Work-state, not transcript.

i tried to hand a session to another harness by giving it the transcript. it did not work, and the reason it did not work is the whole post.

a transcript is a log of how i got here. it keeps the questions i asked in the wrong order, the tool calls that failed, the tangents, and the correction three messages later. it is faithful to the conversation and expensive as a starting point. hand it to a fresh agent and the agent has to reconstruct intent, decide which conclusions still hold, and rediscover where the code actually stands.

that is not resuming work. it is watching someone read your diary and try to become you.

01 / THE UNITThe transcript was the wrong unit

the right unit is work-state: a compact description of where the work is now, without replaying every step that produced it.

IGRIS persists the surfaces that define that state and reads them again at boot. the model can change. the harness can change. the session does not need to impersonate the one before it. it needs the contract the previous session left behind.

that distinction sounds small until an agent stops in the middle of a task. a transcript tells the next agent what was said. work-state tells it what is true, what is claimed, what changed, and what should happen next.

02 / SEVEN FACETSSeven facets have to be reconstructed

i wrote down what a new harness actually needs to reconstruct. seven facets:

• brief — what we are doing and why; the intent, not the chat that produced it.
• phase — where the work sits in the delivery lifecycle.
• atomic claim — the specific unit being worked, claimed so another instance does not take it silently.
• instance identity — which run owns the claim.
• supersession lifecycle — which state replaces which, so stale work can lose cleanly.
• working tree — the staged and unstaged files; the real shape of the change.
• agent log — decisions and reasons worth carrying forward, not a replay of keystrokes.

those facets do not live in one seven-field payload. they are recovered across the brain, session state, claim gate, phase machine, agent log, and the working tree. the portability claim is narrower and more useful: a supported harness can reconstruct enough of them to resume the same work without receiving the prior conversation.

03 / HONEST TIERSEvery harness, honestly

harnesses are not equal, and pretending they are would make the architecture less trustworthy.

IGRIS derives the tier from a capability the harness either exposes or does not: hooks. on a first-class harness, a gate can block a write or an out-of-phase commit. on a bridge harness, the contract, brain, skills, and MCP surface remain available, but enforcement softens where the projected hook set cannot run blocking gates. the workflow still runs. that blocking does not.

1234FIRST-CLASS  Claude Code · OpenCode · Antigravity
BRIDGE       Codex · Gemini CLI · Cursor
TIER TEST    hooks supported → gates block
           no projected hooks → gates advise

Cursor is the sixth declared harness, not a sixth first-class harness. that correction matters. "supported" should describe the surface that exists, not imply guarantees the adapter cannot enforce.

04 / CONTROLLED PROOFThe proof forced an interruption

the useful test was a controlled same-machine proof in a disposable fixture, recorded as B2/G-14 on 2026-06-16.

the controlled sequence tested three mid-workflow handoffs: Claude to OpenCode, OpenCode to Codex, and Codex to Antigravity. the first handoff re-walked architect and forger before reaching sentinel; later handoffs resumed straight to sentinel. a crash-recovery leg then returned the third brief to Claude and exercised held-claim refusal and force-reclaim before quality gates and final recovery. none of the receiving harnesses got a pasted prompt or transcript.

what matters is not that a session was saved and reopened. the result is that the handoff contract was exercised across four tools through recorded state rather than transcript transfer. the repository keeps the B2/G-14 account on the current develop branch and separately labels the public storyboard as an illustration rather than the original proof asset. that distinction stays visible because a polished visual is not a substitute for evidence.

05 / PROJECT HANDOFFA project needs a different payload

a session handoff and a project handoff follow the same principle, but they do not carry the same data.

session resume restores live work-state. project handoff exports a project slice for another installation or another owner:

• export produces a portable bundle from the project's briefs, brief graph, context documents, and goals, with explicit tiers for narrower or wider slices.
• import accepts that bundle only after integrity checks and a reviewed ancestor-based merge. new, unchanged, local-only, incoming, and conflicting rows are classified before anything is applied.
• /handoff wraps both verbs in one guided flow with preview and confirmation, while the architecture records the boundary between continuous same-owner sync and point-in-time cross-owner transfer.

machine-local state does not belong in that bundle. sessions, active instance claims, metrics, secrets, and executable code stay behind. the receiving side gets project knowledge it can review, not a copy of someone else's machine.

that is why the project path is not "the same seven fields, but bigger." it is a separate transport for a separate ownership boundary.

06 / THE BOUNDARYWhat is proven, and what is next

same-machine cross-harness resume is proven. the four-tool path is documented. project export and reviewed import are implemented.

cross-machine continuity is still the frontier being tightened. deterministic phase entry across every harness still has unfinished edges. bridge harnesses still have advisory gates where they lack the hook surface needed for enforcement.

those limits are part of the result. portability is not a sentence in a prompt. it is a contract made of state, transport, capabilities, and boundaries. if one of those is missing, the handoff claim should get smaller too.

07 / CLOSINGState travels. The conversation stays.

the log of how you got here is not the description of where you are. only one of them is portable.

IGRIS resumes the work by reconstructing the brief, phase, claim, instance, supersession, tree, and decisions from their authoritative surfaces. the next harness does not need to become the previous conversation. it needs to know what the previous session left true.

work-state, not transcript.