fix(agent): agentic parity on both surfaces — protocol shape, image delivery, quota status, WAF-aware uploads, history prefix reuse - #168

Merged

Rfym21 merged 56 commits into

Rfym21:mainfrom

maxff77:fix/agentic-parity

Sep 11, 2026

Commits on Sep 7, 2026

fix(images): deliver tool_result images and stop the externalized-context 500

authored and
PEDRO LOBATO CARCAMO
committed
6266b8f

Commits on Sep 8, 2026

fix(images): harvest current-turn images from non-last messages

PEDRO LOBATO CARCAMO
authored and
PEDRO LOBATO CARCAMO
committed
a783f92
fix(openai): harvest current-turn media so a trailing text message cannot drop the image

authored and
PEDRO LOBATO CARCAMO
committed
e3bd245
fix(openai): turn boundary is the last final answer, not any assistant message

authored and
PEDRO LOBATO CARCAMO
committed
5019f04
fix(anthropic): a single tool call no longer erases the image from the turn

authored and
PEDRO LOBATO CARCAMO
committed
526fd49
fix(openai): dedupe after folding, attach to null content, gate t2i, cap the harvest

authored and
PEDRO LOBATO CARCAMO
committed
1c50b43
fix: surface unsupported Anthropic blocks; stop a truncated closer eating a tool call

authored and
PEDRO LOBATO CARCAMO
committed
c6855d5
fix(agent-context): spend the context budget, and tell the client when it was cut

authored and
PEDRO LOBATO CARCAMO
committed
c3188d9
perf(images): reuse one upload across a turn instead of re-uploading per request

authored and
PEDRO LOBATO CARCAMO
committed
264d562
merge: image delivery works on both API paths, plus the context budget and upload cache

PEDRO LOBATO CARCAMO
authored and
PEDRO LOBATO CARCAMO
committed
46106c5
test(probes): add a live agent-loop protocol probe

authored and
PEDRO LOBATO CARCAMO
committed
309da59
fix(tool-protocol): number folded tool calls so results are addressable

authored and
PEDRO LOBATO CARCAMO
committed
a5ea559
feat(agent-turn): give the model a ledger of what it already ran

PEDRO LOBATO CARCAMO
authored and
PEDRO LOBATO CARCAMO
committed
46eca74
feat: inject the executed-call ledger on both API paths

authored and
PEDRO LOBATO CARCAMO
committed
7c97771
feat(agent-turn): seed the tool ledger from history and warn on repeats

PEDRO LOBATO CARCAMO
authored and
PEDRO LOBATO CARCAMO
committed
e2eb41c
fix(anthropic): a truncated turn reports max_tokens even when a tool call was emitted

authored and
PEDRO LOBATO CARCAMO
committed
127ae21
fix(anthropic): tool_use ids use the toolu_ prefix

authored and
PEDRO LOBATO CARCAMO
committed
ee18c78
fix(openai): strip tool-call residue from delivered text

authored and
PEDRO LOBATO CARCAMO
committed
4d87e70
fix(anthropic): keep tool history readable when the request sends no tools

authored and
PEDRO LOBATO CARCAMO
committed
62ba427
fix(anthropic): keep inbound thinking blocks in history

authored and
PEDRO LOBATO CARCAMO
committed
6d9d15c

Commits on Sep 9, 2026

fix(tool-protocol): recognise the numbered closer the folded history teaches

PEDRO LOBATO CARCAMO
authored and
PEDRO LOBATO CARCAMO
committed
b479529
fix(agent-turn): the ledger can no longer forge a line or misattribute a result

PEDRO LOBATO CARCAMO
authored and
PEDRO LOBATO CARCAMO
committed
0108eac
fix(agent-turn): the ledger tells the truth, identically on both paths

PEDRO LOBATO CARCAMO
authored and
PEDRO LOBATO CARCAMO
committed
f2b5a6f
fix(openai): a turn that is only protocol residue fails instead of delivering nothing

PEDRO LOBATO CARCAMO
authored and
PEDRO LOBATO CARCAMO
committed
ab89650
fix(T8 repair): land the OpenAI twin and stop forged ordinals colliding

PEDRO LOBATO CARCAMO
authored and
PEDRO LOBATO CARCAMO
committed
23c5b2c
fix(T9 repair): the delimiter stops being forgeable and the budget stops being per-message

authored and
PEDRO LOBATO CARCAMO
committed
8a68433
chore(probes): drop the dead initializer eslint flags in the agent-loop probe

authored and
PEDRO LOBATO CARCAMO
committed
4238d0c
test(probes): add the duplicate-onset replay harness

authored and
PEDRO LOBATO CARCAMO
committed
f0407ce
fix(probes): make the replay harness measure what it claims to measure

