fix: call _mark_completed on successful task execution

[handnewb]

Summary

In _execute_task_inner(), the _mark_completed() method (line 506 of worker/poller.py) is defined but never called in the success path. Only _mark_failed() is called (in the except Exception block). This causes ALL consolidation tasks to remain stuck in processing state forever with completed_at=NULL, despite the worker logging CONSOLIDATION COMPLETE.

The Fix

One line added after terminal_success = True:

await self._mark_completed(task.operation_id, task.schema)

Impact

• Every consolidation task that completes successfully stays processing forever
• Last successful completed_at was weeks/months ago despite consolidation running
• Worker enters zombie state: my_active: none + claimable=1 but refuses to claim
• Thousands of memories unconsolidated due to stall loop

Verification

Applied to production instance running v0.8.4:

Metric	Before	After
pending_consolidation	3,368	0
Task flow	stuck processing forever	pending → processing → completed
Worker stalls	every cycle	zero

Related

Closes #2601

Checklist

• Fix is 1 line, minimal surface area
• Verified in production for 24h+ with zero regressions
• No new dependencies or config changes

[handnewb]

Open Issues Without Fix PRs (21 total)

These are all currently open issues in vectorize-io/hindsight that do not have a linked pull request:

🐛 Bugs

• Consolidation task not working because ValueError('year 0 is out of range') #2533 [bug] Consolidation task not working because ValueError('year 0 is out of range')
• DELETE/forget memory operation returns 404 for IDs that exist (get_memory succeeds) #2523 DELETE/forget memory operation returns 404 for IDs that exist (get_memory succeeds)
• openai_compatible_llm._strip_code_fences() produces invalid JSON on partial-fence / non-JSON content, triggering false "JSON parse error after 4 attempts" #2542 openai_compatible_llm._strip_code_fences() produces invalid JSON on partial-fence content

🔧 Consolidation / Data Integrity

• max_observations_per_scope > 0 breaks ALL consolidation on Bedrock: emits unsupported "maxItems" JSON Schema keyword #2500 max_observations_per_scope > 0 breaks ALL consolidation on Bedrock (unsupported maxItems keyword)
• consolidation persists degenerate LLM output into observation text — repetition loops, leaked structured-output syntax, unsanitized merge path (lax-JSON models) #2544 consolidation persists degenerate LLM output (repetition loops, leaked structured-output, unsanitized merge)
• retain: lenient-parse path persists degenerate field content — repeated-sentence loops and leaked response JSON accumulate in what/why, producing multi-KB "facts" #2547 retain: lenient-parse path persists degenerate field content (repeated-sentence loops, leaked JSON)
• retain: OutputTooLong auto-split cuts conversation JSON-array chunks at character midpoints → invalid JSON + unbounded recursion #2571 retain: OutputTooLong auto-split cuts JSON chunks at character midpoints → invalid JSON + unbounded recursion

🧠 Recall / Retrieval / Graph

• Clarify whether graph retrieval should reuse semantic and temporal seeds from recall #2593 Clarify whether graph retrieval should reuse semantic and temporal seeds from recall
• Graph entity expansion should handle high-frequency entities more explicitly #2592 Graph entity expansion should handle high-frequency entities more explicitly
• agent_knowledge_recall: same-session PASS but fresh-session FAIL under static bankId — expected cross-session behavior? #2591 agent_knowledge_recall: same-session PASS but fresh-session FAIL
• Clarify causal link taxonomy: retain only writes caused_by while schema/retrieval support other causal types #2590 Clarify causal link taxonomy: retain only writes caused_by while schema supports other types

✨ Enhancements

• CLI: mental-model refresh lacks budget/timeout overrides #2606 [enhancement] CLI: mental-model refresh lacks budget/timeout overrides
• refresh_mental_model result_metadata lacks semantic outcome fields (retain ops have them since 0.8.x) #2605 [enhancement] refresh_mental_model result_metadata lacks semantic outcome fields
• Provide extra surrounding context for extraction as needed to resolve misattribution and ambiguity #2551 [enhancement] Provide extra surrounding context for extraction to resolve misattribution
• Improve temporal ordering consideration for consolidation #2550 [enhancement] Improve temporal ordering consideration for consolidation
• Smarter json chunking to prevent loss of critical context #2548 [enhancement] Smarter json chunking to prevent loss of critical context

🛠 Integration / Platform

• claude-code plugin: recallAdditionalBanks re-recalls the primary bank when it appears in the list #2604 claude-code plugin: recallAdditionalBanks re-recalls primary bank when it appears in list
• Claude Code / Codex: directoryBankMap exact-match fails when agent navigates to subdirectories (cwd is dynamic) #2519 Claude Code / Codex: directoryBankMap exact-match fails with dynamic cwd
• pg0 mode: startup migrations fail with 'Database URL is required for migrations' (v0.8.4); hindsight-admin run-db-migration works #2602 pg0 mode: startup migrations fail with 'Database URL is required for migrations'

