Repository navigation
feat: expand message ref format from 4-digit to 5-digit (9999→99999) - #1
Merged
Merged
Conversation
- MESSAGE_REF_WIDTH: 4 → 5
- MESSAGE_REF_MAX_INDEX: 9999 → 99999
- MESSAGE_REF_REGEX: /^m(\d{4})$/ → /^m(\d{4,5})$/ (backward compat)
- All test files updated to 5-digit format (m00001, m00002, etc.)
- Added boundary tests for 99999 and 100000
- Kept 4-digit backward compat parse tests (m0001 still resolves)
Critical fixes: - state.ts: migrate 4-digit refs (m0001) to 5-digit (m00001) during state loading Fixes lookup failure when parseBoundaryId normalizes old refs - compress-message.ts: restore XML tag template (accidentally destroyed in prior edit) - prompts/tool descriptions: mNNNN → mNNNNN in all LLM-visible text - Error messages: mNNNN → mNNNNN in search.ts and message-utils.ts - message-ids.ts: error message range 0-99999 → 1-99999 - AGENTS.md: fix duplicate section 5.3 numbering, mNNNN → mNNNNN Tests: +4 backward compat tests (350 total, 0 failures)
ranxianglei
force-pushed
the
feat/msgid-expansion
branch
from
May 19, 2026 00:52
33bc110 to
5e54496
Compare
ranxianglei
added a commit
that referenced
this pull request
Jul 5, 2026
Review #1 🔴: dcp.schema.json had nudgeGrowthTokens:6000 in both the field default and the compress default object. If any config layer applies schema defaults, it would override resolveAdaptiveNudgeGrowth. Removed from both locations — field has no default, description documents adaptive behavior. Review #2 🔴: formatBlockAge was duplicated in status.ts and nudge.ts. Extracted to lib/ui/utils.ts as formatAge, imported in both modules.
ranxianglei
added a commit
that referenced
this pull request
Jul 5, 2026
…tool (#51) * v1.8.0: Principle-driven prompts, hybrid Tips, compress improvements * feat: nudgeGrowthTokens — token-based Tips gating (default 6K) * refactor(nudge): extract computeShouldNudge for testability + review fixes - Extract pure computeShouldNudge() helper (no state/side effects) - Add 11 unit tests covering gating logic (floor, growth step, overMaxLimit bypass) - overMaxLimit now bypasses frequency gating (urgent for small models) - Fix dcp.schema.json missing trailing newline Addresses PR #51 code review. * feat(notify): show context before→after level in compress header Replace misleading header (cumulative removed vs active summary — apples vs oranges) with a clear context transition: ▣ ACP | Context 90%→50% (180K→100K of 200K) tokensBefore captured in pipeline; tokensAfter = before - removed + summary. Per-batch Compression line unchanged (already correct). * fix(notify): drop context limit/percentage from header to avoid model anchoring Showing total + % lets the model learn its ceiling and revert to waiting until 80-90% before compressing. Now only absolute before→after is shown: ▣ ACP | Context 180K→100K The model can still see it saved ~80K, but cannot anchor to a ceiling. * feat(nudge): adaptive nudgeGrowthTokens — 5% of context limit, clamped [6K, 50K] Default nudgeGrowthTokens was 6000 (fixed), which for a 1M-context model is only 0.6% — Tips fired far too frequently, making the model feel over-pressured to compress even at low context levels (e.g. 26%). New default is adaptive: 5% of modelContextLimit, clamped to [6000, 50000]: - 10K-100K context → 6K (floor) - 128K → 6.4K, 200K → 10K, 500K → 25K, 1M → 50K - 2M+ → 50K (cap, prevents explosion for future large models) Uses absolute tokens (not percentage) so the model sees a concrete growth threshold. Users can still override via explicit nudgeGrowthTokens config. * fix(inject): gate injectContextUsage by shouldNudge — stop every-message spam injectContextUsage (buildContextUsageGuidance) was called unconditionally every turn, bypassing the graduated nudge mechanism (nudgeGrowthTokens). This meant the urgent 'Context: XK tokens. compress waste promptly.' text