[Feature]: User-controlled context reset (threshold-based) + external spec-file continuity

[ShaoSK]

Related, but not covered by, existing issues: context-usage observability is already proposed in #2038 and #4984. This issue is specifically about user-controlled context reset and how work continuity survives that reset — a workflow angle those issues don't cover.

Deployment type

Official App

What do you want and why?

Background — how I work with a single coding agent today

When I use a coding agent (e.g. Claude Code) directly, my standard workflow is:

1. I keep an eye on the session's context-window utilization.
2. Once it reaches ~30%, I deliberately restart a fresh session — a full context reset, not a compaction — because response quality degrades well before the hard limit.
3. Project progress does not live in the agent's conversation memory. It lives in local spec files (open-spec style: plan / tasks / decisions written to disk). Each fresh session re-reads those files and continues exactly where the previous one left off.

So for me, a large context window is not something to "fill up" — I actively want short, clean sessions, with continuity handled by external files rather than by an ever-growing transcript.

The problem on Multica today

In the current version I cannot reproduce this workflow:

• No context-usage visibility per agent run. I cannot see what % of the context window a running agent is consuming, so I have no signal for when to reset. (This half overlaps [Feature]: Agent Context Window Observability #2038 / [Feature]: Improve context management in issues #4984.)
• No user-controlled reset. I cannot say "start this next run from a clean context." Comment follow-ups appear to resume the prior session by default (see Comment resume should fall back to low-token mode for long or tool-heavy sessions #4754), and the only reset-ish mechanism is the /fresh directive PR (feat(server): /fresh comment directive to skip session resume #1298), which is a manual per-comment flag with no threshold and no UI.
• No first-class continuity contract. There is no supported "the agent's durable progress lives in these spec files, the transcript is disposable" model. Resetting the session today risks losing progress, because progress and transcript are conflated.

The net effect: I'm forced to let sessions grow unbounded and hope compaction/resume behaves, instead of running the short-session + external-spec loop that I trust.

Proposed solution

Three parts, roughly in priority order. Part A depends on the observability work in #2038 / #4984; B and C are the new asks.

A. Context-usage signal (prerequisite — may be satisfied by #2038 / #4984)
Surface current-run context utilization (used / limit, %) somewhere I can see it before/while an agent runs.

B. User-controlled reset, ideally threshold-based

• A manual "reset context / start fresh" control on the issue or agent-run panel (not only a typed /fresh comment).
• Optionally an auto-reset threshold I can configure (e.g. "start a fresh session whenever utilization would exceed 30%"), so I don't have to babysit the number.
• Reset means: drop the prior transcript, start a clean session — distinct from /compact.

C. External spec-file continuity (the part that makes reset safe)

• A supported convention where the agent's durable progress lives in on-disk spec files (plan / tasks / decisions), and every fresh session is guaranteed to re-read them first.
• This decouples "progress" from "transcript", so resetting the context is cheap and lossless. This is the key difference from compaction: compaction tries to preserve the transcript; I want to discard it and rely on files.

Prior art / references

• [Feature]: Agent Context Window Observability #2038 — Agent Context Window Observability (usage %)
• [Feature]: Improve context management in issues #4984 — Improve context management in issues (usage visibility + compact entry)
• Comment resume should fall back to low-token mode for long or tool-heavy sessions #4754 — Comment resume should fall back to low-token mode
• feat(server): /fresh comment directive to skip session resume #1298 (PR) — /fresh directive to skip session resume (closest existing mechanism)
• Open-spec style workflows: progress tracked in local plan/task files rather than model memory.

Screenshots / mockups (optional)

Rough sketch of what the signal + control could look like on an agent run:

Context: 58% used (116k / 200k)   [ Reset context ]   Auto-reset at: [ 30% ▼ ]
Continuity: progress written to ./specs/*.md  (re-read on every fresh session)