feat(timeline): compression parity upgrades (P2+P3+P6)

[ngoclam9415]

Summary

Close three compression gaps vs OpenClaude while staying single-tier, LLM-agnostic, KISS. Always uses len(str)/4 heuristic — zero provider coupling. System+tools coverage via optional caller-supplied callbacks.

Implements plan 260419-2241-compression-parity-dana-runtime (Phases 1–4).

What's in

Phase 1 — Heuristic threshold (P3)

• DANA_COMPACT_TRIGGER_TOKENS env knob (default 150000, clamp [8k, 2M])
• CompressedTimeline(system_tokens_fn, tools_tokens_fn) callbacks folded into needs_compression() estimate
• star_agent wires len(system_prompt)//4 + len(json.dumps(tools))//4

Phase 2 — Cheap client-side shrink (P6)

• cheap_shrink_tool_results() stubs old tool_result bodies to [cleared for context budget] preserving tool_call_id
• Predictive gate blocks mutation when shrink-alone can't close gap → avoids vacuous summaries over stubs
• Idempotent via content equality; opt-in via enable_cheap_shrink_tool_results (default off)

Phase 3 — Reactive compact + circuit breaker (P2)

• PromptTooLongError typed exception; provider mapping for Anthropic (invalid_request_error + "prompt is too long"), OpenAI-compat (context_length_exceeded), Gemini (post-hoc WARNING on MAX_TOKENS)
• llm_caller._invoke_llm_sync/async wraps with PTL catch → reactive_compact(attempt) → retry with 1s/3s backoff
• reactive_compact drops {1:5, 2:10, 3:20} oldest + forward-orphan pruning + full re-summary (no shrink-bypass)
• Per-session circuit breaker with cooldown recovery (DANA_CIRCUIT_COOLDOWN_SECONDS, default 300s) + half-open probe + reset_circuit() ops hatch
• Kill switch DANA_DISABLE_REACTIVE_COMPACT=1
• star_agent._maybe_compress_timeline re-raises PTL explicitly (no swallowing)

Phase 4 — Telemetry

• CompressionLogFields TypedDict authoritative allowlist + new_compaction_id()
• AST-based test asserts logger.*(..., extra={...}) keys stay within allowlist (no prompt-content leakage)

Files

Source: 2 new (compact_trigger.py, telemetry.py), 8 modified (providers, types.py, compression_engine.py, compressed_timeline.py, llm_caller.py, star_agent.py).
Tests: 6 new files, 44 new unit tests — all pass.
Fixtures: 3 provider PTL JSONs under tests/fixtures/provider_ptl/.
Docs: system-architecture.md updated, project-changelog.md created.

Known gaps (follow-up PRs — report filed)

Review report: plans/reports/code-review-260420-1112-compression-parity-triggering.md

• CRITICAL-1 PTL retry closes over the captured messages list; post-compact retry re-sends the stale oversized payload → circuit opens on genuinely-recoverable turns. Tests pass because fake LLM ignores payload content.
• CRITICAL-2 Recent huge tool_result is unreclaimable (shrink keep_recent=10 blocks it; reactive_compact drops oldest only).
• HIGH Multi-compression does not preserve prior summary text; stubbed content persisted across reload produces vacuous summaries on re-compression.

These should be addressed before relying on reactive recovery in production. The exception types, circuit breaker, telemetry, and pre-turn heuristic trigger are usable as-is.

Test plan

• Unit: uv run pytest tests/unit/test_compact_trigger.py tests/unit/test_compressed_timeline_callbacks.py tests/unit/test_cheap_shrink.py tests/unit/test_reactive_compact.py tests/unit/test_llm_caller_ptl_retry.py tests/unit/test_log_field_allowlist.py → 44/44 pass
• Full unit suite regression: 1064 passed (1 pre-existing failure in test_star_agent_streaming.py verified on develop, unrelated)
• Compile: uv run python -m compileall dana/core/timeline dana/core/agent dana/common/llm dana/core/llm → OK
• Integration test for CRITICAL-1 (stale messages) — follow-up PR
• Integration test for CRITICAL-2 (huge recent tool result) — follow-up PR