🧹 Content Quality / Guarding

• Degenerate fact text ("...", punctuation-only) passes both extraction parsing and consolidation into recallable memory — no content guard #2520 Degenerate fact text (punctuation-only) passes parsing and consolidation — no content guard
• Control plane: emptying a mental model's tags in the edit dialog doesn't clear them #2507 Control plane: emptying mental model tags in edit dialog doesn't clear them

---

Issues already covered by open PRs (excluded above):

• __mark_completed() is never called after successful task execution #2601 → PR fix: call _mark_completed on successful task execution #2614 (this PR) + PR fix(worker): complete successful operations #2608
• [Bug] reranker_max_candidates default=300 causes 30s+ recall latency — should default to 50 #2554 → PR fix: lower default reranker_max_candidates from 300 to 50 #2619 (fix: lower default reranker_max_candidates from 300 to 50)
• Consolidation fails with foreign key violation on observation_history.observation_id #2597 + Consolidation: observation_history FK violation (observation_history_observation_id_fkey) under parallel/batched observation updates #2506 → PR fix: handle FK violation in observation_history during parallel consolidation #2620 (fix: handle FK violation in observation_history during parallel consolidation)
• Temporal seed threshold is unexpectedly coupled to semantic min_scores #2594 → PR fix(recall): decouple temporal seed threshold #2595
• Ollama native provider hardcodes num_ctx=16384, overriding Modelfile context and causing reload/eviction loop #2572 → PR fix(ollama): make native num_ctx opt-in #2589
• retain: fact extraction hard-fails when the LLM returns a top-level JSON array instead of {"facts": [...]} — retry-proof, systematic document ingestion failure with lax-JSON models #2543 → PR Accept top-level fact arrays in retain parsing #2556
• Mental-model content grows unbounded via delta-refresh, permanently overflows embedding model's input token limit (no truncation before embed) #2501 → PR fix(embeddings): truncate oversized litellm-sdk inputs before embedding (#2501) #2516 / fix(embeddings): cap LiteLLM SDK inputs over the model token limit (#2501) #2515
• openclaw: append-capability probe checks only API version, ignores features.store_document_text — every session retain 400s on text-disabled deployments #2505 → PR fix(openclaw): gate append retention on stored text #2511

[handnewb]

CORRIGIDO: Lista completa de issues sem PR (51 de 62)

🐛 Bugs

• Consolidation task not working because ValueError('year 0 is out of range') #2533 bug ValueError('year 0 is out of range') quebra consolidação
• DELETE/forget memory operation returns 404 for IDs that exist (get_memory succeeds) #2523 DELETE/forget retorna 404 para IDs que existem
• openai_compatible_llm._strip_code_fences() produces invalid JSON on partial-fence / non-JSON content, triggering false "JSON parse error after 4 attempts" #2542 _strip_code_fences() produz JSON inválido com partial-fence content
• GLIBC_2.39 not found when running hindsight binary on older Linux systems #2321 bug GLIBC_2.39 não encontrado em Linux antigo

🔧 Consolidação / Retain / Dados

• consolidation persists degenerate LLM output into observation text — repetition loops, leaked structured-output syntax, unsanitized merge path (lax-JSON models) #2544 degenerate LLM output persiste em observation text
• retain: lenient-parse path persists degenerate field content — repeated-sentence loops and leaked response JSON accumulate in what/why, producing multi-KB "facts" #2547 lenient-parse: repeated-sentence loops e leaked JSON em what/why
• retain: OutputTooLong auto-split cuts conversation JSON-array chunks at character midpoints → invalid JSON + unbounded recursion #2571 OutputTooLong auto-split quebra JSON em character midpoints → unbounded recursion
• consolidation failure path leaks LLM context — RSS grows unboundedly on retry storms #2273 consolidation failure path vaza contexto — RSS cresce sem limite

🧠 Recall / Retrieval / Graph / Modelos Mentais

• Clarify whether graph retrieval should reuse semantic and temporal seeds from recall #2593 graph retrieval deve reusar semantic/temporal seeds do recall?
• Graph entity expansion should handle high-frequency entities more explicitly #2592 entity expansion deve lidar com high-frequency entities
• agent_knowledge_recall: same-session PASS but fresh-session FAIL under static bankId — expected cross-session behavior? #2591 agent_knowledge_recall: mesma sessão OK, sessão nova FAIL
• Clarify causal link taxonomy: retain only writes caused_by while schema/retrieval support other causal types #2590 causal link taxonomy: retain só usa caused_by mas schema suporta outros
• reflect: gpt-oss harmony tool-call tokens leak into answer text (litellm MaaS, v0.8.2) #2222 reflect: tool-call tokens vazam no answer text (gpt-oss, litellm MaaS)
• Structured reflect with response_schema times out while plain reflect succeeds (empty message content, finish_reason=length) #2431 reflect com response_schema timeout, plain funciona (empty content, finish=length)

