[wolfy-j]

Problem

Claude binds thinking blocks and the prompt cache to the exact request prefix: top-level system, tools, and every earlier message. Content rebuilt or moved between requests, while sitting before already produced thinking blocks, invalidates them from that point. With #235 the API drops those blocks instead of returning 400, but the reasoning is then lost on every later request and the cache is rewritten. Sources in the framework:

• The time_aware trait put the current time (rounded to 15 minutes) into the system prompt.
• Per-step notes moved: memory recall was a per-step developer message at the tail that hosts persisted at a different position; Claude rendered developer notes differently depending on where the rolling cache marker sat; Bedrock hoisted developer notes into system.
• After checkpoints and handoffs, thinking blocks bound to an old prefix were replayed with no shared way to remove them.

Change

llm

• A SYSTEM or DEVELOPER message with metadata.preserve_position = true is a positional note (prompt.is_positional_note). Claude and Bedrock append it as its own text block to the preceding user turn (including a tool-result turn), or open a new user turn when none precedes it; it is never merged into earlier blocks and never placed in top-level system. Google renders it as a user turn; OpenAI and OpenAI-compatible keep it positional.
• Developer notes on Claude and Bedrock always render as their own text block <developer-instruction>…</developer-instruction> appended to the preceding user turn, independent of cache markers, never edited into earlier blocks. On Claude a note that followed plain user text now sits in its own block instead of being joined with "\n"; notes after a cache marker, an assistant turn or tool results now carry the wrapper too. Bedrock renders developer messages as user content instead of system.
• builder:drop_stale_thinking(is_current) removes thinking_blocks as a leading run: through the last stale message, refusing (history unchanged) when that would split an assistant turn with an unanswered call. Stripped messages are copies; caller tables are not mutated. An empty list carries nothing.
• Google's mapper no longer clears metadata on caller messages.
• Golden request-body tests for all five drivers; Claude and Bedrock developer cases are recorded from this branch, everything else from master.

agent

• time_aware contributes agent_options.turn_context.time = { timezone, format? } (trait context keys timezone and time_format); wippy.agent:turn_context.render(agent_options, now) renders the note to the exact minute for hosts to persist once per user turn. time_interval is rejected with an error naming the replacement.
• Trait build-function errors fail compilation (previously swallowed); a failing prompt function fails the step, keeping kind, retryability and details.
• Memory recall runs whenever the agent has a memory contract: the note is a positional developer note at the tail and is returned in response.context_notes for the host to persist in place. response.memory_prompt is removed.
• response.prefix_fingerprint: sha256 of the sent Claude/Bedrock prefix projection (system texts incl. hoisted conversation system messages; tools by name, description, schema and native definition), computed before the request. With runtime_options.drop_stale_thinking the agent strips blocks whose stored fingerprint differs (or is absent) before sending; runtime_options.thinking_is_current(msg, fingerprint) lets hosts add their own epoch (e.g. checkpoint id). The fingerprint never reaches the wire.
• The conversation-tail cache marker precedes trailing developer and positional notes. The sorted-key encoder from attention_read moves into wippy.agent:canonical.

Tests

llm 2083 pass (9 skipped), agent 560 pass (4 skipped); repo-wide tests, lint and manifests pass. Every review finding is pinned by tests that failed before its fix.

Release coordination

• Merge after fix(llm): let the API drop thinking blocks bound to another conversation #235: the Bedrock developer move and the Claude developer rendering change stored histories' prefix once on upgrade; fix(llm): let the API drop thinking blocks bound to another conversation #235's drop_block turns that into a drop instead of a 400 on bound requests (thinking: always, or adaptive/budget with effort > 0). Unbound requests on enforced accounts can 400 once until hosts strip with drop_stale_thinking.
• Session and dataflow must persist context_notes, the turn-context note, prefix_fingerprint storage and drop_stale_thinking in the same release; until then they have no time note, and recall still reaches the request but is not persisted.
• The llm changes release as wippy/llm 0.5.13; the agent requires >=v0.5.13 (earlier llm lacks prompt.is_positional_note). llm and agent publish together. Squash with an explicit commit message without !/BREAKING CHANGE markers so release-please cuts patch releases; the behaviour changes are listed above.
• Upstream follow-up: go-lua errors.wrap context is not visible from Lua (tostring/:message() return the parent message).

Merge instructions

• Merge fix(llm): let the API drop thinking blocks bound to another conversation #235 first, then this PR, then refresh the release PR (chore: release master #236) and confirm it lists both under llm 0.5.13 before merging it. If the release PR merges first, 0.5.13 ships without prompt.is_positional_note and the agent requirement must name the version that actually carries it.
• Squash with this commit message (no ! / BREAKING CHANGE markers, so release-please cuts patches: llm 0.5.13, agent 0.5.10):

feat(agent): keep the request prefix stable across steps

Positional notes (system or developer messages with
metadata.preserve_position) render in place on every driver; Claude and
Bedrock developer notes render as their own <developer-instruction>
block independent of cache markers, and Bedrock developer messages are
sent as user content instead of system blocks. builder:drop_stale_thinking
removes thinking blocks bound to another prefix as a leading run.

The agent returns prefix_fingerprint and, with drop_stale_thinking, strips
stale thinking before each request. Memory recall returns its note in context_notes for hosts to persist in
place; response.memory_prompt is removed. time_aware contributes agent_options.turn_context instead of
system prompt text and rejects time_interval; turn_context.render renders
the note for hosts. Trait build-function errors fail compilation and a
failing prompt function fails the step. The agent requires wippy/llm
>= 0.5.13.