agy-staff

agy-staff 0.9.0 — DeepSeek Harness plugin & visual subagent sidebar

Highlights

agy-staff 0.9.0 adds native support for DeepSeek Harness (dsh) (#54 / #55), pairing DeepSeek Harness sessions with Google's Antigravity CLI (agy) through three coordinated layers: a visual right-sidebar subagent companion, native Cordis lifecycle tools, and hydrated /agy-* skills.

1. Visual subagent right-sidebar (dsh/client.js)

• Role capability picker: Quick-insert cards for /agy-brain (Argon), /agy-lead, /agy-staffer, /agy-researcher, /agy-reviewer, and /agy-implementer (3.8 Flash) with crispEdges pixel-art SVG icons and a persistent 中 / EN language toggle.
• Multi-turn conversation cards: Groups linked continue --job <id> runs into a single expandable thread card with per-turn Query and Reply viewports, distinct pixel status badges (done, running, error/crashed, canceled, attention), and one-click continuation / progress check insertion into the session input box.
• Locked single-line live activity strip: While a subagent is running, .agy-live-strip stays mounted at a fixed 20px single-row height above Query / Reply, collapsing consecutive identical tool steps into ×N counts (newest step first) so the cards below never shift on 5-second polls.
• Automatic completion transition & low-overhead polling: Expanding a running card never locks turnDetails[id].loaded early; when the turn transitions to a terminal state, the sidebar automatically loads <id>.result.md. Idle polling reads only state.json (1 read/5s), caching .spec.json per job ID and reading .progress.json exclusively for running / queued jobs.

2. Native Cordis lifecycle tools (dsh/index.mjs)

DeepSeek Harness's command sandbox blocks agy from binding its localhost language-server port and reading its OAuth keychain credentials. The server plugin registers three native Cordis tools that invoke companion/agy-companion.mjs outside the command sandbox with array arguments (spawn(process.execPath, [...])):

• agy_brain: Starts a staffer --brain job (defaulting to gemini-4-argon-medium) or continues an existing brain thread via job_id, returning { ok, job_id, status, output }.
• agy_worker: Dispatches research, review, implement, staffer, or synchronous ask workers (or continues a thread via job_id). Passing restricted: true opts into --restricted; omitting or passing false defers to .agy-staff/config.json.
• agy_job: Full background job lifecycle tool supporting wait (60-second soft timeout with timeoutMs: 75000, treating exit code 0 as done and exit code 2 as running with ok: true), observe, result, cancel, and status.

3. Single-source /agy-* skills & isolated bundle packaging (dsh/, package.json)

• Isolated dsh/ subpackage: Root package.json declares dsh.bundle.patch (./dsh/cordis.patch.yml) and exports . (./opencode.mjs), ./icon (./dsh/icon.svg), and ./package.json, keeping the OpenCode V1 and Pi entrypoints untouched while displaying the custom pixel rainbow arch icon in DeepSeek Harness's Installed plugins list.
• Hydrated skills from opencode-skills/: dsh/index.mjs registers all eight agy-* skills directly from opencode-skills/, replacing <skill-dir> and <plugin-root> at registration time and prepending a DeepSeek Harness note directing models to use agy_brain, agy_worker, and agy_job.

Validation

• Passed npm run check:skills (Pi & OpenCode generated skills verified, 0 drift).
• Passed all 243 offline tests across 30 suites (npm test), including tests/dsh-plugin.test.mjs (covering bundle manifest exports,

agy-staff 0.8.2 — First proposal from an external brain

Use agy:brain when you want Argon to frame the problem before the host settles on an approach. The host supplies the goal, constraints and existing evidence, tests the proposal's assumptions, and owns execution and delivery. For example, if a migration proposal assumes every historical job has a conversation ID, check the old records and return any counterexample to the same brain conversation before implementing it.

The brain skill starts a background staffer --brain job. It uses an advisory prompt and defaults to gemini-4-argon-medium; an explicit model selection is supported. A native Flash research child may gather read-only evidence. Regular workers keep their no-delegation rule. Follow-ups use continue --job <id> as in 0.8.1.

This first iteration makes no comparative cost or quality claim. For the mechanism probe and its limits, see #50 and #53.

Live acceptance check

On November 6, 2026, a host ran the shipping staffer --brain path in a disposable repository using authenticated AGY 1.5.2, without a model override. The first brief requested a design for migrating job history and supplied only the one-to-one pattern seen in standalone runs. Argon Medium recommended keying reports by job ID, but proposed mapping historical link keys to a single report, and asked the host to check cardinality.

The host found four real, linked jobs with separate reports sharing one conversation ID; 15 completed jobs with no conversation ID; and a 0.8.0 foreground path that could record a conversation without a job. It sent those observations through continue --job. Argon retained the job ID as the report address, dropped the single-report conversation alias, separated jobless pointers from reports, and deferred URL rules pending an inventory. The host accepted the changed identity policy but treated the suggested leaf-job handle as a UI default, not a CLI requirement: the companion looks up the explicitly selected job and checks its conversation and running status.

Both job records show brain: true, gemini-4-argon-medium, and native success; the second has the first as parent and the same AGY conversation ID. This was one real two-turn design task, not a model comparison, a Flash delegation test, or a migration implementation.

agy-staff 0.8.1 — Clearer diagnostics. Better continuity.

Changes since 0.8.0

Two quality improvements: #48 / #49 preserve AGY's native diagnostics, and #51 / #52 give every direct persona shared dispatch guidance and an explicit job-based continuation interface.

