fix(agent): persist image-shrink recovery

[embwl0x]

Summary

• record the exact data-URL replacements produced by reactive image-shrink recovery
• mirror those replacements into canonical session history for OpenAI and Anthropic image shapes
• persist the repaired history before retrying, preventing later turns from re-sending and re-encoding the same oversized images
• avoid a second decode/re-encode by reusing the already successful provider-facing repair

Current-main proof

On current main, try_shrink_image_parts_in_messages() mutates only the provider-facing api_messages copy. The new regression failed before the fix because the helper could not expose any replacement for canonical history; the original oversized payload therefore remained in messages for the next turn.

After the fix, the helper returns the exact old-to-new payload mapping, detached canonical OpenAI and Anthropic messages are rewritten, and the repaired session is persisted before the retry. The retry-loop regression records the first rejected request, repaired canonical persistence, and repaired retry in order while asserting the image encoder runs exactly once.

Duplicate / sibling-path audit

Reviewed open image-shrink PRs #37412, #42509, #49124, #52444, and #53884. They cover 413 routing, classification, downscale acceptance, or wrapped-error parsing. None persists a successful repair into canonical session history.

Sibling shapes audited: OpenAI image_url dict/string parts and Anthropic native base64 image.source blocks.

Verification

• tests/run_agent/test_image_shrink_recovery.py: 25 passed
• 413 and image-rejection suites: 42 passed
• image routing, compression persistence, and retry-state suites: 109 passed
• tests/run_agent/test_run_agent.py: 416 passed
• Ruff passed on all changed files
• git diff --check passed
• pushed-range gitleaks passed

No live Hermes state or provider credentials were used.

Fixes #61994

[teknium1]

Thanks for tracing the recovery state through canonical session persistence. The premise is confirmed on current main: Anthropic cache preparation deep-copies the API request (agent/prompt_caching.py:84-99), while the image-too-large retry only shrinks api_messages (agent/conversation_loop.py:2632-2647).

Problems

• The new loop-level persistence behavior lacks a regression test. tests/run_agent/test_image_shrink_recovery.py:12-15 explicitly documents that retry-loop wiring is not tested, and the added tests only validate mapping collection and standalone canonical replacement. They do not assert the new _persist_session() call in the retry path.

Suggested changes

• Add a focused retry-path test with detached canonical and API message objects. Assert the canonical payload is rewritten, _persist_session(messages, conversation_history) is called before retry, and the retry uses the repaired image without another encode.

Automated hermes-sweeper review.

[embwl0x]

Addressed in 31ae7be on current main (a0a6cd80). The new retry-loop regression uses detached canonical/API image objects and asserts the order first rejected request -> canonical persistence -> repaired retry, with exactly one resize call (25 focused tests pass). The branch update is published on the earlier #62005 PR; no new PR was opened. The later competing #62051 remains documented for maintainer selection, and its sweeper review points to the reactive persistence design in #62005 rather than a global ingestion cap.

[embwl0x]

Resolved the current-main conflict in e76f996 while preserving both invariants: image-shrink remains copy-on-write for API-local messages, and a confirmed provider rejection still mirrors the exact successful repair into canonical history before persistence and retry.

Local verification:

• 27 focused image-shrink tests passed
• 185 adjacent image routing, prompt caching, retry-state, and compression persistence tests passed
• 446 run-agent tests passed with isolated HERMES_HOME
• Ruff, git diff --check, publish gate, and gitleaks passed

GitHub now reports the PR mergeable; replacement CI is running.

[embwl0x]

Resolved the new current-main conflict in 63e5b4d56 while preserving both image-shrink invariants: API-local rewriting remains copy-on-write, and a confirmed provider rejection still mirrors the exact successful repair into canonical history before persistence and retry.

The conflict was in tests/run_agent/test_image_shrink_recovery.py. I retained the PR's persistence regressions and current main's stronger copy-on-write/dimension coverage, without resurrecting the redundant dimension test removed upstream.

Local verification:

• 16 focused image-shrink recovery tests passed
• 73 adjacent image routing, prompt caching, retry-state, and compression persistence tests passed
• 220 full run-agent tests passed in an isolated HERMES_HOME
• Ruff, Python compilation, git diff --check, publish gate, and PR-only gitleaks passed

GitHub reports the updated head mergeable; replacement CI is running.

[embwl0x]

Replacement CI completed with the conflict resolution itself clean: 7/8 Python slices passed, both Docker builds passed, Desktop E2E passed, and all lint/security/lockfile checks passed.

Slice 8 failed only in tests/tools/test_vercel_sandbox_environment.py: 16 tests hit FeatureUnavailable because terminal.vercel is absent and lazy installs are disabled. This is an upstream-base failure introduced by the current-main Vercel restoration, not by this PR:

• Exact base tested: 2d404942471633d5338a8ff514ea7da24549274f
• Untouched detached base worktree: the same 16 tests fail with the same FeatureUnavailable path
• Updated PR head: the same 16 tests fail
• The PR diff remains limited to the four image-shrink files

I did not import an unrelated Vercel fixture repair into this image-shrink PR. The conflict remains resolved and GitHub reports the head mergeable.

[embwl0x]

Upstream landed the unrelated Vercel CI fixture correction in 8eb06e75b. I merged current origin/main into this branch at be96ff449, preserving the image-shrink diff without additional source changes.

Verification:

• 16 focused image-shrink recovery tests passed locally
• 16 Vercel sandbox tests that previously failed passed locally
• focused Ruff, git diff --check, publish gate, and gitleaks passed
• replacement CI is fully green: all 8 Python slices, both Docker builds, Desktop E2E, lint, security, supply-chain, and the required aggregate gate

GitHub reports the refreshed head mergeable.