Make CodeAF the default harness, installed automatically with AgentField by AbirAbbas · Pull Request #1128 · Agent-Field/agentfield

AbirAbbas · 2026-10-23T21:18:49Z

Summary

CodeAF (the renamed aforge-v2, Agent-Field/CodeAF) becomes the default harness in the Python, Go and TypeScript SDKs, and it is installed automatically wherever AgentField is. A provider-less app.harness(...) call runs codeaf exec --json headless with nothing else to set up.

New codeaf provider in all three SDKs. It runs codeaf exec --json --dir <root> with the prompt on stdin (AGENTFIELD_CODEAF_COMMAND=do opts into codeaf do). It reads CodeAF's current envelope (ok, stop, answer, spend_usd, tokens, model) and exit codes rather than the legacy aforge fields CodeAF keeps for one more release.
Nothing a user already chose changes.
- provider="aforge" (or AGENTFIELD_HARNESS_PROVIDER=aforge) is an alias that runs CodeAF.
- A legacy pin keeps the original aforge adapter and binary, both with provider="aforge" and when no provider is named (as on main). The pins are AFORGE_BIN, Python aforge_bin, TS aforgeBin, and Go Options.BinPath with the default provider.
- AGENTFIELD_AFORGE_COMMAND and AFORGE_MODEL are honored as fallbacks for AGENTFIELD_CODEAF_COMMAND and CODEAF_MODEL.
Pinned, checksum-verified binary. CodeAF v0.8.2 is pinned, with all six release assets' SHA-256 embedded in each SDK and in af. It installs only into $AGENTFIELD_HOME/bin/codeaf and never touches a user's own CodeAF.
Installed by every path that installs AgentField:
- install.sh / install.ps1
- Desktop launch
- af skill install --all
- a successful af install
- the python-agent, go-agent and control-plane-cloud images
- af codeaf ensure, the explicit command
- for pip/npm/Go-only users, a lazy download on the first harness call
The installers are served from main but install the latest stable af. When that af has no codeaf command yet, they provision aforge exactly as before, so there is no gap between merging this and the next stable release. Opt out with --no-codeaf, -NoCodeaf or AGENTFIELD_SKIP_CODEAF=1; the old aforge opt-outs still work.
Quiet and isolated in automation.
- CodeAF state lives in $AGENTFIELD_HOME/codeaf.
- The update check is off.
- Telemetry is off under CI, DO_NOT_TRACK or AGENTFIELD_TELEMETRY_ENABLED=false.
- The default worker model is ~deepseek/deepseek-v4-flash-latest.

Behavior (validation contract)

With no provider named, CodeAF runs and usage is attributed harness=codeaf. An explicit provider or AGENTFIELD_HARNESS_PROVIDER still wins, and a blank value is ignored.
aforge without a pin runs CodeAF and logs one line. With a pin, it runs the legacy adapter unchanged.
Binary lookup order: explicit option, then CODEAF_BIN, then the managed binary, then a pinned download, then PATH, then a legacy aforge binary (with a warning). If none is found, an "unavailable" error suggests af codeaf ensure and includes the download error.
- Failed lookups are not cached, and a download is retried at most once a minute.
- The first-use download counts against the harness timeout. Explicit binaries never wait on it.
A download is SHA-256-verified before an atomic rename, and the version marker is written last, so a mismatch leaves nothing installed. An unwritable $AGENTFIELD_HOME/bin is reported as a download failure.
Exit and stop mapping:
- done → success
- deadline → timeout
- unchecked, budget or turn-cap with an answer → success, logged as partial
- question, incomplete, error or price → failure carrying CodeAF's own error text
The doctors (af doctor, af harness doctor and the SDK doctors) use the execution lookup order and never download. For aforge, they report what would actually run.