appeared on EVERY message regardless of context level or growth. Now it only fires when shouldNudge is true (same gate as Tips), so the context usage text respects the adaptive nudgeGrowthTokens interval. * fix(inject): gate ALL suffix content behind shouldNudge — zero compression noise between nudges blockGuidance (Compressed blocks list) and visibleIdRange were still injected unconditionally every turn, bypassing the graduated nudge mechanism. Even without explicit 'compress now' text, the model saw 'Compressed blocks: 9 (...)' every turn — a constant indirect reminder that compression is active, creating low-level pressure to compress. Now ALL suffix injections (context usage, block list, Tips, visible ID range) are gated by shouldNudge. Between nudges, the suffix is empty — the model sees zero compression-related metadata. When shouldNudge fires (every nudgeGrowthTokens), the full info appears in one consolidated burst. Updated test to reflect gated behavior (added token info to trigger shouldNudge=true). * feat: add acp_status tool + simplify suffix to compact summary Add acp_status tool (lib/compress/status.ts) for on-demand block details: block IDs, summary sizes, ages, and topics in a formatted table. Simplify the per-message suffix from detailed block list ('b1 (750t), b59 (750t), ...') to a compact one-liner: 'Compressed blocks: 9 (4.2K summary, last 5m ago). Use acp_status.' Aging warnings and consolidation suggestions remain in the suffix (triggered only when usage > 50% or > 50 blocks respectively). * feat: gate system prompt behind nudgeGrowthTokens frequency System prompt (SYSTEM constant + decompress extension + protected tools) was injected EVERY request via experimental.chat.system.transform hook, adding ~50 lines of compression philosophy per turn regardless of context level. At 26% context the model still saw 'BE FRUGAL — compress waste promptly' every turn, creating false urgency. Now gated: injectCompressNudges stores shouldInjectThisTurn in state.nudges after computing computeShouldNudge. The system prompt handler reads it on the NEXT request (1-turn lag, since system hook runs before messages hook). - shouldInjectThisTurn === undefined (first turn) → inject (initial awareness) - shouldInjectThisTurn === true → inject - shouldInjectThisTurn === false → skip entirely (model sees nothing) - manualMode always injects (needs manual trigger instructions) * fix: nudgeGrowthTokens dead code + duplicate context usage injection Code review #8 fixes: 1. resolveAdaptiveNudgeGrowth was dead code — config default (6000) made nudgeGrowthTokens always defined, so the ?? fallback never triggered. Fix: make nudgeGrowthTokens optional in config type, remove hardcoded default from DEFAULTS, simplify merge to pass-through. Now undefined when user doesn't configure it → resolveAdaptiveNudgeGrowth runs. 2. applyAnchoredNudges duplicated context usage info — it appended buildContextUsageGuidance output to contextLimitNudge, while injectContextUsage also injects the same info. Removed the duplicate from applyAnchoredNudges; injectContextUsage is now the single source. * fix: compress resetting lastNudgeTokens to 0 bypassed growth gate After compress, lastPerMessageNudgeTokens was reset to 0, which caused computeShouldNudge's lastNudgeTokens===0 sentinel to fire every turn (bypassing nudgeGrowthTokens growth check). This created a feedback loop: compress → Tip fires immediately → model compresses again → repeat. Fix: set lastPerMessageNudgeTokens to post-compress currentTokens so the growth counter restarts from the correct level. Also hardened getModelInfo to handle messages without model info (optional chaining). * test: add integration tests for post-compress nudge state transitions Unit tests of computeShouldNudge (pure function) passed but missed the real bug: injectCompressNudges