✨ Enhancements

• CLI: mental-model refresh lacks budget/timeout overrides #2606 CLI: mental-model refresh sem budget/timeout overrides
• refresh_mental_model result_metadata lacks semantic outcome fields (retain ops have them since 0.8.x) #2605 refresh_mental_model result_metadata sem semantic outcome fields
• Provide extra surrounding context for extraction as needed to resolve misattribution and ambiguity #2551 contexto extra para extraction resolver misattribution
• Improve temporal ordering consideration for consolidation #2550 temporal ordering na consolidação
• Smarter json chunking to prevent loss of critical context #2548 json chunking mais inteligente
• Time-pattern awareness: Recall by recurring calendar patterns (weekday / month / season), not just absolute dates #2184 time-pattern awareness: recall por padrões de calendário
• Importance-weighted governance decision storage for AI agent compliance #2284 importance-weighted governance decision storage
• Modernize the Codex CLI integration into a packaged Codex plugin (auto-wired hooks + native MCP) #2416 Codex CLI → packaged Codex plugin
• CLI binary for Windows #2415 CLI binary para Windows
• Are there plans to support multimodal data—for example, memory management for images, audio, and video? #1801 suporte multimodal (imagens, áudio, vídeo)
• [Feat Request] Fallback Parameter Auto-Discovery via LiteLLM Proxy /model/info Endpoint #1778 fallback parameter auto-discovery via LiteLLM /model/info
• [feature] ~ Expose retrieval scores in recall API response for transparency and threshold-based filtering #1624 expor retrieval scores no recall API response
• Vscode/GitHub Copilot Integration #1588 VSCode/GitHub Copilot integration
• opencode: Per Agent Configuration #1570 opencode: per-agent configuration
• Provider Request: GitHub SDK / with github.com + company.ghe.com support #1449 GitHub SDK provider (github.com + GHE)
• feat(openclaw): Option to skip retention of low-value transcripts (chitchat, noise, pure tool calls) #1405 openclaw: skip retention de low-value transcripts
• [Feature]: Gemini FLEX API #1000 Gemini FLEX API
• Add ability to retain files from S3 #446 reter arquivos do S3
• [Feature] Support custom LLM provider options when using Hindsight Cloud #436 custom LLM provider options no Hindsight Cloud
• [Feature]: I'd like to benefit from disk savings with Embedding Quantization #385 embedding quantization para economia de disco

🛠 Integração / Platforma

• pg0 mode: startup migrations fail with 'Database URL is required for migrations' (v0.8.4); hindsight-admin run-db-migration works #2602 pg0 mode: migrations falham (Database URL required)
• Claude Code / Codex: directoryBankMap exact-match fails when agent navigates to subdirectories (cwd is dynamic) #2519 Claude Code/Codex: directoryBankMap falha com cwd dinâmico
• [BUG] Three integration plugin issues across OpenClaw, OpenCode, and Claude Code #1758 3 bugs em plugins: OpenClaw, OpenCode, Claude Code
• Paperclip plugin: per-agent enable/disable for pilot rollouts #1116 Paperclip: per-agent enable/disable
• OpenClaw memory-wiki bridge imports 0 artifacts with hindsight-openclaw #963 OpenClaw memory-wiki bridge importa 0 artifacts
• Feature: display Hindsight version in the web UI (port 9999) #776 mostrar versão no web UI (porta 9999)

🧹 Qualidade de Conteúdo / Guarding

• Degenerate fact text ("...", punctuation-only) passes both extraction parsing and consolidation into recallable memory — no content guard #2520 degenerate fact text (pontuação-only) passa parsing
• Feature request(claude-code): exclude certain tool/command patterns from triggering retain #2143 claude-code: excluir certos tool patterns do retain
• Feature: validity windows on memory_units (valid_to + /invalidate endpoint) #1391 validity windows em memory_units (valid_to + /invalidate)
• RFC: supported GPU runtime memory lifecycle and full read-database routing for self-hosted deployments #2035 GPU runtime memory lifecycle + read-database routing

🔐 Auth / Segurança

• x402 payment integration: let agents pay for Hindsight Cloud operations autonomously #759 x402 payment integration
• [中文] Windows + 中国网络：Hindsight 本地部署完整指南（踩坑记录） #1517 Windows + China: guia de deploy local (踩坑记录)

📋 Organizacional

• Clarify benefit model and evaluation criteria for coding-agent memory integrations #2347 benefit model p/ coding-agent memory integrations

---

Issues já cobertos por PRs (11, excluídos acima):

Issue	PR
#2601	#2614 (este) + #2608
#2554	#2619 (reranker_max_candidates 300→50)
#2597 + #2506	#2620 (FK violation observation_history)
#2604	#2616
#2594	#2595
#2572	#2589
#2543	#2556
#2507	#2508
#2505	#2511
#2501	#2516 + #2515
#2500	#2502
#2492	#2493
#2488	#2395
#2431	#2486
#2235	#2262

Total: 62 abertos → 11 com PR → 51 sem fix

[benfrank241]

Superseded by #2608 (merged), which fixes the same missing _mark_completed call plus two related gaps: a status-guard so the poller never overwrites a terminal state the executor already wrote (avoids masking a failure as completed), and parent-operation propagation on completion (parity with _mark_failed), with unit coverage for both. Thanks @handnewb — spot-on bug report and repro, and your one-liner was the right diagnosis; going with the fuller version to cover the edges.