Rollback and compatibility

Set AFORGE_BIN=<path> to pin the old harness per process.
The aforge stage stays in the agent and cloud images.
af aforge ensure and its download mirror are unchanged.
Products (pr-af, sec-af, cloudsecurity-af, contract-af, SWE-AF, video-essay-af) select provider="aforge" without a pin, so after they pick up this SDK they run CodeAF through the alias. Follow-up PRs in each repo raise the SDK floor and ship CodeAF in their images.

Testing

Every run below uses the real CodeAF v0.8.2 and aforge v0.1.0 binaries. The model was a local OpenAI-compatible endpoint (Ollama, qwen3:14b, via CODEAF_BASE_URL), so no paid model was involved.

Zero-config, fresh home, Python, Go and TypeScript (packed npm tarball): first call downloads and verifies CodeAF, the model writes the requested file, and schema output is parsed.
Windows (native): the Python SDK downloads codeaf.exe and completes the same run.
Control plane + agent nodes (Python and Go): each reasoner runs app.harness() / agent.Harness() through the control plane, the result comes back through the execution API, and execution_usage records harness=codeaf with token counts.
Provider and env matrix (all three SDKs):
- the alias, via argument and via env
- the explicit legacy pin, and the implicit default with AFORGE_BIN (both run aforge v0.1.0)
- AFORGE_MODEL alone selects the model; AGENTFIELD_AFORGE_COMMAND=do launches codeaf do with all seats pinned
- an existing aforge user upgrading, online (gets CodeAF) and with the download blocked (falls back to aforge)
- the opt-out with nothing installed, an unwritable home, HTTPS_PROXY set for TypeScript, a missing CODEAF_BIN
- three concurrent calls (one download), and a forced timeout (exit 3 → timeout)
CLI:
- af codeaf ensure: concurrent installs (one download), idempotent, --force, opt-out
- af harness doctor / af doctor
- the skill-install hook and its opt-out
- af aforge ensure unchanged
Installers:
- install.sh with stable af v0.1.142 → installs aforge; with this branch's af → installs CodeAF; --no-codeaf → installs neither
- install.ps1 on Windows PowerShell 5.1 with main's and this branch's af.exe: both paths work, the caller's session env is untouched, and the skip variables are honored
Docker: the CodeAF stage script runs with empty (classic builder), amd64 and arm64 TARGETARCH; the cloud image is built and smoke-tested in CI.
CI gates, run locally with CI's literal steps on the final tree:
- Python: ruff 0.15.22, mypy, full pytest on 3.11 and 3.10 (2,804 / 2,803 passed)
- TypeScript: lint, full vitest (1,256 passed)
- Go SDK: tidy, build, -race on the new resolver tests
- control plane: build, linux/darwin/windows compile matrix
- desktop: typecheck, tests (673 passed)
- coverage gate passed; patch coverage: control plane 93%, Go SDK 96%, TypeScript 98%, Python changes 99%+ (measured separately; the gate's tracked-module list excludes the harness)
Three independent adversarial reviews of the landing tree. The first found two blockers (legacy pins and legacy command/model settings being dropped) and four should-fixes. The second found four smaller edge cases. All of them are fixed with tests from the contract above, and the final review's verdict was merge.

Not covered or not fully covered:

No paid OpenRouter call. The default model is reached the same way through CodeAF; only the endpoint differed.
codeaf do: launched correctly from the SDK, but a 14B local model does not converge on do within ten minutes, so it was not run to completion.
macOS: cross-compiled only.
python-agent and go-agent images: not built end to end (no Docker on the test machine). Their CodeAF stage script was run directly.
Unrelated failures seen:
- TestDispatchAgentRequestStopsWaitingOnceTheNodeIsDemoted fails intermittently on main too (2 of 10 runs).
- Python agents' harness usage is not stored for successful control-plane executions, sync or async; only failed ones get a usage row. main's SDK behaves identically (tested with a fake harness binary). The status handler returns early on an already-terminal execution before ingesting usage, which is the likely cause. Go agents are not affected. Worth a separate fix.

