feat: auto-compact and retry on context window errors

[TheArchitectit]

Problem

When the model API returns a context window exceeded error, only the interactive REPL path (run_turn()) had auto-compact retry logic. The non-interactive paths (run_prompt_json, run_prompt_compact, run_prompt_compact_json) simply propagated the error with result? and no retry, causing a hard stop for users running claw -p or claw -p --json commands.

Additionally, context window detection in the existing run_turn() used ad-hoc string matching (contains("context_window") || contains("no parseable body")) instead of the canonical is_context_window_failure() method from the api crate, and the "no parseable body" marker (added for OpenAI-compat backends in PR #3214) was missing from CONTEXT_WINDOW_ERROR_MARKERS.

Solution

1. Added "no parseable body" to CONTEXT_WINDOW_ERROR_MARKERS in the api crate, so is_context_window_failure() now covers OpenAI-compat backends that return 400 with an un-parseable body.

2. Added RuntimeError::is_context_window_failure() in the runtime crate. Since ApiError is erased into a string message at the runtime boundary, we need a runtime-level marker check that mirrors the api crate's detection. This replaces the ad-hoc inline string matching.

3. Extracted auto-compact retry into LiveCli::auto_compact_retry() -- a shared helper method that:

   • Detects context window errors via RuntimeError::is_context_window_failure()
   • Compacts progressively (preserve 4 -> 2 -> 0 recent messages)
   • Retries the same user input with the compacted session
   • Is bounded by MAX_COMPACT_RETRIES = 3 to prevent infinite loops
   • Logs user-facing messages: "Context limit reached, auto-compacting session... (attempt N/3)"

4. Extended auto-compact retry to ALL turn execution paths:

   • run_turn() (interactive REPL) -- now uses shared helper
   • run_prompt_compact() (-p --compact) -- auto-retry added
   • run_prompt_compact_json() (-p --compact --json) -- auto-retry added
   • run_prompt_json() (-p --json) -- auto-retry added

Changes

• rust/crates/api/src/error.rs: Added "no parseable body" to CONTEXT_WINDOW_ERROR_MARKERS
• rust/crates/runtime/src/conversation.rs: Added RUNTIME_CONTEXT_WINDOW_MARKERS constant and RuntimeError::is_context_window_failure() method
• rust/crates/rusty-claude-cli/src/main.rs: Added LiveCli::MAX_COMPACT_RETRIES = 3 constant, LiveCli::auto_compact_retry() method; replaced inline retry logic in run_turn() with delegation to the shared helper; added auto-compact retry to run_prompt_compact(), run_prompt_compact_json(), run_prompt_json()

Diff Verification

rust/crates/api/src/error.rs             |   1 +
rust/crates/runtime/src/conversation.rs  |  37 +++++
rust/crates/rusty-claude-cli/src/main.rs | 273 +++++++++++++++++--------------
3 files changed, 184 insertions(+), 127 deletions(-)

The 127 deletions in main.rs are the inline retry code being extracted into auto_compact_retry() (refactor, not revert). No upstream commits were lost.

Testing

1. Unit tests: cargo test -p api -p runtime -- all pass
2. Compile check: cargo check --workspace -- clean (no warnings from modified code)
3. Format: cargo fmt -p api -p runtime -p tools -- no changes needed

Manual testing scenario:

• Run claw in a session with many messages until context limit is hit
• Observe auto-compact message: "Context limit reached, auto-compacting session... (attempt 1/3)"
• Session should automatically compact and retry the turn without user intervention
• Test with claw -p "long prompt" and claw -p --json "long prompt" as well

Related: Builds on error detection from PR #3015 and the "no parseable body" marker from PR #3214.

[1716775457damn]

Nice improvement — extending auto-compact retry from REPL-only to all execution paths (prompt, compact, json variants) fixes a real UX pain point. Using RuntimeError::is_context_window_failure() via the canonical api crate markers is the right approach, much cleaner than ad-hoc string matching. The progressive compaction strategy (4→2→0) with MAX_COMPACT_RETRIES=3 is well-bounded.\n\nOne note: I see cargo fmt is failing in CI — consider running cargo fmt --all before merge to keep checks green.