Clearer diagnostics. The companion retains native termination reasons and bounded typed error metadata alongside delivered responses, including partial answers. Hosts can distinguish normal completion, print timeout, interruption and API failure without inferring the cause from prose. Normal warning-free runs keep their low-noise output and intermediate-file cleanup. CLI version checks run only when an option or output format is rejected without a structured API error; plugin updates do not update the separate AGY binary. Errors never trigger an automatic retry or update.

One source for persona rules. templates/persona-skill-rules.md maintains four short host-facing rules: reuse helpful context for related follow-ups, start fresh when the task or role changes or an independent opinion is needed, pass new host-chat decisions without expanding authorization, and follow the existing execution and recovery boundaries. npm run generate:skills embeds them into all five canonical persona skills before generating Pi and OpenCode adapters. Claude Code and Codex load the canonical files directly. npm run check:skills and CI detect stale output or malformed shared blocks. Host dispatch guidance remains separate from worker prompts and guardrails.

Continue the right job. Natural follow-ups should reuse the matching completed job when its context helps, even if the user does not say “continue.” The host retains each job ID with its assignment and sends intervening decisions in the follow-up brief. This is guidance for the host agent, not an automatic topic-matching system. New topics, different personas, independent reviews and explicit fresh starts use new sessions.

The single continuation entry point is continue --job <id>. It retains the original persona and execution style, inherits the recorded model, permission profile and working directory unless a supported explicit override is supplied, and creates a linked invocation. Synchronous ask now records a job ID too: its answer remains on stdout, while stderr carries the job ID. Continuing an ask stays synchronous. Interleaved ask calls do not redirect default background result collection.

Continuation migration

Use the job ID returned by the earlier invocation:

node "<plugin-root>/companion/agy-companion.mjs" continue --job <id> --prompt "Check the alternative discussed above"

• Persona --continue, --conversation <id>, persona --job <id>, and bare continue are rejected before invoking AGY. Errors provide a job-based migration command when a matching record exists.
• A conversation ID is not a job ID. Use status to inspect recorded jobs; a continued job keeps its original persona. To change roles, start the target persona with the relevant findings in its brief.
• Pre-0.8.1 ask conversations without a job record cannot be resumed through the new interface. Start a fresh ask to obtain a job ID; other existing jobs with recorded conversation IDs can be continued.
• A running job still must finish or be canceled before continuation. Execution-timeout recovery still requires explicit confirmation; a wait soft expiry only requires collecting the same job again.

Validation

• The implementation passed 233 offline tests, including job-ID-only continuation, synchronous ask records, legacy-selector migration, shared-source generation and packag

agy-staff 0.8.0 — Bench v1 findings in the lead, jobs and worker instructions

Changes since 0.7.4

agy-staff Bench v1 ran 0.7.3 on Claude Code and Codex over SWE-bench Verified, SWE-Atlas QnA and DeepSearchQA. Its trajectories showed where delegation lost time, tokens or correctness. #43 moves those findings into the instructions and adds the 20-task Codex 0.8.0 validation to Bench v1. The companion runtime and the job lifecycle are unchanged.

Waiting. Every wake-up rereads the host's whole context, so the jobs skill now asks the host to wake as rarely as it can. When the host is known to wake the model on a background command's completion (such as Claude Code), the host ends its turn after launching the wait instead of sleeping or polling. Otherwise it stays in its turn and sets the longest supported block explicitly (5–10 minutes, or the tool's maximum such as 300,000 ms in Codex write_stdin / // @exec, or a foreground wait <id> --timeout 10m in synchronous-shell hosts like Pi and OpenCode). In v1, Codex made 872 collection calls over 20 trials, most returning after the default of about 30 seconds; the skill now names the Codex parameters (write_stdin yield_time_ms: 300000, or the code-mode // @exec pragma).

Briefs. The host copies the user's constraints and prohibitions into the brief verbatim, in their own section, and keeps orchestration instructions such as "use $agy:lead" out of it. In v1, paraphrased briefs dropped a benchmark rule 7 times in 20, and forwarded orchestration instructions led one worker to launch a nested agy job in the same worktree. Briefs now separate settled decisions from open ones after the host checks the facts they rest on, and a deadline sets the job's --timeout and appears in the brief.

Delegation and acceptance. Lead still delegates substantive work by default and now states the trade-off: delegation saves the host's context and quota, not total time or tokens. A tightly coupled change stays in one assignment with its investigation and verification. Acceptance reads the worker's status and reported conflicts first, checks interface choices, departures from the brief and figures reported as verified, and stays well below the cost of redoing the work. The host fixes small, understood defects directly and sends larger gaps back with continue.

Workers. The staffer, research and implement templates share a short set of working rules:

• do the assignment without delegating further;
• follow the brief's scope and decisions, and report confirmed contradictions between the brief's factual premises and what the worker observed under ## Conflicts with the task instead of building on a false premise;
• respect a stated deadline and prefer a partial report with named gaps;
• write the final message only after all work has finished, beginning with Status: DONE, DONE_WITH_CONCERNS or BLOCKED.

The review template gains the delegation, deadline and final-message rules. The research template drops its defensive framing, asks for full source documents once a search finds them, never repeats a query, and stops searching after two similar results; in v1, timed-out research jobs spent up to 93 searches on snippets. The implement and research templates ask for whole reads of small files, searches with context lines, an exploration stop once the change can be named, and reusable probe scripts; in v1 the worker took a median of 104.5 steps per SWE-bench job, one tool call per step.

docs/DESIGN.md records the shared worker rules, the brief boundary for orchestration instructions, the cost of each wake-up, and the revised lead defaults. Package, Claude Code, and Codex manifests are versioned at 0.8.0.

Validati