🤖 Generated with Claude Code

…ed install CodeafProvider runs `codeaf exec --json --dir <root>` (or `do` via AGENTFIELD_CODEAF_COMMAND), parses CodeAF's current envelope and exit ladder, and isolates its state under $AGENTFIELD_HOME/codeaf. The binary is resolved lazily: explicit override, CODEAF_BIN, the managed $AGENTFIELD_HOME/bin/codeaf, a checksum-pinned download of CodeAF v0.8.2, PATH, and finally a legacy aforge binary. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

…alias Provider-less calls now run CodeAF. `provider="aforge"` keeps working and runs CodeAF too, unless a legacy binary is pinned with AFORGE_BIN or aforge_bin, which keeps the original aforge adapter. Harness results carry the provider that actually ran so usage is attributed correctly. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

…anaged install CodeafProvider runs CodeAF headless (`exec` by default, `do` via AGENTFIELD_CODEAF_COMMAND), parses the current envelope and exit ladder, and resolves the binary lazily, downloading the checksum-pinned v0.8.2 release into $AGENTFIELD_HOME/bin when it is missing. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

… an alias Provider-less calls run CodeAF. 'aforge' still works and runs CodeAF unless aforgeBin or AFORGE_BIN pins the legacy binary. Usage is attributed to the provider that actually ran. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

…nstall ProviderCodeAF runs CodeAF headless (`exec` by default, `do` via AGENTFIELD_CODEAF_COMMAND), maps its current envelope and exit ladder onto harness results, and resolves the binary lazily, downloading the checksum-pinned v0.8.2 release into $AGENTFIELD_HOME/bin when missing. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

DefaultProvider is now ProviderCodeAF. ProviderAforge still works and runs CodeAF unless Options.BinPath or AFORGE_BIN pins the legacy binary, which keeps the original aforge adapter. Usage follows the provider that ran. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

internal/codeaf installs the CodeAF v0.8.2 release binary into $AGENTFIELD_HOME/bin, verified against embedded SHA-256 pins, with a lock, a version marker and the AGENTFIELD_SKIP_CODEAF / AGENTFIELD_SKIP_AFORGE opt-outs. `af aforge ensure` is unchanged. Both doctors gain a codeaf row. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

…l and af run The skill catalog install now provisions CodeAF instead of aforge, and `af install` / `af run` repair a missing CodeAF best-effort, so upgrading af alone is enough to get the default harness. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

install.sh / install.ps1 provision CodeAF (opt out with --no-codeaf / -NoCodeaf; the old aforge opt-outs still work). The python, go and cloud images bake the pinned CodeAF binary and fail the build if it cannot be verified; the aforge stage stays as a rollback hatch. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

…e homes and proxies A failed lookup or download is no longer cached for the life of the process: later calls re-check every location and retry the download at most once a minute. CodeAF's state falls back to a private temp directory when $AGENTFIELD_HOME is read-only, the doctor honours CODEAF_BIN, and downloads go through HTTP(S)_PROXY with NO_PROXY support. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

…empty env values Failed lookups and downloads are no longer cached for the life of the process; later calls re-check every location and retry the download at most once a minute. CodeAF's state falls back to a private temp directory when $AGENTFIELD_HOME is read-only, and an explicitly empty per-call variable (such as CODEAF_TELEMETRY_ENDPOINT="") now reaches the CodeAF child instead of being dropped. Legacy adapters keep their existing env handling. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

… homes Failed lookups and downloads are no longer cached for the life of the process; later calls re-check every location and retry the download at most once a minute, and a truncated download now falls through to PATH and the legacy binary instead of raising. CodeAF's state falls back to a private temp directory when $AGENTFIELD_HOME is read-only, and the doctor honours CODEAF_BIN and the temp-managed binary in the same order as execution. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

