fix(prompt): stop invalidating provider prompt caches on every turn

[devlikebear]

Closes #920. Phase 0 / LP-001 of #919.

What was wrong

internal/prompt/builder.go:70 wrote a second-resolution timestamp as the first line of every system prompt. Prompt caching is prefix-matched, so nothing ever matched:

• the cache_control marker on the Anthropic system block wrote a fresh entry every turn and never read one,
• OpenAI and Gemini automatic prefix caching never engaged,
• the whole static body (identity, planning, formatting, long-running command rules, workspace bootstrap) was re-charged at write rates on every turn.

Two things the issue did not name turned up during the audit and are fixed here as well:

1. The builder's tail was not the assembled prompt's tail. buildContextFromResult and handler_chat_context.go append the memory rule, skills, session override, style, goal, and critic after the builder's output — so the per-turn ## Prior Context recall sat mid-prompt with static text behind it. Moving only the clock would have passed a prefix unit test while delivering almost none of the win.
2. The chat memory cache stored the fully assembled prompt. The prefetch goroutine (handler_chat_prefetch.go) builds without the session's WorkDirs/CurrentDir, so once its entry landed in the cache the next turn's prompt lost ## Working Directories entirely — a correctness bug, not just a caching one.

What changed

Area	Change
internal/prompt/builder.go	Clock removed from the head; emitted as a trailing ## Current Time block, truncated to the minute. BuildResult exposes StaticPrompt / DynamicTail (Prompt is still the concatenation). Ordering invariant documented on BuildResultFor. PresetRelevantMemory lets a caller replay cached recall without re-searching.
internal/tarsserver/handler_chat*.go	The dynamic tail is carried separately and emitted last, after skills / override / style / goal / critic. buildLLMMessagesWithTail sends it as a second, adjacent system message; an empty tail collapses to one message as before.
internal/llm/anthropic.go	System messages are emitted as one text block each, with cache_control on the first (stable) block instead of after the joined text. A single system message behaves exactly as before.
internal/tarsserver/memory_cache.go	Stores only the recall payload. The prompt is always rebuilt from live options, so cache hit ≡ cache miss by construction — immune to the next BuildOptions field somebody forgets to plumb.

Decisions the issue asked to document

• System-prompt tail, not the first user message. Decorating the outgoing user turn would either pollute persisted history or leave the sent message differing from the stored one — which breaks the message-prefix matching feat(llm): add rolling cache breakpoints to Anthropic message history #921 (LP-002) is about to rely on. The system tail has neither problem.
• Minute resolution. Second resolution is more precision than any assistant answer needs and re-breaks the prefix on every retry, tool-loop restart, and rapid follow-up — exactly the burst where a hit is worth most. Hour resolution would buy little more, since recall churn dominates in memory-bearing sessions, and it makes the time visibly stale.
• Anthropic needs the block split, not just reordering. One breakpoint on a joined system block is all-or-nothing: a volatile tail anywhere inside it means zero reads. Reordering alone fixes OpenAI/Gemini (automatic longest-prefix) but not Anthropic. Breakpoint count stays at 2 of the allowed 4 (system + last tool), leaving room for feat(llm): add rolling cache breakpoints to Anthropic message history #921.

Observed, deliberately not changed

• helpers_agent.go:496 — the cron/minimal agent prompt still embeds a second-resolution timestamp. One-shot prompts with no cross-turn reuse; nothing to cache.
• Single-system-message paths (telegram, sub-agents, previews) stay all-or-nothing on Anthropic, exactly as before. They get the reordering benefit on OpenAI/Gemini.
• BuildPriorContextPreview never included the prompt head, so it was unaffected.

Acceptance criteria

Criterion	Status
Leading region byte-identical across two consecutive turns	✅ TestBuildResult_StaticPrefixSurvivesClockChange
Model still knows the current time	✅ TestBuildResult_KeepsCurrentTimeInTail, ..._SubAgentPromptStillCarriesTime
Prefetched and live prompts remain identical	✅ TestPrepareChatContext_CacheHitRebuildsLiveStaticRegion
Second Anthropic turn reports non-zero cache_read_input_tokens	⬜ needs a live two-turn run against a real Anthropic tier — not runnable in CI, see below

Validation

• bash scripts/windows_test.sh — green (the CI Windows job's exact selection).
• golangci-lint run --enable=errcheck --enable=staticcheck --new-from-rev on the changed packages — 0 issues.
• scripts/check_diff_coverage.sh — 95.0% (96/101 changed coverable lines), threshold 80%.
• Note that windows_test.sh excludes internal/llm and internal/tarsserver wholesale, so locally those two packages were verified by targeted runs (all chat/memory/prompt/context/anthropic tests pass) plus a base-commit comparison confirming every remaining failure is a pre-existing Windows environment issue (POSIX paths, owner-only file modes, symlink privilege, macOS/Linux notifiers). The full-package proof lands in the Linux CI jobs.
• Still outstanding: the manual two-turn session against a real Anthropic tier, recording cache_read_input_tokens before and after. Worth recording the before number too — the existing last-tool breakpoint may already produce a non-zero read, so the delta is what attributes the win to the system prefix.

Out of scope

Cache breakpoints on the message history (#921) and any change to what the static sections contain.

🤖 Generated with Claude Code

[sonarqubecloud]

Quality Gate passed

Issues
0 New issues
0 Accepted issues

Measures
0 Security Hotspots
93.7% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud