feat(cli): agents label their own tab, and the task row collects it (#358)

[GabrielDumbrava]

Closes #358.

What

Agents label their own tab with small key/value properties, and the task's sidebar row collects them.

• termic prop <key> <value> sets a property on the caller's OWN tab (termic prop ticket ABC-1).
  termic prop <key> "" clears it, a bare termic prop lists the task's properties, termic prop <key> one key.
  --tab targets another tab (the usual id / index / title selector, reaching any strip tab like tab close), --task / --project another task.
• MCP task_prop does the same (set, clear, list) and defaults to the caller's task, like the CLI.
• list --json and status --json carry the collected view (props on the task) and each tab's own (props on each tab).

Sidebar

• Task row, after the name on the left (never in the badge/kebab slot): one entry per key, keys in the order they were FIRST set anywhere in the task, each key's distinct values joined with , in tab order.
  Two agents on two tickets read ABC-1, ABC-2; with a status too, ABC-1, ABC-2 · review.
  An update keeps its key in place. The row's tooltip lists every key: value with the tab it came from.
• Each tab's row (task expanded) shows only what that tab set.
• The sidebar filter (Sidebar: per-project task filter (text + notifications bell) #324) matches property values, so typing ABC-2 finds the task working on it.
• Nothing changes for tasks without properties, and nothing is shown in compact mode.

How

• Properties belong to the tab, so agents in one task never overwrite each other.
• TERMIC_TAB_ID is injected into every agent PTY, on the host and through the env overlay docker::build_spec forwards into the container, so termic prop knows the caller's tab.
  Only agent PTYs get it (the aux terminal has no tab); it is reserved against extra named ports.
• The webview store is the only writer (setTabProp, reached by a tab_props RPC).
  persisted_tabs is rewritten whole by syncDurableTabs, so a Rust-side write would be overwritten by the next sync; the store applying one write at a time is what makes a single key atomic.
  An unchanged value writes nothing (bear trap 8).
• Persisted with the tab (props on PersistedTab, each entry with the time its key was first set), so durable agent tabs keep them across a relaunch.
  Closing a tab drops them; writing into a task that is not open is refused, the tab close rule.
• One collect rule, lib/tabProps.ts collectTaskProps: the sidebar renders it and cliAgentState pushes it to Rust for list / status, so the CLI and the window cannot disagree.
• Validated, not escaped: keys [a-z0-9][a-z0-9_-]* up to 32, values one line up to 40, 8 keys per tab.
  One rule in termic-proto, used by the CLI before it touches the socket, the server and the MCP tool.
• Wire protocol v18 (v17 was taken upstream); an older app answers "rerun your command".

Tests

• termic-proto (limits), termic-cli (parsing, pre-socket guard), cli_server.rs (routing to the resolved tab by every selector, list and key filter, refusals before any RPC, error mapping), mcp_server.rs (task_prop), lib.rs (only an agent PTY gets TERMIC_TAB_ID).
• lib/tabProps.test.ts (validation, apply, collect order and dedup), store/cliTabProps.integration.test.ts (the real handler on the real store: per-tab writes, persistence, no-op, clear, typed errors, reading an unopened task), taskFilter.test.ts (property values match).
• e2e cli.e2e.ts: 6 new cases over the real socket and the real CLI binary with $TERMIC_TAB_ID: the task row and tab rows, key order and in-place updates, status --json, refusals and clear, the filter. mcp.e2e.ts's pinned tool list gains task_prop.
• After rebasing on main at 6143e38: cargo test 1235 + 81 + 21 pass; cli.e2e.ts 44/44, mcp.e2e.ts 20/20.
• npm test: the 18 failures are the same 18 that fail on main (profileScope, taskLabel, bottomTerminal.integration, pr, usageUnknownDismissed).
• Full make e2e before the rebase: the files that failed were checked one by one; credentials and terminal-find fail identically on main, and the rest pass run on their own (the agent.e2e.ts IIP image case passed 3 runs out of 3).

tools_list_size_is_a_conscious_diff moves from 17800 to 18400 bytes for task_prop, with the reason next to the number.

Manual test

The author tested it by hand in the dev app and reports it works.

Docs: docs/data-model.md (the tab property entity), docs/cli-agent-instructions.md, the TERMIC_CLI_HELP hint agents get, docs/e2e-coverage.md.

🤖 Generated with Claude Code

[GabrielDumbrava]

[simion]

Reviewed the termic prop / MCP task_prop implementation: the TS/Rust validation rule is genuinely shared, the store remains the sole writer, TERMIC_TAB_ID is correctly scoped to agent PTYs, and the v18 negotiation degrades gracefully against older clients. No blocking issues found.

[simion]

Merged — thanks for this one, Gabriel. termic prop is exactly the kind of small, well-scoped primitive that makes the agent-tab story click into place. Appreciate the contribution.