…install `af run` no longer waits: when CodeAF is missing it starts a detached `af codeaf ensure` in the background and launches the agent immediately. `af install` and skill installs honour the command context and give up on a busy install lock after 30 s. Downloads go through HTTP(S)_PROXY, and the doctors find CODEAF_BIN and the managed binaries in the same order as execution. `af codeaf ensure --bin-dir` installs beside a custom af. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

… and custom installs The installers resolve a single managed home for the whole run (staging, explicit AGENTFIELD_HOME, or ~/.agentfield) and pass it to every child, so a staging install no longer downloads CodeAF twice, and a custom install directory gets CodeAF beside af. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

…ecked The temp-directory fallbacks for the binary and its state now live under <tmp>/agentfield-codeaf-<uid> and are only used when the directory is a real directory owned by the current user with mode 0700, so another local user cannot plant a binary or share the state directory. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

The temp-directory fallbacks for the binary and its state now live under <tmp>/agentfield-codeaf-<uid> and are only used when the directory is a real directory owned by the current user with mode 0700, so another local user cannot plant a binary or share the state directory. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

The temp-directory fallbacks for the binary and its state now live under <tmp>/agentfield-codeaf-<uid> and are only used when the directory is a real directory owned by the current user with mode 0700. An existing shared or world-writable directory is no longer accepted. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

…per-user temp copies `af run` now takes a cross-process lease before starting the background `af codeaf ensure`, so many launches while a download is in flight start one installer, and a failed attempt is not retried automatically for ten minutes. The child is reaped instead of released. The doctors read the temp fallback only from an owner-checked per-user directory, and `--bin-dir` is gone: CodeAF always installs into $AGENTFIELD_HOME/bin. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

Custom install directories no longer get a second copy beside af; the skill hook and the explicit ensure now target the same file, so a fresh install downloads CodeAF exactly once. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

…before reading stdin A provider CLI that fails at startup (a missing key, a bad flag) can close its stdin before the prompt is written. The write error was unhandled and surfaced as an uncaught exception in the agent process; the exit code and stderr already report the failure, so the stream error is now ignored. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

+          response.end(bytes);
+        }
+        if (kind === 'http') response.writeHead(404).end();
+        if (kind === 'redirect-loop') response.writeHead(302, { location: request.url }).end();


github-actions · 2026-10-23T21:23:29Z

Performance

SDK	Memory	Δ	Latency	Δ	Tests	Status
Python	9.1 KB	-	0.26 µs	-26%	✓	✓
Go	232 B	-17%	0.68 µs	-32%	✓	✓
TS	410 B	+17%	1.92 µs	-4%	✓	⚠

⚠ Regression detected:

TypeScript memory: 350 B → 410 B (+17%)

github-actions · 2026-10-23T21:24:53Z

📊 Coverage gate

Thresholds from .coverage-gate.toml: per-surface ≥ 84%, aggregate ≥ 85%, max per-surface regression ≤ 1.0 pp, max aggregate regression ≤ 0.50 pp.

Surface	Current	Baseline	Δ
`control-plane`	88.20%	87.40%	↑ +0.80 pp	🟡
`sdk-go`	93.60%	92.00%	↑ +1.60 pp	🟢
`sdk-python`	94.72%	93.73%	↑ +0.99 pp	🟢
`sdk-typescript`	92.37%	90.42%	↑ +1.95 pp	🟢
`web-ui`	84.89%	84.79%	↑ +0.10 pp	🟡
aggregate	86.11%	85.75%	↑ +0.36 pp	🟡

✅ Gate passed

No surface regressed past the allowed threshold and the aggregate stayed above the floor.

github-actions · 2026-10-23T21:24:54Z

📐 Patch coverage gate

Threshold: 80% on lines this PR touches vs origin/main (from .coverage-gate.toml:thresholds.min_patch).

