

---

Design Decisions

1. Provider-Aware Token Heuristic (Bug #17)

Different providers report token counts with different semantics:

Field	Anthropic	Other providers (e.g. GLM-5.1)
`input`	Full context tokens (includes cached)	Only new non-cached tokens
`cacheRead`	Subset of `input` (redundant)	Cumulative cached tokens ≈ actual context size

The heuristic: if `cacheRead > input && input > 0`, the provider uses split reporting where `cacheRead` is the better proxy. Otherwise `input` already contains the full context. This is deliberately simple and observable.

2. Anchor Persistence Strategy (Bugs #1, #19)

DCP now maintains its own persisted `byMessageId` index, decoupling block validity from opencode's message windowing:

• Current messages → primary check (fast, in-memory)
• Persisted index → fallback (covers historical anchors)
• Deactivation → only via explicit user action or confirmed absence from both sources

3. Compaction Coexistence (Bug #2)

`resetOnCompaction()` now preserves `prune.messages` (compression blocks) while resetting transient state. After compaction, DCP reassigns refs and re-syncs blocks — but compression summaries survive.

---

Files Changed

File	Bugs	Nature of Change
`sync.ts`	#3, #10, #19	Remove compressMessageId check, add persisted anchor fallback, save on deactivation
`persistence.ts`	#1, #9, #13	Persist messageIds/lastCompaction, remove error swallowing, fix activeBlockIds loading
`prune.ts`	#6	Fallback to anchor message when no preceding user message
`hooks.ts`	#7, #18	Second assignMessageRefs pass, use raw context limit
`inject/utils.ts`	#5, #20	overMinLimit default, compress cooldown
`token-utils.ts`	#8, #17	Provider-aware token heuristic, token fallback estimation
`utils.ts`	#2, #12	Preserve prune.messages, fix isMessageCompacted boundary
`decompress.ts`	#11	Recursive deactivatedByUser for consumed blocks
`reasoning-strip.ts`	#14	Undefined guard on modelID/providerID
`compress-range.ts`	#21	Strengthened compress tool prompt

---

Notes for Reviewers

• `[FIX Bug N]` markers: Each code change is annotated with the bug number. Search for `[FIX Bug` to navigate.
• Debug logs removed: Temporary diagnostic statements have been removed.
• No new dependencies: All fixes are logic changes within existing code.
• Migration: Existing session files without `messageIds` in persistence are handled gracefully — fields default to empty and are populated on the next message transform.

Reproduction Recipes (for QA)

Bug #1: Start session → compress messages → restart opencode → attempt compress → `startId not available`
Bug #3: Start session → compress → continue 20+ turns → observe blocks deactivating
Bug #6: Compress messages near conversation start (first 5 messages) → observe messages vanish with no summary
Bug #17: Use non-Anthropic provider → observe CRITICAL WARNING at low context usage
Bug #18: Use model with large output budget (204K context, 131K output) → observe 45% threshold firing at ~17%
Bug #20: Compress at high context → observe model immediately re-compressing

[Tarquinen]

Thank you for your work, but I can't really review a PR that fixes 20 bugs in 1 go, this needs to be many much smaller PRs. There's a lot of good stuff here I think but its very hard for me to replicate bugs or check how your fixes work with a branch that has so many changes.

Please consider breaking this down into much more easily digestible PRs, and make sure to verify all of the AI code is good, i'm seeing things like sdk v1 types used and your code/comments don't line up with descriptions in this PR. Also note that when you #n in github you're referencing this PR to previous PRs/issues.

[ranxianglei]

Do as you please, I won't disassemble it anymore. These bugs are ones I actually discovered and confirmed through long-term operation; failing to fix them would render the system inoperable. This plugin is excellent, but I simply don't have the energy to invest more in the community. I will maintain my own version. Finally, thank you very much for your plugin and the inspiration it provided.