set lastPerMessageNudgeTokens=0 after compress, which re-triggered the 'first turn' sentinel every turn. New integration tests verify the full state machine: - Post-compress: lastPerMessageNudgeTokens = currentTokens (not 0) - Post-compress small growth (<threshold): shouldInjectThisTurn = false - Post-compress large growth (>threshold): shouldInjectThisTurn = true * feat: add scripts/dev-deploy.sh — one-command build + deploy Generic, portable dev-deploy script for local ACP development: - Auto-detects node/npm (nvm/fnm/volta/manual installs) - Builds (tsup bundle + tsc types) + deploys to opencode plugin cache - --check flag runs tests before building - --no-build deploys existing dist/ only - No hardcoded user paths — uses $HOME and script-relative paths AGENTS.md section 3.4 updated with correct cache path, verification commands, and common-mistake warning. Fixes the stale-dist bug: deploy-acp.sh (~/scripts/) only copied dist/ without building, leading to 4 commits missing from deployed bundle. * fix: remove schema default 6000 + dedup formatAge Review #1 🔴: dcp.schema.json had nudgeGrowthTokens:6000 in both the field default and the compress default object. If any config layer applies schema defaults, it would override resolveAdaptiveNudgeGrowth. Removed from both locations — field has no default, description documents adaptive behavior. Review #2 🔴: formatBlockAge was duplicated in status.ts and nudge.ts. Extracted to lib/ui/utils.ts as formatAge, imported in both modules. * fix: review suggestions — sentinel semantics, idWidth, nullglob, acp_status tests 1. lastNudgeTokens===0 sentinel → undefined (types.ts, state.ts, utils.ts, persistence.ts, inject.ts, inject/utils.ts). Uses undefined as 'never nudged' sentinel instead of 0. Old persisted state with 0 loads safely (growth = currentTokens - 0 = huge → nudges once, then corrects). 2. shouldInjectThisTurn 1-turn lag comment in hooks.ts. 3. dev-deploy.sh: shopt -s nullglob for node PATH glob detection. 4. status.ts: idWidth computed from displayed blocks (not activeIds). 5. tests/acp-status.test.ts: 7 dedicated tests for acp_status tool. * docs: update changelog + devlog for v1.8.1 (adaptive nudgeGrowthTokens)
ranxianglei
added a commit
that referenced
this pull request
Jul 15, 2026
- notification.ts formatEntryRanges: use effectiveMessageIds.length instead of startId/endId (4th leak site) - system.ts: update stale 'message-ID ranges' description to 'message counts' - recap.ts + status.ts: singular grammar (1 msgs → 1 msg, 1 messages → 1 message) Addresses Oracle #1 (FAIL→expected PASS) and Oracle #2 (PASS) consensus findings. Ultraworked with [Sisyphus](https://github.com/code-yeongyu/oh-my-opencode) Co-authored-by: Sisyphus <clio-agent@sisyphuslabs.ai>
ranxianglei
added a commit
that referenced
this pull request
Jul 16, 2026
- notification.ts formatEntryRanges: use effectiveMessageIds.length instead of startId/endId (4th leak site) - system.ts: update stale 'message-ID ranges' description to 'message counts' - recap.ts + status.ts: singular grammar (1 msgs → 1 msg, 1 messages → 1 message) Addresses Oracle #1 (FAIL→expected PASS) and Oracle #2 (PASS) consensus findings. Ultraworked with [Sisyphus](https://github.com/code-yeongyu/oh-my-opencode) Co-authored-by: Sisyphus <clio-agent@sisyphuslabs.ai>
This was referenced Sep 3, 2026
This was referenced Sep 17, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
m0001(4-digit, max 9,999) tom00001(5-digit, max 99,999)/^m(\d{4,5})$/accepts both old 4-digit and new 5-digit refsChanges
lib/message-ids.ts:MESSAGE_REF_WIDTH4→5,MESSAGE_REF_MAX_INDEX9999→99999, regex accepts 4-5 digitsm0001→m00001format referencesAGENTS.md: added mandatory dual-agent review for code + tests, PR workflowTest Results
346/346 tests pass, typecheck clean
Review Checklist
as anyor type suppression