Surface	Touched lines	Patch coverage	Status
`control-plane`	334	93.00%	✅
`sdk-go`	568	96.00%	✅
`sdk-python`	0	—	➖ no changes
`sdk-typescript`	291	98.00%	✅
`web-ui`	0	—	➖ no changes

✅ Patch gate passed

Every surface whose lines were touched by this PR has patch coverage at or above the threshold.

santoshkumarradha · 2026-10-24T01:28:58Z

@AbirAbbas can we do a round of checks for sweaf and praf as well ? 🙏🏻

install.sh and install.ps1 are served from main but install the latest stable af. Until that release has `af codeaf`, `af codeaf ensure` fails and the machine ended up with no harness at all, while the released SDKs still default to aforge. Detect the command in `af --help` and, when it is missing, provision aforge exactly as earlier installers did. install.ps1 also stops changing AGENTFIELD_HOME and the skip variables in the caller's session (iwr | iex runs in that scope), keeps main's install directory, and drops the -Staging switch it never had before. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

The classic (non-BuildKit) builder leaves TARGETARCH unset, which made the fail-closed CodeAF stage exit with "Unsupported architecture". Fall back to dpkg --print-architecture. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

af run no longer spawns a detached `af codeaf ensure` (and its lease and failure cooldown files); the curl installer, skill install, `af install`, Desktop and the SDKs' own lazy download already provision CodeAF, so af run is byte-identical to main again. The CLI lookup and doctors also stop consulting the per-user temp directory, which the SDKs no longer use. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

Remove the per-user temp-directory fallback. An unwritable home bin directory is now an ordinary download failure (reported in the unavailable error, retried at most once a minute), and CODEAF_HOME is created best-effort without a temp fallback. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

Same change as the Python SDK: no temp-directory fallback for the binary or for CODEAF_HOME. Without any home directory, CODEAF_HOME is left to CodeAF's own default instead of a cwd-relative path. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

…fallback The downloader is a plain https.get with redirects again. When HTTPS_PROXY is set the SDK skips the download (a direct attempt would hang) and points at `af codeaf ensure`, which honors proxies. The per-user temp-directory fallback is removed as in the other SDKs. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

An older af's `aforge ensure` fallback only knows AGENTFIELD_SKIP_AFORGE, so AGENTFIELD_SKIP_CODEAF=1 still downloaded aforge there. Treat either skip variable as CODEAF_MODE=none up front, as install.sh already does. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

… the timeout - No provider named + a legacy pin (AFORGE_BIN, aforge_bin) keeps the legacy aforge adapter, as on main, instead of switching to CodeAF. - AGENTFIELD_AFORGE_COMMAND and AFORGE_MODEL are honored when the CodeAF variables are unset (per-call env first), so the alias no longer replaces a chosen command or model with the defaults. - The harness timeout now covers waiting for the first-use download; a caller that runs out of time gets a timeout result while the shared download keeps going for later calls. Execution gets the remaining budget. - harness_doctor reports what provider 'aforge' would actually run. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

…across downloads Same compatibility rules as the Python SDK (implicit default + AFORGE_BIN or Options.BinPath keeps the aforge adapter; legacy command/model fallbacks; timeout covers provisioning; doctor follows the alias). Explicit binaries no longer wait on the resolver lock, and callers waiting for a shared download honor their own context. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

…g by the timeout Same compatibility rules as the Python and Go SDKs: implicit default + AFORGE_BIN or aforgeBin keeps the aforge adapter, legacy command/model fallbacks, the timeout covers first-use provisioning, and the doctor follows the alias. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

Without AFORGE_BIN the aforge provider runs CodeAF, so the doctor checks CodeAF first and falls back to the legacy binary; with AFORGE_BIN it checks that binary. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

…fig; doctor honors CODEAF_BIN A HarnessConfig whose provider was filled from AGENTFIELD_HARNESS_PROVIDER at construction keeps that provider (as on main) instead of being treated as the implicit default. The aforge doctor no longer falls back to the legacy binary when an explicit CODEAF_BIN is missing, matching execution. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

