Action Items

Bugs

Release Review

Next (v0.0.15 release)

Bug fixes (must clear before tagging)

• ollama.go:486 — FastProbeModel sets cap.ToolMode = ToolModeStructured unconditionally for tool-capable models. ModelCache.Set UPSERTs this, silently overwriting any mode the user set via /model mode. Fix: removed cap.ToolMode = ToolModeStructured from FastProbeModel; toolsReliable() falls through to SupportsTools==CapYes correctly.

• terminal.go:1297 — After option-2 retry, toolCallRecords still holds records from the first-pass RunToolLoop (populated at line 1141) even though those history entries were rolled back at line 1181. Fix: set toolCallRecords = nil at the top of the option-2 block before rolling back history.

• terminal.go:1321 — After option-2 retry, noToolCalls was always true. Fix: capture hadToolCalls := len(a.History) > histLenBeforeChat before option-2 fires, then compute noToolCalls := !hadToolCalls && len(a.History) == histLenBeforeChat.

• terminal.go:1185 — Option-2 retry called a.Client.Chat directly, bypassing RunToolLoop. Fix: when useStructuredTools is true, option-2 retry now uses RunToolLoop (with a fresh spinner) and updates toolCallRecords from the retry’s history slice.

• terminal.go:1203 — After option-2 retry with RunToolLoop, a.History[histLenBeforeChat:] now correctly contains the retry’s tool-call messages (populated by the fixed retry path). groundingCheck uses these automatically.

• file_inject.go:116 — cantReadPhrases entry "please provide the file" was too broad. Fix: tightened to "please provide the file content".

• model_cache.go:47 — Exported ToolMode* constants now have /** ... */ block-doc comments with Description and Example sections as required by CLAUDE.md.

Release steps (after bugs cleared)

• Bump codemeta.json version to 0.0.15, update releaseDate and releaseNotes
• Regenerate version.go: cmt codemeta.json version.go
• Commit and tag v0.0.15

Deferred to v0.0.16

Option 2 reactive retry — surgical rollback

See audit-trail-plan.md W1 and the small-model tool-use mitigation work (file_inject.go option 2).

• The current retry in terminal.go calls Client.Chat directly and rolls back the full history by one message. When RunToolLoop added intermediate tool-call/tool-result messages before the refusal, those are silently dropped. Implement a surgical rollback that only removes the assistant refusal message and re-adds the augmented user message, preserving prior tool loop history.

/model mode MODEL auto — reset to auto

See audit-trail-plan.md option 3 (/model mode command, model_cache.go).

• /model mode currently accepts structured, prose, inject, and none but has no auto value to clear an explicit override and return the model to capability-detected defaults. Add auto as a valid mode that sets tool_mode = '' (empty) in model_capabilities, restoring modelToolMode() fallback to CapabilityStatus.

Retraction monitoring service

See scholarly-provenance-plan.md S2 and scholarly-provenance-design.md.

• The sources table in knowledge.db already has retracted INTEGER and retraction_note TEXT columns for manual marking (via /kb retract). Add a periodic background check against the Retraction Watch API (retractionwatch.com) that flags retracted DOIs automatically. A /kb check-retractions command (or scheduled task) should query registered sources with identifier_type = 'doi' and update retracted/retraction_note on hits.

llama.cpp Apertus tool-call format

See henry project (henry-handoff-20260622-llamafile-factory.spmd).

• When llama.cpp gains stable custom token support, update templates/apertus-4b-toolcall.jinja in the henry project to use structured tool-call tokens instead of the current prose JSON fence workaround. Retest with Apertus 4B via bin/assay --llamafile.

Dual RAG injection audit

See DECISIONS.md (2026-06-02 — Dual RAG injection audit, deferred).

• Users with both memory.enabled and rag.enabled receive RAG content twice per turn: once via UnifiedMemory.Recall() at session start and once via ragAugment() per prompt. Audit the overlap and either (a) skip RAG chunks in UnifiedMemory.Recall() when a.RagOn is true, or (b) make ragAugment a no-op when UnifiedMemory already injected from the same store this session.