Decide the Harness Seam's tool, command, diff, and reasoning-summary events

[Sandstorm831]

Question

What must the Harness Seam's Turn events carry so the Run Workbench can draw each tool call as one row that changes in place, typed by kind with its main input and result count, show command output and exit code, show file diffs, and show a collapsed reasoning summary? Decide the call identity that pairs start and end, the running, failed, and declined states with error text, command output and diffs as data rather than summary strings, the reasoning-summary event, and which native signals (rate limits, token usage, account notices, unknown methods) stop becoming generic activity while both Adapters fill context usage.

[Sandstorm831]

Starting context

Input	Why it matters	Authority
ADR 0036 and its ADR 0022 amendment	The transcript needs paired tool rows, command output, file diffs, reasoning summaries, and no protocol noise; raw chain of thought stays private	Canonical
Secant Run lifecycle glossary	Turn, Harness Session, Harness Request, and Reasoning summary vocabulary	Canonical
Establish whether Claude Code's direct stream-json transport emits reasoning summaries, and under which setting and its research	The direct Claude Code transport's observed reasoning frames and controls, compared with Codex's summary and raw streams; version and model qualifications matter	Research
Study how T3 Code turns Claude Code and Codex events into a conversation	Tool identity and lifecycle, command output and diffs, reasoning, noise filtering, and the documented gaps in Secant	Research
Harness Interface, Claude Code Adapter and frame model, Codex Adapter and runtime protocol	The current event contract and exactly where native facts are dropped or flattened	Canonical (current behaviour)
Module design, topology, dependencies, and Harness Module guidance	Keep native types behind the Adapters and the semantic event contract owned by the Harness Seam	Canonical

Required methods

• /grilling and /domain-modeling; consult /codebase-design for Seam and ownership decisions.

Reading rule

Read ADR 0036's transcript requirements and reasoning-summary paragraph, then the reasoning research's conclusion and qualifications. Use the T3 Code note to compare each tool, command, diff, context, and noise requirement with the current Harness Interface; follow its pinned primary-source links only where the decision needs more detail. Resolve the semantic event contract with the human. This ticket does not implement the Adapters or decide the Run Projection's durable/live history and headless JSON contract; that is the separate dependent decision.

[Sandstorm831]

Resolution

The human agreed the seven decisions below, including the final requirement that live output stay bounded. This resolves the Harness Seam contract; production implementation and the separate Projection/headless decision remain outside this session.

The accepted contract is recorded here and in ADR 0038: Carry observed tool and summary facts through the Harness Seam, with explicit ADR 0022 amendments and ADR 0036 clarifications. The human amended and pushed the documentation as a signed commit verified by GitHub. The documentation PR is merged into main at merge commit 16fb18e.

1. Summary settings: inherit the Harness's existing behavior, including unset defaults. Secant does not enable summaries or change thinking/effort for Thought rows. Display supplied provider summaries; raw reasoning stays private.
2. Thought rows: use the first nonempty summary line as the shortened collapsed label; preserve the complete expandable body. Show Thinking and a spinner while streaming. Show a duration only when the Harness reports a trustworthy reasoning duration, never from summary-delivery timing.
3. Command output: preserve native command text, supplied output, and observed exit code as separate facts. Optional native live output updates one identified, replaceable preview; completed output reconciles without duplication. Both live and completed panels are bounded by default, with ADR 0036's ten-line collapse, width-aware display, and explicit human expansion. Streaming never takes over the screen. No polling or forced settings to obtain unavailable live output.
4. Tool identity and outcomes: stable opaque call identity within the Turn, observed parent relationship, and reported running/completed/failed/declined states with useful errors. Turn settlement stops live indicators; a tool without a result remains unconfirmed. No invented tool success, failure, or interruption.
5. Typed rows: read, search, command, file change, web, MCP tool, subagent, and other; meaningful main inputs and optional reported counts with units. Unknown counts remain absent. Native details are translated behind the Adapter, not guessed by clients from summary prose.
6. File changes: display only supplied paths and patches at their reported call or Turn scope. A cumulative Turn diff is not attributed to an arbitrary call. No Git/filesystem reconstruction, and a requested edit is not proof that it happened.
7. Context and noise: show Harness-reported context/usage values with their meanings preserved. No tokenizing, summing usage into occupancy, derived percentages, guessed window sizes, or repairs to inconsistent numbers. A percentage appears only if reported. Usage/account/rate notifications and unknown methods do not become transcript activity; genuine semantic failures/requests remain meaningful facts.

Architecture: the Harness Module owns native parsing, classification, correlation, and source-specific output assembly. The application Projection Module owns bounded per-call reconciliation, preview replacement, and Turn-derived liveness behind the existing Projection Port. Presentation owns wrapping, collapse, expansion, and colour. The agreed features share identified rows and replaceable previews; they introduce no new execution controller or raw-provider client.

Investigation assets: the inspected OpenCode shell renderer, OpenCode owned cleanup, T3 Code command-delta filtering, and T3 Code call collapse. The extended OpenCode note and T3 Code note are published in the same signed commit. Source and existing tests were inspected, not executed. OpenCode's existing TUI shows bounded live shell previews; T3 Code drops live output and shows shortened completed summaries. Both use stable call correlation. Their synthetic cleanup outcomes are not copied into an external-Harness contract. Recorded Codex output/exit fields are established; Claude shell exit-code extraction still needs native fixture qualification.

Hand-off: Decide how a Turn's history grows mid-Turn in the Run Projection and headless --json now owns retained preview budgets, interrupted partial-output retention, final reconciliation, durable publication, cumulative diff placement, and frozen headless contract exposure. Its Starting context links this resolution and the pinned documentation. No other decision is resolved here.