…commands, doctor honors CODEAF_BIN Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

…s CODEAF_BIN Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

The F3c tests compared two timestamps taken at different moments with 10-50 ms tolerances, which failed under load (seen locally with coverage on). Assert the ordering instead: at least the provisioning time is deducted from the budget, and the inner CodeAF timeout is the remaining budget minus five seconds, taken no later than the outer one. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

AbirAbbas and others added 23 commits October 23, 2026 16:22

feat(desktop): provision CodeAF on launch instead of aforge

9c6d9fd

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

docs: document CodeAF as the default harness and aforge as an alias

cb08af7

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

ci(docker): smoke-test the CodeAF binary in the cloud image

fa386a2

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

AbirAbbas requested a review from a team as a code owner October 23, 2026 21:18

github-advanced-security Bot found potential problems Oct 23, 2026

View reviewed changes

Comment thread sdk/typescript/tests/harness_codeaf_install.test.ts

response.end(bytes);

}

if (kind === 'http') response.writeHead(404).end();

if (kind === 'redirect-loop') response.writeHead(302, { location: request.url }).end();

AbirAbbas closed this Oct 23, 2026

AbirAbbas reopened this Oct 23, 2026

AbirAbbas closed this Oct 23, 2026

AbirAbbas reopened this Oct 23, 2026

AbirAbbas and others added 19 commits October 23, 2026 22:10

docs: describe CodeAF provisioning without the removed fallbacks

9b7f963

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

docs: document legacy-pin, fallback-variable and timeout behavior

e768a6b

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

fix(sdk/go): timeout result on the legacy fallback, case-insensitive …

743c16b

…commands, doctor honors CODEAF_BIN Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

fix(sdk/typescript): case-insensitive command variables; doctor honor…

b9d153f

…s CODEAF_BIN Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

fix(cli): aforge doctor fails when an explicit CODEAF_BIN is missing

ce0b7cb

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

test(sdk/go): require the inner --timeout argument in the F3c test

1a430ef

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

AbirAbbas merged commit 647e31c into main Oct 24, 2026
50 checks passed

AbirAbbas deleted the feat/codeaf-default-harness branch October 24, 2026 13:03

AbirAbbas mentioned this pull request Oct 26, 2026

fix(sdk): fail fast with a clear message when the CodeAF harness has no model key #1130

Merged

4 tasks done

Navigation Menu

Repository navigation

Make CodeAF the default harness, installed automatically with AgentField - #1128

Make CodeAF the default harness, installed automatically with AgentField#1128
AbirAbbas merged 42 commits into
mainfrom
feat/codeaf-default-harness

AbirAbbas commented Oct 23, 2026 •

edited

Loading

Uh oh!

github-actions Bot commented Oct 23, 2026 •

edited

Loading

Uh oh!

github-actions Bot commented Oct 23, 2026 •

edited

Loading

Uh oh!

github-actions Bot commented Oct 23, 2026 •

edited

Loading

Uh oh!

santoshkumarradha commented Oct 24, 2026

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Conversation

AbirAbbas commented Oct 23, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Behavior (validation contract)

Rollback and compatibility

Testing

Uh oh!

github-actions Bot commented Oct 23, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Performance

Uh oh!

github-actions Bot commented Oct 23, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

📊 Coverage gate

✅ Gate passed

Uh oh!

github-actions Bot commented Oct 23, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

📐 Patch coverage gate

✅ Patch gate passed

Uh oh!

santoshkumarradha commented Oct 24, 2026

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

AbirAbbas commented Oct 23, 2026 •

edited

Loading

github-actions Bot commented Oct 23, 2026 •

edited

Loading

github-actions Bot commented Oct 23, 2026 •

edited

Loading

github-actions Bot commented Oct 23, 2026 •

edited

Loading