PEDRO LOBATO CARCAMO
authored and
PEDRO LOBATO CARCAMO
committed
785486d
test(probes): record that the replay design cannot reproduce the bug

PEDRO LOBATO CARCAMO
authored and
PEDRO LOBATO CARCAMO
committed
8e45452
fix(images): stop the folded tool result claiming an image read returned nothing

PEDRO LOBATO CARCAMO
authored and
PEDRO LOBATO CARCAMO
committed
03b0587
test(probes): make the tool_result image probe score what it claims to score

PEDRO LOBATO CARCAMO
authored and
PEDRO LOBATO CARCAMO
committed
5e62b00
fix(images): bound the upload cache by the URL's own signed expiry, not a guess

PEDRO LOBATO CARCAMO
authored and
PEDRO LOBATO CARCAMO
committed
3f0f48b
fix(openai): stop the turn gate turning a correct answer into HTTP 429

authored and
PEDRO LOBATO CARCAMO
committed
a9c9f57
fix(images): say an image is here only when it is, and only when we said it

authored and
PEDRO LOBATO CARCAMO
committed
08e504a
fix(images): a cache entry that vanished mid-read must re-upload, not ship null

authored and
PEDRO LOBATO CARCAMO
committed
e80b54b
test(probes): give the tool_result image probe a cell that can fail

authored and
PEDRO LOBATO CARCAMO
committed
8411eb4
fix(openai): close the residue leak the tolerant unwrap opened, and stop it faking closes

authored and
PEDRO LOBATO CARCAMO
committed
aedd124
fix(images): one HARVEST_MEDIA_CAP, and a test that fails when the twin scans diverge

PEDRO LOBATO CARCAMO
authored and
PEDRO LOBATO CARCAMO
committed
2575b71
test(agent-turn): pin the ledger's default entry capacity at 40

authored and
PEDRO LOBATO CARCAMO
committed
1c31c6f
test(probes): make the A/B driver's tool simulator honest, and pin it

authored and
PEDRO LOBATO CARCAMO
committed
ec0de7e
fix: map Qwen's daily-quota refusal to 429 on both API paths

authored and
PEDRO LOBATO CARCAMO
committed
08c724f
fix(agent-turn): raise the ledger byte cap to 12000, and pin the knob that governs it

authored and
PEDRO LOBATO CARCAMO
committed
e5babf7
test: make a short run fail instead of exiting 0

PEDRO LOBATO CARCAMO
authored and
PEDRO LOBATO CARCAMO
committed
5ca8e60
fix(quota): stop re-picking the dead account, and tell the truth about the 429

authored and
PEDRO LOBATO CARCAMO
committed
41b2e4d
fix(agent-turn): give the ledger its own inline budget, so 12000 reaches the model

authored and
PEDRO LOBATO CARCAMO
committed
258a658
fix(test-gate): a skipped test is not a pass, and bless can go down

authored and
PEDRO LOBATO CARCAMO
committed
a493ada

Commits on Sep 10, 2026

revert(agent-turn): the ledger goes back to 6000 B, its benefit was never measured

authored and
PEDRO LOBATO CARCAMO
committed
40df21e
test(agent-turn): pin what the model READS of the ledger when the prefix is truncated

authored and
PEDRO LOBATO CARCAMO
committed
fb904db
test(agent-turn): medir la ceguera del brazo por defecto en vez de contarla

authored and
PEDRO LOBATO CARCAMO
committed
ef1b6b2
chore(probes): the replay harness stops shipping an operator's transcript path

authored and
PEDRO LOBATO CARCAMO
committed
b24335f
fix(context): attachment failure with tools is a retryable 529/503, never a silent compacted 200

PEDRO LOBATO CARCAMO
authored and
PEDRO LOBATO CARCAMO
committed
f43af96
upload: fail fast on Aliyun WAF captcha at /files/parse + circuit breaker

PEDRO LOBATO CARCAMO
authored and
PEDRO LOBATO CARCAMO
committed
16c9207
fix(upload): rate-limit /files/parse before the WAF does

PEDRO LOBATO CARCAMO
authored and
PEDRO LOBATO CARCAMO
committed
f73850f
feat(context): reuse baked history prefix across turns

PEDRO LOBATO CARCAMO
authored and
PEDRO LOBATO CARCAMO
committed
4c26239

Commits on Sep 11, 2026

Fix prefix-cache misses and 529 fallback gaps found in code review

PEDRO LOBATO CARCAMO
authored and
PEDRO LOBATO CARCAMO
committed
e7e14d2